Datasets:

---

Srijan-Srivastava

/

webtext-super-tiny

like 1

ERROR: type should be string, got "https://youtu.be/5vp9ypOUgMw\n\nVibe Coding is For Senior Developers\n\nAI coding is really, really cool if you suck at writing code yourself. That's why we're only seeing junior devs who are using these tools. People like Dan Abramov, the creator of React, or DHH,\nDan Abramov\nthe creator of Rails, or Lionus Torva,\nthe creator of Linux, or Anti-Res, the creator of Reddus. Wait, they're not junior. What the hell's going on? Oh boy. It feels like there's just been an\nexplosion in these really talented developers suddenly embracing and taking on these AI tools. It's been wild to watch. On one hand, they're a bit late,\nbut on the other, it still feels like we're all so so early. And a lot of these people, even the ones that I'm not necessarily the most fond of, have\nhistorically cared a lot about their craft and the details and doing things right. Which is why it's kind of crazy\nto see all of them embracing AI the way they are now. If I'm being honest, I don't know many devs who are shipping a\nlot of code that aren't embracing these tools at this point, especially experienced developers. And there's a lot of reasons for that. I've been\nthinking a lot about this post from Eric over at Cursor. Turns out senior engineers accept more agent output than juniors. This is a very interesting\nobservation. And I have so much to talk about. One thing I have to talk about first though is today's sponsor. AI models are pretty smart, but they get way smarter when they have access to a\nbrowser. And no, using curl and bash doesn't count. I'm talking an actual browser where they can click buttons,\nsign in, and do real work. That's why I love today's sponsor, Kernel. They're the fastest way to get a browser out for your agents to use. And I mean fast in a\nlot of different ways. They spin up in under 400 milliseconds. It takes under 10 minutes to get started using them.\nYes, really. And also, just a quick anecdote, I reported a bug with an older version of their package because I hadn't upgraded yet. And within 10\nminutes, mind you, at 900 p.m., within 10 minutes, they had shipped a fix. That kind of response time is essential for\nreal businesses, which is why companies like Cash App have already made the move over to kernel. Here's an example action from one of their example runners. We give it a name, create browser for\ntesting. In this case, it is a promise that takes in the context. It creates a browser and then leaves it on because this is a browser we want to be able to\nuse for things. So, you run the command and now you have a URL to a reall life browser running in the cloud that you or your agents can use. Give your agents access to the entire web at soy.v.link/\nlink/ colonel. And yes, I'm still sick.\nSorry. I uh have the terrible upper respiratory thing going around SF and the bay right now. Hopefully, I'll have my voice back proper very soon. I\nstarted thinking about this video when DHH made this blog post about promoting AI agents. What he's specifically referring to is that the role these\nagents have in his day-to-day work has changed and he is promoting them effectively to a higher level of involvement in his work and day-to-day.\nAs he says here, \"I'm ready to give the current crop of AI agents a promotion.\nThey're no longer just here to help me learn or answer my questions or check my work. They are fully capable of producing production-grade contributions to real life code bases.\" Pure vibe\ncoding remains an aspirational dream for professional work for me for now.\nSupervised collaboration though is here today. I've worked alongside agents to fix small bugs, finish substantial features, and get several drafts on\nmajor new initiatives. The paradigm shift finally feels real. Yeah. So, if you are talking about these things as though they're only useful for people\nworking on small side projects or everyone's favorite comment, Theo doesn't know anything about real big code bases. He only works on startup\nscale. Your god and save your DHH is now on our side. It's happened. These tools are no longer just ways to quickly edit\nsome CSS on a page or look into how something works or answer questions.\nThey are commanding real systems and doing real work. How you choose to use and embrace that can vary a lot. For\nexample, Lionus who recently said that vibe coding is okay as long as it's not used for anything that matters.\nSpecifically using the definition of vibe coding where you aren't reading the code, you are just generating it to go do something. And he is putting his\nmoney where his mouth is and does this himself. He put up a repo recently for digital audio effects audio noise and he wanted to do some visualizers in it. And\nwhen he did that, he decided that his knowledge of Python and visualizations with it aren't really his strength. And rather than try to learn it and do it\nhimself or go do a bunch of Googling, he actually vibecoded it himself. I know more about analog filters, and that's not saying much than I do about Python.\nIt started out as my typical Google and do the monkey see monkey do kind of programming, but then I cut out the middleman himself and just use Google\nAnti-gravity to do the audio sample visualizer. While I don't necessarily think anti-gravity is the best tool to embrace the modern AI coding experience,\nthe fact that he is seeing value here is huge. If you haven't already watched my video on not falling behind with AI stuff, I highly recommend watching that\nbecause I go in depth on a lot of pieces on how to embrace and catch up on these things. One of the things that I really tried to push here is this idea of\nthinking out of the box. What are some things that you would build if you had more time and knowledge, but don't? so you can't like just little one-off tools\nthat you wish existed. Those could be a library that exists but doesn't quite do what you want or some web app to manage something in a game that you play. What\ntools, what accessories, what pieces of software don't exist that you wish did.\nAnd the more you can think in this way where you realize that a problem you have in a big codebase might not be best solved in that big codebase, maybe it's\nbetter solved in a small random repo next to it that lets you debug something or look into something or experiment. If you have a theory about a certain\nlibrary you want to play with in your project, but implementing it in that project is too complex, you go spin up something to test it out and test the limits. If you're curious what different\nAI models are capable of, you can build your own evals and scaffolding to test those things without having to actually look at write or much less read the code\nyourself. The thing to understand here is that AI can write code that works and solves useful problems. How big of\nproblems you let it solve and how much you embrace that code is for you to understand and decide on yourself. One way of thinking about it is how much are\nyou willing to let that code execute before it's worth reading. If you're writing some code to just run once or twice on your own machine, who cares? If\nyou're writing code to actually ship to prod, maybe you go the more DHH way where you're treating it like a co-orker that you're working with and reading\nwith. It's also worth noting that DHH just a few months ago didn't think that these things were particularly good. He did a Lex Freedman podcast where he\ntalks about this and said that he doesn't think these are great beyond the AI autocomplete. Actually, he says that that's not even that great overall. But\nnow with tools like open code, he is much more heavily using these things.\nAnother great example comes from this PR that was just filed by anti-res, the creator of Reddus. There's a dependency in Reddus for fast float, which is a C++\ndependency. He is annoyed having this giant pile of C++ in the codebase,\nespecially if you use C++, you know how annoying it is to manage external code.\nIt's rough. He's trying to get more and more of the C++ out. and he replaced this 3,800 line C++ template library with a minimal pure C implementation.\nAnd he has this note here. This code was written by Claude Code using Opus 4.5 and tested carefully, both hand tested and tested against the original\nimplementation in a synthetic way. The code review was independently performed by Codeex GPT 5.2. Isn't that wonderful?\nHe did thoroughly review it with a different model. Yeah. And he posts all the results. It ends up being significantly faster and building faster and having less steps. all good things.\nThis also touches on a topic I'm probably going to do a whole separate video on, which is the idea that libraries, as we know and understand them now, are going to start dying in\nfavor of prompts and specs. So, why is it that these specific senior devs are all embracing these tools, but a handful of other ones aren't? This is where my\nhot takes are going to come in, and we're going to start with Eric's post here. As Eric said, senior devs tend to accept more agent output than juniors do. This is for a handful of reasons.\nThey write higher signal prompts with tighter specs and minimal ambiguity.\nThey decompose work into agent compatible units. They have stronger priors for correctness, making reviews faster and more accurate. And juniors\ngenerate plenty, but they lack the verification heristics to confidently greenlight output. Also, you'll notice here management and executives tend to\naccept quite a bit more, especially on the higher bound here. Staff does even more as well. The amount of code being accepted varies depending on your level.\nLet's talk a bit about why this is the case. Going to ask you a hypothetical question. What makes a junior dev into a\nsenior dev? Just think about this. How would you define the difference between a junior dev and a senior one? I'll let chat participate for this as well. Seen\na lot of pieces here. Mostly people saying experience differences confidence and knowledge. Seniors better at\neverything. One interesting one here is talks less which I don't necessarily agree with. The way I would frame the gap is a combination of two big things.\nCapability and clarity. These are the key skills you level up as you improve as a developer. Your capability goes up\nwhen you can do more work in less time with a better understanding and a lower failure rate. The clarity is really important though as your communication\nskills improve and you get better at talking about the things you're working on and sharing them the right ways with the right people at the right level of\ndetail. The next question, how about the gap from senior to staff? What makes this this one's a lot harder? Some\npeople are saying a title has been employed at Fang. You're now a business manager able to find large scopes making\nmore stuff in parallel. Some people are catching on to it. responsibility.\nInteresting. The ability to ask for pay and title rises. Eh, I would argue it's delegation and orchestration. The harsh\nreality is that one programmer can only do so much. And as you scale up and level up, you realize quickly the cap of\nyour own capability as an individual typing code in a keyboard into a computer. It does not matter how fast a\ntyper you are or how good you are at whatever programming languages and tools you use at a certain point. And if you continue trying to just skill up in your\nindependent individual capability, you will hit a ceiling. Some people have an inherently higher ceiling for that than\nothers. But it is effectively impossible for one really talented person to go as\nfast and as thorough as a team of well organized and orchestrated people being led by somebody with a clear vision. So\nthere's a couple pieces here that I want to focus on. Capability is what most people think of when they think of the level of an engineer. Like the more\nsenior you are in your title, the better you are at writing code. And hypothetically speaking, the more well you'd be able to perform in like a\ncompetitive programming exercise, the faster you could build a solution to a problem independently. I would argue all of these skills are different as you\nlevel up. Here's a lame way of looking at this. If you were to measure these capabilities by level, this is rough\naverages from my experience working with people in the industry. The capability of a fresh out of college junior dev not\ngoing to be super high. Their clarity is in their ability to clearly describe what they're working on, why they're doing it, what problems they're running\ninto, and what is important about the work they're doing also going to be very low because they don't know what details do and don't matter yet. I can't tell\nyou how many times I was reviewing a junior developer's plan for some work and they talk about things like the\nsyntax or the language or whatever. Like none of that matters, especially if you're working in an existing codebase.\nLike I don't need a section on how you expect to format the hooks in your React code. I don't care. Very common thing\npeople do early in their careers. And their ability to delegate and orchestrate is obviously zero. They don't know how to break up work yet because they don't even know what the work they're doing is yet. Senior, you\nstart to see massive jumps in capability and meaningful jumps in clarity. But delegation orchestration usually still aren't strengths yet. But you'll notice\nhere, the senior is already very close to tapped out both on capability and clarity. So getting better at writing\ncode and getting better at writing about the code you're writing does not guarantee you anything. It's not going to get you a promotion. Being better at\nwriting code is not going to be a meaningful improvement because like junior to senior, that's a 3x. senior to staff. Even if they were to be like 10\nout of 10 or an 11 out of 10, you went from 3xing between these two to a 20%\nincrease here at best. It's just not worth it. There's obviously a midle that goes between these, but I don't care. That's what we're here to talk about.\nI'm trying to clarify the differences between these roles. It doesn't really matter how good the staff dev is at just writing code themselves because most\nquality staff devs aren't doing that anymore. They are now much more focused on how the work gets done, not how they\nwill do the work. Ideally, a good staff dev is not diving into the details anywhere near as much. The best feeling\nI had early in my career was when the senior and staff devs on the teams I was working with and around me when I was a junior would have their moment of\nrealization that I knew enough of what I was doing that they could hand things off to me. And a lot of that came from when I went above the expectation for\nclarity in my earlier career stage when I was better at describing what was happening and why and why a specific\nsolution wouldn't work. That built massive confidence out of the staff in principal engineers that I was working with. This seems like a weird tangent to\ngo on in a video about AI software development, right? But what if I told you this skill does not matter anywhere near as much when you're doing AI code?\nAnd these skills matter significantly more. Huh? Turns out being good at clarity when you're describing what you\nwant to build, what's going wrong with it, what problems exist, and how they should be solved, is really useful when you're prompting an agent to do it.\nTurns out delegation, when you're trying to break up big pieces of work into small chunks that can be independently resolved by various parties, be it an\nagent or earlier stage developer, is really useful for agents. and orchestration obviously very key for all\nof this. The harsh reality here is that capability has been overindexed on for a long time in our industry. The amount of\ncode that you can independently write has been this weird flex people love to do and even still to this day people love bragging about their GitHub\ncontribution chart on their GitHub profile because how much code you ship has always been how we measure your\ncompetence as a developer. I've never liked this, even though it was a measure that made me look pretty good because I used to ship a ton of code. Learning that these other pieces in particular,\ndelegation and orchestration are essential to success as you level up and make bigger, more important software was a really, really tough thing for me to\nget over. But I also think this is why I have been enjoying these tools so much because I got over this in 2021 when I\nquit my job, made a startup, and now was blocking myself constantly and blocking my team too. These are skills I needed\nto level up on in order to do the work that I do. And believe it or not, everyone we've just talked about,\nanti-res, Lionus, DHH, Michael, the creator of Effect, all of these people have also really focused in on these\nskills. Many of them still love writing code themselves. But Lionus doesn't write a whole lot of code in the Linux kernel anymore. It's very rare he actually writes a contribution. Usually,\nhe sends an idea to the email list, and somebody who he trusts will go make the changes, email him the patch, and if he likes it, he'll apply it. He's\neffectively the first vibe coder in this sense. He's actually reading the code.\nSo that definition of vibe coding doesn't necessarily match, but the idea being writing down a clear, concise description of what you want and how you\nwant it done, seeing a result, and then bringing it into the codebase. It doesn't matter if it's an AI doing it or if it's a human doing it. These are the\nsame thing. I see people having this realization in chat right here. It's true. People overindex on code don't understand what it means to be an engineer rather than just a programmer.\nYep. Software engineering is moving towards actual engineering. This orchestration super important and that's what's changing. And I don't want to\ncall anybody out in particular, but if you do go look at the people who are really weirdly anti- AI code stuff, I\nwon't say a thing about their capability because a lot of them are very very capable engineers and have built incredible things. But if you look at their history, you look at the work\nthey've done, you look at the ways that they've worked, the job roles they've had, and the teams they've run, if they've run teams, they tend to be\nlacking in these other areas, especially delegation and orchestration. That doesn't mean they're junior devs. It simply means that the skills that make\nyou really capable with these tools are skills that are not their strengths. And if you're a n out of 10 capability engineer where you can write incredible\ncode, but you're not as good at running a team or delegating across employees,\nthis is going to be a lot harder for you. Thankfully, this is how we have mostly done titles in software dev where senior is as high as you can get purely\non your ability to write code and then the rest tends to come from your ability to communicate, operate, and run teams.\nAnd agents are a really cool way to level the skill up early. And here's where we get to my favorite part. You can become a good manager this way. I\nthink this is going to force a ton of developers to get better at things like clarity in their requirements, at\nmanaging work being done in parallel, at rejecting work that isn't quite at the bar. The only thing that it won't teach\nthem is politeness. Because it's a lot easier to tell an agent, \"Your PR sucks.\nI'm closing it.\" than it is to tell a person on your team. But these skills are so important. Just like on a human level, if you want to do more and be\nmore capable, you have to get good at delegation and orchestration. You have to improve your ability to bring others\ninto the work who might not be as good as you are at writing the code. Cuz that's one of the weirdest feelings you'll have the first time you have it.\nI still have the weird feeling here and there when I have work that I know exactly how I would do it that I could go build in a day and instead of doing\nthat, I choose to give it to somebody on the team who will not take a day.\nThey'll take three or four days and the result the first time won't be quite as good as what I did and then I'm going to have to give them a bunch of feedback\nand then wait for the next rendition and this thing that would have taken me a day might take a week or two instead.\nThat feels like a failure. And it feels even worse if you measure your capability and you measure your value by how much code you're shipping. If you\nspend your day in meetings taking tasks that would have taken you a day and breaking them out to take multiple weeks for other engineers, you're going to feel like you're not doing and it's\ngoing to suck. But then one of those things works and it ships and there's a bug and the person who wrote the code goes and fixes that bug. That's when it\nclicks why you do it this way because you can't possibly do all of those one-day tasks and then maintain those\none-day pieces of work over time. But if other people build it in ways they understand, it spreads ownership and\nunderstanding across the team in a way that lets you ship more with more confidence and spread the work around.\nIt's not just the work with a team, it's the ownership, too. And that's the one piece you lose with the AI. Claude Code isn't going to own a mistake that it\nmade. It's not going to understand why there's a problem and fix it for you automatically. You can try to orchestrate tooling to make that kind of happen, but just due to how context\nWindows works, they're not going to remember why they did a thing. So, there are some skills you won't learn this way. But the general idea of letting go\nand realizing your time isn't best spent writing the code in the code file in your editor. It's better spent thinking about the system and how the pieces come\ntogether and making the details less your problem. That is key. And I think the best AI developers are the ones who\nhave realized this. The people who are going the furthest with these tools aren't the people who just learned to code. It's the people who have been\ncoding for a long time and realize these tools are way, way easier and cheaper than hiring a team of 20 people and\ngetting them all onboarded. So, as crappy as it is to reduce things this way, I really think the anti-AI people just haven't run teams well and don't understand this part. And that's fine.\nMost developers don't. Historically,\nthere have been like one person in this column for every 50 people in this column. The majority of devs don't have\nthese skills. That has to change. More and more devs need to level up their ability to clearly explain what they're doing and why they're doing it. More\ndevs need to get better at spreading their work out across people or agents,\norchestrating all of it, and keeping all of it together. We're all managers now. If you don't want to be a manager,\nthat's fine. I don't know how long your job's going to last, but we all need to do this. And the more you do this, and the better you get at this, the more\nthese tools will be useful to you, and the bigger the things you can tackle will become. I'm sure this comment section won't be a disaster. Seriously\nthough, I hope you take this as what it is, an opportunity to skill up your comms and your clarity as you describe the work that you want done rather than\nan attack on your capability of writing software. The future is exciting, but only if you let it be. And I'm really excited about where things are going.\nLet me know what you all think. And until next time, peace nerds.\n"

ERROR: type should be string, got "https://youtu.be/Z9UxjmNF7b0\n\nYou're falling behind. It's time to catch up.\n\nI've never felt this much behind as a programmer. The profession is being dramatically refactored as the bits\ncontributed by the programmers are increasingly sparse in between. I have a sense that I could be 10 times more\npowerful if I just properly string together what has become available over the last year. And a failure to claim\nthe boost feels decidedly like skill issues. There's a new programmable layer of abstraction to master in addition to\nthe usual layers below involving agents, sub aents, their prompts, contexts, memories, modes, permissions, tools,\nplugins, skills, hooks, MCP LSP/Comands, workflows, etc. Okay, I'll cut in here.\nThis is a post from Carpathy that I've been thinking a lot about since seeing it because in some ways I feel very\nsimilar, in other ways I don't. And I have a lot that I want to say about it. The ending in particular is fascinating\nto me. Clearly, some powerful alien tool was handed around, except it comes with\nno manual and everyone has to figure out how to hold it and operate it while the resulting magnitude 9 earthquake is\nrocking the profession. Roll up your sleeves to not fall behind. Yeah,\nthere's a lot going on here and I felt it now more than ever. I still remember\nwhen I did the GPT5 video, I felt the first rumblings of this where it no longer felt like, oh, these things could\nhelp with autocomplete and stubbing out API files for me. Now it feels like AI\ncan build real applications and help maintain and do a lot of this type of work that we have been doing ourselves\nfor decades upon decades. I still can't believe the rate at which this change is\nhappening, but it is very, very real. All of the smartest devs I know are building crazy things with AI now,\neither using the AI as part of the product or for the majority of the dev work. And it's easy to feel like you're\nfalling behind. I'd go as far as to say most of us feel this way right now. But there are ways that we can get ahead.\nThere are things that we can all do as developers to be more powerful as a result rather than less. I think it's\nfinally at the point where it is our responsibility as developers, especially if we want to stay employed, that we\nneed to be on top of these things. So, in this video, I'm going to do my best to explain what I've done to stay on top\nof the latest stuff going on in the AI dev world. And I hope it's useful to you. Do you know what else I hope is\nuseful for you? Today's sponsor. Do you know what's slower than GitHub's loading times? You know, when you're trying to open a PR and it takes 10 seconds to\nrun, there are actions. I'm still just floored at how slow GitHub actions are.\nIt's unbelievable. And that's why is killing it. And that's why GitHub is scared of them. They make your actions\ncomically faster and your Docker builds as well. In just the last week, they have saved developers over 143,000\nhours. Companies like Post Hog saw 37x improvement on their build times from\n150 minutes to 4 minutes. And even open source projects like gRPC saw a 12x\nimprovement in their builds with depot. They don't keep the details secret, by the way. They're able to do this so much\nfaster by using way better CPUs, much better networking, and putting your cache on an NVME drive attached to the\nbox that's actually doing the work. All of this results in crazy differences, like 10 times faster throughput. And by\nthe way, that's all at half the cost of GitHub's runners. It's insane how cheapo is. Trying them out is a oneline change,\nand you'll be so happy you did it. Check them out for free, no credit card required at soy. though. And here we\nhave my best attempt at a guide to not falling behind. Historically, I have\nalways said the same thing about most tools. It's better to be late than early. I still believe this. I want to\nbe very clear. I still very much believe this. I've seen so many people go allin on things that turned out to be useless,\nlike building Java outlets for the Kaioera Echo before the iPhone came out or MCP. There's lots of things like this\nthat happen every couple years where something seems like the obvious, oh my god, that's the future and then it\ndoesn't happen. You need to wait until the thing is actually providing value to people every day before it's worth\nbetting on. The harsh reality is that we are there now with AI. I am writing the majority of my code with AI now. I would\nsay more than the majority, like 90%. And for the teams that I run, we're at at least 70% AI generated code,\nsometimes even more. the companies that I'm advising, that I work with, that I invest in, that I talk to every day are\nat similar numbers. The AI isn't great at making decisions. It's good at talking through them still. That could\nchange in the future. So, you should keep an eye on it accordingly. But the the thing I really want to drive home here is that it is useful. This is no\nlonger will AI pan out, will it become something we use. Coding has changed\nforever. We are past that point now. Getting into it now isn't getting into\nit early anymore. getting into it now is getting into it late. So if you were waiting to see if this stuff would be\nuseful, we are now past that point. It is useful. It's time. And if you think that the stuff you're working on is too\ncomplex for AI, tell it to the people building compilers with it. Tell it to the people who are building languages and systems and crazy applications with\nit. Tell it to the CEO of Railway who's rebuilding their deployment system with it. Tell it to Carpathy who's one of the\nsmartest developers that ever lived that feels like he's falling behind. It's here. No more denying it. And this is\ngoing to affect the job market in meaningful ways. We don't know in what directions or in what amounts, but we\nknow it will affect things. So, we've established that you're officially late. You can cope all you want, and I'm sure\nmy comment section is going to be full of it. I don't care. We're here. So, how do you catch up? There's a lot of ways\nto do this. Personally, my favorite way is to just go try what the hottest tool\nis and push it to its limits. Step one is to try and find those limits. Go check out Claude Code or Cursor or Open\nCode or any of these tools. Use the latest and greatest model which at the point of filming this video is Opus 4.5\nor GPT 5.2x high if you have the patience. Give them a shot. Push them to their limits. Take some random feature\nthat you've been waiting for a while to implement in one of your products or projects and ask the agent to do it.\nOpen up cursor, select opus, and say, \"Hey, I want this planet.\" And then see\nif it writes a good plan and read the output. I'm not here to tell you that\ncode doesn't matter anymore. I'm here to tell you how to use these tools to write code more efficiently. And part of that\nis reading the output. I know you might call me a contrarian if you watch my previous videos about how I've been\nusing cloud code recently, but that's a step we'll get to in a bit. This is how we get in. Try to find the limits and\nread the output as you look for those limits. I also recommend using plan mode\nin the tools that offer it. Open codes plan mode isn't great yet. Chances are by the time you're watching this video, it's probably been fixed. They're moving\nvery fast. Cursor and Claude Code both have great plan modes. Plan mode is great because it's plain English back\nand forth. It feels like sitting in a room with a co-orker in a whiteboard planning out how you're going to do this\nthing. And you can also watch the agent as it goes through the codebase to learn about it and figure things out. And you\nlearn how it does that because that's the the ultimate goal here. Should put\nthat at the top. Actually, if we have more goals as we go through, I will add them here. But these are the two key ones. Building an intuition for what\nthese tools can and can't do and increasing the amount of code that you can output without decreasing your\nconfidence in the quality. So, step one, you have to try and find these limits. And as you're finding the limits, you\nshould also be watching how they do things. I had a list of increasingly complex tasks that I wanted to\neventually do in the T3 chat codebase that I would use as my testing ground. I would keep ramping up how hard they\nwould be. One fun example I would give when I would demo the new models and how good they were at UI is I would ask them\nto build a mock image generation studio to showcase their way of thinking about building UI that way. I no longer ask\nthem to build mocks. I just filmed a video about why I like Convex so much and I had Claude Code use Convex and FAL\nto build a real image gen studio that at one shot. This is a fully working image\ngeneration studio with backend and sync in file storage that will generate whatever you want. Here I'll generate an\nimage of a vibe coding corgi using three laptops at once. And in just a moment once uh we get a response from Nano\nBanana Pro, we now have a vibe coding corgi using a whole bunch of laptops with a save button that works, a UI\nthat's good. All these things that were really annoying before it can just do.\nIt's kind of wild and you need to have these systems in your head like knowing your own code bases, knowing what types\nof work you do and just keeping a mental tab of it. If you have some task that took you a week to do that is relatively\nwell documented, make a clone of your repo from before you did it, put it in a folder somewhere, write down the\ndescription of the task, put it in a markdown file in that folder, and then run cloud code on it. See if it can take\nthat as an input and build the thing. If it gets close but not quite, rewrite the prompt a little bit. See if you can\nguide it there. And if it still can't, save that. Put it in a zip and lock it away somewhere. And the moment the tools\nget better, pull it back out from cold storage and try again. Building your own pseudo benchmarks of what capabilities\ndo these tools have so that you know what they're capable of and how far they can go and also what they get caught up\non is super super valuable. So that's the first step. Take the actual day-to-day work you're doing, break it\nup, box it up, and throw agents at it to see if it can make work that meets your quality bar. This is really good for\nwork you've already done because you know how you would do it, and you can compare and contrast how you did it and how the AI would have done it. Now, I\nhave a slightly harder step two, thinking outside of the box. This one took me a bit to get over, and I'm also\npredisposed to do this already for various reasons, primarily content, but also just the way I work. I can't tell\nyou how often I have a problem that code could solve, but it doesn't make sense\nto write code to solve it. I have a bunch of old assets that I backed up from my Android phone when I was in\ncollege, like just videos and photos that I took when I was in school. They are very poorly organized and just a\nrandom set of folders on my computer. And I wanted to have those saved and upscaled using a tool like Topaz. I\ncould have manually tried to do all of that and I did. And I got through about six months of my eight years of archival\nbefore realizing I can't do this. And I stopped. When I realized how good Cloud Code had gotten, I wanted to test its\nlimits on Windows where I have a lot of this stuff stored. So I spun up Cloud Code on Windows and told it to do it. It\nthen wrote a bunch of very long scripts, including a single 30,000 line of codejs\nfile that it used to reorganize and re-encode all of the files I had stored\non the system. And it did a phenomenal job. That's the type of thing I could have written myself, but I never would\nhave because it would have taken too much time for the immediate benefit it gives me. And I've realized there are\ntons of these. For every project I'm working on right now, I have three or four sub projects that are just me doing\nrandom [ __ ] I want for that project. Part of why I do all the image studio stuff is because I want to overhaul how\nImagin works in T3 Chat. So, I keep spinning up sandboxes to play around with different ideas I have outside of\nthe T3 chat codebase so I can find the UX I like. And some other things I'm working on, like my fish slop game that\nis just a very unique insane aquarium clone. It's not that unique. There's a\nlot of little things I have to do with this game to make it work. Like asset management in particular. I vibecoded\nthree separate tools to work on the assets for this game just to make it easier to track them, organize them,\ngenerate them, screw with them, test different treatments on them, stuff like that. Building a full suite of asset\nmanagement tools for a random game you'll probably never release would be mental illness. Now it's a great\nexperiment. Now it's a way for me to try out new tools, learn, build things that\nwould never have made sense before. Building a 10,000 line of code project to manage assets for a 15,000 line of\ncode project is the type of thing that made no sense before. But this is why you have to rewire your brain a bit. You\nhave to think out of the box. You have to think about things that could be solved with code that would have made no sense to solve with code before because\nit's too much code, which is too much effort. The amount of effort it takes to write code has gone down a ton. And once\nyou understand the limits and capabilities of these systems, you start to see the world a little bit different.\nI'm going to do a really weird comparison. When you look at this image,\nwhat do you see? 99.5% of you are going to say a stair set. Some of you are\ngoing to say the El Toro 20. This stair set is very important in skateboarding.\nIt has a legacy that's hard to put into words. There's a reason for this. It's because it is gigantic and jumping down\nit is kind of insane to do. And many skaters have done that. Many have gotten very hurt trying to do that.\nSkateboarders see the world a little bit different. Once you've rode a skateboard for a certain amount of time, you no\nlonger look at the street the same way because you're thinking about how it feels to ride your board on the pavement. You no longer look at stairs\nthe same way. You're counting them to see how many there are and how hard it would be to jump down it. You don't look at ledges or rails or benches or picnic\ntables even the same way. It's rare that I walk into a new room or walk down a new street and don't think about how I\ncould skateboard on it. Once you learn the skateboard, you see the world a\nlittle bit different. Everything is now kind of an obstacle. And when you learn to code, the same thing happens. When\nyou go to a website or open an app or see the weird errors on the screen at the airport, you're not seeing the same\nthing everyone else does. You're thinking it through a little further because you know more about how it works. The way you perceive these things\nis different. When you get an error in an app, you're thinking about what led to that. When someone else gets an error\nin an app, they get annoyed. You have to do that again. Now, I'm sorry, cuz one\nof the hardest things you can do to your brain is rewire it. So, the world looks a little different, but you have to now.\nAnd to other creators, other developers, other people who are watching this video, it sucks and it's hard. And it\ntook me a while. It took me like over a year and a half of using these things every day to have this click. You can\njust build [ __ ] If you have a random idea, it doesn't take 3 days to build it\nanymore. It takes a couple minutes. When I wanted to compare how different AI\nmodels write pros, I had this novel idea. What if I could have a model write\nan essay, another model give feedback on it, and then the original one could rewrite the essay? I wanted to automate\nthis because I was just curious. It took me about 10 minutes of vibe coding and cursor to build up a workspace that\ncould do that. Now I can answer these questions. I can figure these things out. I can do [ __ ] that just never would\nhave been worth doing before. And that's what's so crazy. There's this whole world of things that we encounter every\nday as developers that we didn't bother dealing with before because it was easier to just do it by hand than it was\nto automate the thing. There are so many of these little things that were just obnoxious that are now trivial. If I\nwant to play with a new API, I throw the docs at claude and say, \"What can I do?\" If I want to build a new Chrome extension, I open up cursor and say,\n\"Hey, make this with me.\" It's so freeing to realize any problem that can\nbe solved with software just got a 100 times more solvable. And that isn't just yet another new vibecoded app or\nproject. That could be something as annoying as I have this set of commands I enter over and over. It'd be nice if\nit was one command here. I'll give an example. You know, I I'll even voice the text just to make the point. Make me an\nalias that selects all files and commits them with a commit message and pushes up\nimmediately. Now, it's going to make me a git alias in my Zish RC. Almost certainly didn't do it, but apparently I\ncan just configure this with a git global directly. Get ACP. Let's tell it\nto do something a bit different. I would like it to be just one word. Can you maybe put this in my Zish RC instead?\nThere we go. Now I have this new command added. This type of thing that just wasn't necessarily worth the time for\nmost people is now trivial. You can see I even did this earlier where I wanted to always YTDL things as an MP4. So here\nit will just auto postprocess. I can use this custom command for YouTube DLP and it will come out as an MP4. I have so\nmany of these appearing in my Zish RC because whenever I have a thing I do a lot. There we go. Once you start\nthinking this way, you're most of the way here. But the next step is where things get really fun. Okay, to be fair,\nstep two is actually really fun. But the next piece is the orchestration. And I'll be very honest here. I'm still\nlearning this part. This is the part that things really start to accelerate. And I'm at the point where I kind of\nwant to rethink my whole operating system around what capabilities exist here. figuring out how to spin up\ndifferent agents, how to link pieces together, how to write glue code that you're not even really writing yourself\nthat lets you attach these cool tools and pieces and figuring out all the parts. Things like when I was working on\nmy game, I found a tool for generating pixel art that was okay, not great, but\nokay. And orchestrating between that and the game and all the different features that I want to work on and getting this\nall pulled together was challenging. One of the pieces I came up with, and it's really silly, the game is the fish game.\nSo, in order to keep track of all of the different fish in the game, I made the fish bible. The fish bible is a markdown\nfile that describes all of the different fish and pets in the game. And whenever a change happens in the game, this file\nshould be updated. Whenever a change happens in this file, the game should be updated. I enforce that via the claude\nMD where I specify the fish bible file is the authoritative reference for all creatures, fish, pets, aliens in the\ngame. It contains all of these details and it needs to be kept in sync. This is in my claude MD. Whenever I'm using\nclaude code, it will do its best to honor this. This also means I'm not looking at the code as much, but that's\nthe point of this project. Having a a range of how much you do and don't care about code in a project is essential.\nAnd part of this thinking outside of the box piece is that there are a lot of things you might want to do in your project that aren't worth the time to\nwrite or even read the code. But if these are just things like setup scripts or automations for things that you were\ndoing by hand before, if that code isn't actually running on production servers, you're just using it to benefit yourself\nin your life, your bar for the quality will be different, and it should be. Figuring out where in your life you can\ninsert a bunch of slop code will give you a great opportunity to play with these things more. And I promise you,\neverybody watching this video has places in their life where a little bit of slop could actually be useful. I've been\namazed at how many places we're finding for this in my own work. Even just like building a tool to keep track of all of\nmy thumbnails and which types of thumbnails tend to get the most views. Visualizing that type of thing was\nannoying. Now I can build it in a day or less. But again, the orchestration is the hard part. How do you keep track of and organize and combine these things to\nbuild awesome [ __ ] There's a lot of tools being built to do this and a lot of independent people building their own things around this. Claudebot is a great\nexample of this. Peter's been doing all sorts of crazy [ __ ] and Claudebot is one of his attempts to pull it all together\nto make all of his tools accessible to an AI agent that he can talk to via Telegram or WhatsApp that's controlling\na computer. It's so cool. Somebody said in chat that I'm finding places for slot cuz I'm forcing it. You have to force it\ninitially and then it clicks. That's the key. You realize how much random [ __ ] is\ntedious in your life that you could automate that was just not worth automating before and now suddenly\nabsolutely is. There are so many things like that. I am still amazed just how many there are and how many more I'm\nfinding. I'm keeping a list of all of the ideas I have. Like one that I really want to start working on is a canban\nboard for managing tasks that I want to generate code for between the planning stage, the implementation stage, and the\nreview stage. And there are so many opportunities like this and at real companies too. A big part of what inspired this video is this post from\nRaul who is the head of applied AI at RAMP. Ramp's a real company doing real\nchallenging work and they're doing some super cool stuff with AI especially with the orchestration of tools like open\ncode. This is him quote tweeting the original Carpathy post that started the video. He makes some bold statements at the top here. You're guaranteed to lose\nif you fall behind. The no unforced error AI leader playbook is the following. So these are great ideas on\nhow to start getting ahead for your team. Use coding agents. Give all engineers their pick of harnesses,\nmodels, and background agents. Cloud code, cursor, Devon with closed and open models. Meta engineers used to be forced\nto use Llama 4, which is hilarious. They no longer are. They can use whatever they want. Now open 4.5 is the baseline\nnow. That's probably going to change very soon. This video hopefully won't get too dated beyond that. We will see.\nI will be sure to make new videos if that does happen. By the way, if you watch this much, you should hit the sub button if you want to keep up. It's a\npretty good way to help and uh helps the channel out a ton too. Next piece, give your agents tools to all dev tooling.\nThings like linear, GitHub, Data Dog, Sentry, any internal tooling. If agents are being held back because of the lack\nof context, that's your fault. Very bold. Still coming around to this one\nmyself, but I've seen so many compelling examples that it's hard to ignore. Here's an example of an engineer at RAMP\nwho was demoing one of the cool bots they built, their internal inspect bot. They added it, what are the 20 most\ncommon sentry issues in core, spin up child sessions fixing them. So it found\nthe 20 most common errors and then spun up a PR for every single one of them.\nThat's insane. And this isn't some tool they're selling. This isn't some crazy product they want to put out and say, \"Hey, everyone, go install inspect.\"\nThis is just their own internal explorations and usage because doing this is now way more justifiable than\never because it's not that expensive time-wise. Another big piece, invest in\ncodebase specific agent docs. Stop saying that it doesn't do X well. If that's an issue, try better prompting\nand agents MD file, linting, and code rules. These are actually really big pieces. Agents get a lot smarter when\nthey get feedback on what's wrong. Set up your LSP and use it. Please, for the love of God, use a typed language if\nyou're not yet, because then when the code has errors that are checkable via types, you can feed those errors back to\nthe agent and it can fix them. Open code does this by default. Cloud code added it. It's not fully supported just yet.\nCursor has this working really well by default. When you have good linting and code rules and type safety, all of a\nsudden these agents can fix a lot of the errors that they would have on that first pass and they will come to you\nwith a result that is more likely to be working initially. And then the agent MD in particular is a really important\npiece too. I know the Claude code team apparently changes their Claude MD file for the internal repo multiple times a\nday. If someone on the team notices that Claude went down the wrong path for something, they don't hope it won't the\nnext time, they go to the CloudMD file and add instructions saying, \"Hey, don't do that.\" Good way of thinking of this\nis that every manual edit that you make from generated code is an opportunity for agent MD improvements. Next piece,\ninvest in robust background agent infra. Get a full development stack working on VMs and sandboxes. It's hard to set up,\nbut it'll be worth it. Your engineers can now run multiple in parallel. Code review will be the bottleneck soon.\nThat's another thing I would actually add. Maybe as like a step zero. If you're really struggling to like get\ninto this, at the absolute least, please add some AI code review tools to your\nrepos. There's a bunch of them. There's controversy around some of them. I personally mostly use Grapile and Code\nRabbit. I use Grapile on all my personal work and side projects. The team still slightly prefers Code Rabbit for T3\nchat, but they're both really good. There's lots of other options that are pretty good as well. These things are very helpful to get into. It's a great\nway to see the value of AI really quickly. Yeah, code review will be the bottleneck, but please use AI code\nreview too, not as the only code review, just to augment it. It's it's so nice to have an AI catch a mistake before a\nhuman has to bother reviewing it. Also, important call out, figure out the security issues. Stop being riskaverse\nand do what is needed to unblock the access. Absolutely agree here. And even with earlier career engineers, I would\nsay the same thing like I would have so much worry like oh if we give these people access to the deployments they\ncan break all of prod.\nAnd now when you're actually building product with AI important pieces, always use the latest generation models in your\nfeatures. move things off of last gen models ASAP unless robust evals indicate otherwise. It's very rare unless you're\nlike on Gemini or something that the generation change is actually a downgrade. You'll have to make changes every couple weeks. Like the GitHub\nCopilot mobile app is still offering code review with GPT 4.1 and Sonnet 3.5.\nThat's hilarious. You're leaving money on the table by being on Sonnet 4 or GPT40. Absolutely. Stay on top of these\nthings. Use embedding semantic search instead of fuzzy search. Any general embedding model will do better than\ntraditional fuzzy search heristics. Yes, like search is one of those big pieces.\nA lot of these harnesses now, things like cursor and claude code aren't using traditional search anymore. They tell\nthe model they are so the model doesn't have to learn anything new. And then they quietly behind the scenes use better search semantics to get better\nresults. Leave no form unfilled. Use structured outputs in whatever context you have on the user to do a better best\neffort prefill. That's a cool idea. I should consider that more for things. Allow unstructured inputs on all product\nsurfaces. You must accept free form text and documents. Forbes are dead. Also very interesting. Another one that I\ntotally agree on is that custom fine-tuning is dead. Stop wasting time on it. The frontier is moving too fast\nto invest 8 weeks into a fine tune. Costs are dropping too quickly for price to matter. Better prompting will take\nyou very far. This is only getting more true as instruction following improves. Really want to emphasize this point to\ngo back to my list here. When you find a limit, when you write a prompt for the model and it can't do the thing, try to\nget past it. Strategies for this are improving your prompt, giving more context in particular in the prompt,\nadjusting claude MD/ aents MD just to give better resources on the codebase\nand how you want things to work. Adding tools to give better feedback to model.\nThis is things like LSP support, linting, etc. See if you can make the\nmodel go past the limit when you find it. I have been incredibly surprised by how much you can do by just giving some\nslightly better tools and adjusting your claim MD file. The exact same prompt on the exact same codebase could have\nbetter results. There are some people who swear by the strategy of never reverting and just prompting until it\nworks. I know Pete, for example, the guy who made QuadBot, swears by the strategy. Yeah. As he says here, he\nbasically never reverts or uses checkpointing. If something isn't how he likes it, he asks the model to change it. I am not there yet. Personally, I\nstill like to revert mostly so I can learn the structure to get these things\nright first try. But this varies from person to person. Work on it, play with it, experiment. You really should try\npushing the limits of not just these tools, but yourself. It's going to feel weird and uncomfortable. If you're not\nseeing the value, you're not there yet. Keep pushing until you do. So many incredibly talented devs are doing\nabsurd [ __ ] with these tools. You can get there, but play with it more. Speaking of really smart people, Nean is\nin chat. Nean is the creator of Stylex, which is one of the coolest styling libraries I've ever seen that was used\nto handle styling at meta at scale. Library is super cool. Nean is super smart. Nean just said in chat that the\nrevert thing is kind of model specific. Codex is much better at unfucking a bad change it makes and Opus is less good at\nthat and benefits from the reverts. This is knowledge you get from experimenting with these things and playing with them more. You got to just do it. Get in the\nweeds. Get your hands dirty. Get uncomfortable. If you're not at least a little bit uncomfortable, you're not\ntrying hard enough. It's a harsh reality, but it's one that we're in now. You need to push past these limits in\nyour head and in your work. More fun stuff from Raul in this post because there's more stuff here I want to talk\nabout. Building evals is essential. You should build lots of evals to make quick model upgrade decisions. They don't need\nto be perfect, but at least need to allow you to compare models relative to each other. Most decisions become clear\non a Pareto cost versus benchmark perf plot. Yep. Yep. It is so easy to vibe\ncode a benchmark. It's one of the best things to start with. If you notice something that AI is or isn't doing well\nfor you and your work, break that off into a bench. Make a quick weird oneoff benchmark. It's so fun customizing and\nbuilding your own evals. I've had a ton of fun with it. I was going to put out tools to make it easier for others to do, but I don't feel like I need to\nanymore. Just vibe code it. Tell it to use the ASDK and open router and go ham. The last piece here, and this is\nessential for leaders at companies. I'm trying to do this the same myself. Encourage all engineers to build with\nAI. Build primitives to call models from all code bases, structured output, semantic similarity endpoints, sandbox\ncode execution, etc. So much good [ __ ] One last piece. This is one I'm really internalizing. Stop worrying so much\nabout inference spend. Your inference profit and loss needs to be significantly higher. Costs are dropping\ntoo quickly for this to matter. It's not x mill a year. It's x divided by 52 per week for the next y weeks. Yep. If your\ninference bill is $3,000 a week this week, it is as likely to go down as up\nweek over week as new things ship and change and behave differently. Things are changing really, really, really\nfast. And you should be going out of your way to try all these new things. Now, I know a lot of this was focused on\nleaders and not everybody watching, in fact, the majority of people watching probably aren't managers. You're\nprobably IC's, aspiring devs, or people just working independently on code. You\nmight not be able to do all of these things if you don't have that type of buyin at your workplace. Find every opportunity you can to do it\nindependently. Try sneaking it into your workplace if you can. And if that doesn't go well, find a better job. I\nknow that Ramp is hiring, for example. RA actually asked me to shout that out. So, shout out Ro for writing this. You\nguys should check him out and talk to him if you're interested. The link will be in the description for this post. Find a place where you can do this. You\ncan make the place yourself independently on side projects. You can find a way to sneak it into your workplace if they allow for that type of\nthing. Generally, like now more than ever, ask forgiveness, not permission feels almost essential. Like if your\nworkplace doesn't let you use these tools, use them anyways. Either you will now be way ahead of your co-workers and\nbe seen as an evangelist at the company or you'll get fired for it and you have an incredible story to tell in your job\ninterviews for other places. I promise you if you went to somebody like Raul and said, \"Hey man, I just tried all the\nthings in your post about how to like bring AI to the workplace. I got fired because nobody wanted to see the light.\nCan we talk about what working at RAMP would look like?\" I guarantee you that'll be a good conversation. Push the\nlimits in every sense. That is really the theme I want to drive home here. Find the limits of how much you can use\nthe tools of what the tools are capable of and what the people around you in your workplace is willing to allow. Push\nas hard against those limits as you can and see where it brings you. You will be\namazed first off at how far away those limits are, how capable these systems are, how well you can work around those\nlimits, and how much [ __ ] you can get done as a result of all of this. And if your manager is not convinced, send them\nthis section. Hey, you manager at tech company X. It's so important to let your\nengineers use these things. The best engineers have already made the move to using AI tools to help with the majority\nof their work for the majority of the time. And if you're not letting your employees do this, you are intentionally\nletting them fall behind and they are going to leave and your company's going to fail. Standing against AI in the development world isn't some crazy noble\nstance or some way to protect your users. It's a way to piss off the best devs at your company and keep them from winning. If you want to win, you need to\nlet your team use the best tools. And right now, most of those tools include AI. Get over your [ __ ] Let your\nemployees use the things they want to, and you'll be amazed at how much more productive they become. I think I've said all I have to here. This was a very\nfun deep dive on how to stay ahead. I know what the top comments are all going to be. Hey Theo, I'm a student. How do I\nkeep up? I have no idea. I am thankful I'm not a student at this point in time.\nIf I figure out an answer, I will be sure to do a video about that in the future. But for now, keep building, keep\nshipping, keep doing things, and keep pushing limits, and you'll probably end up somewhere great. That's all I got for this one. Good luck keeping up. And\nuntil next time, peace nerds.\n"

ERROR: type should be string, got "https://youtu.be/gh6aFBnwQj4\n\nAnthropic just burned so much trust...\n\nI don't even know where to start with this one. Anthropics officially started cutting people off for using their Claude code plan with things that aren't Cloud Code itself. On one hand, I saw this coming for a while, but on the other, it's just kind of a crappy move to play. And I Yeah, this sucks. If you're not already familiar, Enthropic put out these plans for Claude Code when people started using it heavily because their API prices are absurd. So, you pay a flat monthly fee of $100 for the 5x plan or $200 for the 20x plan and get way more usage. The amount of usage you get is nuts. I wrote some scripts to keep track of how many tokens I was using and what it would cost over API usage. And it's easy to do thousands of dollars of API usage through these plans because Enthropic subsidizes them with their other sources of income. People using the plans and not paying and using the full value of it, etc. Since these plans have such generous limits, a lot of thirdparty builders have been implementing them in their own services. things like Open Code, things like Cladbot, which we'll talk about a bunch in a bit, all have been using your O through Cloud Code to take advantage of the token limits you're getting. So, you're still paying anthropic the money. You're just using the API key effectively in other apps. At least you were until they locked it down. Anybody who is using their Cloud Code plan through apps like Open Code saw this error yesterday. This credential is only authorized for use with Claude Code and cannot be used for other API requests. Oh boy. I think this is a massive misplay that's going to keep burning the little bit of good faith Anthropic still has in the developer community. And I I can't believe they did it. Honestly, I have a lot to say about this, but I also am going to have much higher bills if I cancel my Anthropic plan here and use other apps. As such, we're going to do a quick word for today's sponsor. We're getting tired of hopping between tools all day. It's just frustrating. Maybe I have an idea for a feature, so I kick it off using Devon or Codeex on my phone in the chat GPT app. Then I go to my computer to see how it came out and I see it's mostly working, but there's a few things missing. So then I have to grab the branch, pull it into my machine, spin up a different coding agent, lose all of the context and everything else going on and hope that I can actually iterate. Imagine a world where it's all in sync. The job you started on your phone can be opened in your editor, your terminal, or the web. All however you choose, whenever you want. You can kick off a job in Slack and then finish it in VS Code. Wouldn't that be great? You'd probably have to give up a lot of control though, right? Not if you're using today's sponsor, Kilo Code. fully open source, access to every model as soon as it drops because they literally just let you bring your own with open router is how I test new models already. But now it's going to be a lot more than that because the ability to kick off a job on my phone and finish it on my computer is just like mind-bogglingly cool. As they say, it doesn't matter where you're using their tools, VS Code, Jetrain, CLI, or kicking off the job via Slack. It doesn't care. It will keep your history and your context all in sync so that you can access it across all of these different platforms. fully open source, so if there's anything you don't like, you can change it yourself, see how it works, and do what you need. If you're a big company trying to audit these things, it's a lot easier when they're fully open. And they even have a Slack integration in early access, which again, very excited about. Hey, uh, friends at Kilo, can we get this? Thanks, guys. And they have so much other fun stuff coming like they just published their road map publicly on the site. Again, really open source down to like the roots of the company. They have an iOS and Android app coming, a full agentic engineering course and certification. Bring your own key centralized so that all employees, all people on the same team have access to the same keys, as well as the same shared MCP servers, prompt libraries, and more. This makes a ton of sense because of how much you can do with their tools. They provide you with five default agent configurations, and you can add your own yourself. You can specify which models and which tools each of these has access to, and it makes it so much easier to control the way that you're using your AI. You're trying to get deeper in these tools, understand different models, and really control how they access your codebase and do things. There's no better place to start than Kilo. Check them out now at get $13.37 of credit if you use code Theo. Where do we even begin with this? I heard about this initially because Peter, who created Cloudbot, posted the following error. I think he was one of the first people to see it because he was actively working on Claudebot when this happened. The error you saw was the LLM request rejected, the credentials only authorized for use with Cloud Code and cannot be used for other API requests. You're not familiar with Cloudbot. I've been playing with it a whole bunch. I actually have a Mac Mini set up in the corner here just to run Claudebot. So I can send messages via Telegram to this bot on my computer that runs Claude Coder, in my case, Codeex, to do different things on my machine. It is super super cool, but also uses a ton of tokens. Believe it or not, I actually think Claudebot is a significant portion of why this happened, even more so than something like Open Code, because it's very tokenhungry. It burns through tokens unlike almost anything else I've used before. It's super powerful, but it's very token hungry. And these plans work largely because you expect a large number of the users to not hit the usage limits. Like, I've been paying the $200 a month for OpenAI's $200 tier for a while now and have certainly not done $200 a month of inference in it. Apparently, the average person on that plan does cost OpenAI money, like they use more than they pay for. But regardless, the way that this works is economies of scale. I'll get into the economics and some numbers in a sec. But first, I want to make sure anybody who thinks that this is an accident or unintentional realizes the truth here. Another open-source CLI for doing code stuff got hit as well. And they actually had in the PR where they removed cloud code support this comment. The patch removes cloud code support following a request from Anthropic to align with their terms of service. This is very Anthropic. They do a lot of stuff like this. They're actually the single company that sent the most DMCAs out on GitHub because Anthropic was upset that they left the source code in through a source map in Cloud Code, which is closed source. By the way, reminder, just because there's a GitHub repo with a lot of stars does not mean Cloud Code is open source. That is for a handful of pieces in community management. Cloud code itself is still closed source. They accidentally shipped source maps. People published the source maps that they published themselves and got DMCA for it. And countless DMCA requests have been sent by Anthropic on GitHub for this reason. They are not scared to hit people up and tell them that they're in violation of whatever and suppress what they're doing. It's almost Applelike in that way. This is just kind of how Anthropic operates. They want things done their way. They don't want you doing things your way. They want you to use the systems that they put in place to do things. They don't want you using other SDKs. They want you using theirs. They don't want you using other APIs. They want you using theirs. They don't want you using other CLI tools. They want you using theirs. Before we get into the long-term plans they have, I want to talk about the economics here. Let's say you have multiple different subtiers. You have the $20 a month tier, you have the $100 a month tier, and you have the $200 a month tier. You have these different tiers. Let's say the theoretical max that you can use in the $20 a month tier is $30. For the sake of making this digestible, let's say the $20 month here you could use $30 max. They said it's 5x more for the $100 plan. So that would be $150 can be used on the $100 plan. And they said 20x more for the $200 a month. That's 20x more than the 30. So that's 600. That's wrong because I know you could use significantly more than that, but we will just go with it. So, let's say this is the maximum you can use for each of these. Generally speaking, people on the cheaper plan are not going to come close to their limits. They're going to use significantly less. I have these numbers for T3, and I would be surprised if Anthropics were significantly worse in terms of the split. My guess is the average $20 a month users probably using $8 a month of usage average. And as you get to the higher tiers, the percentage that they're using is going to get higher. So the $100 a month tier is probably going to be closer to using $70 a month average. And we'll do the bad numbers. We'll assume the average $200 a month tier is using $300 a month average. So how do these economics work out? Let's just do easy numbers. We'll say we have a,000 users on this tier. We'll say we have a 100 users on this tier. And we'll say we have the same 100 on this tier. I'm too lazy to do all of this math. So, I'm going to use a better chat app, T3 chat, which to any model pretty much, we'll use I don't know, just for fun, we'll use Sonnet. Do the math on these. Turn off search. We don't need it. So, for the,000 users at 20 a month, averages $20,000 in, $8,000 of cost, $12,000 a month profit. The tier two users, $7,000 a month cost, $10,000 a month revenue, $3,000 of profit. And then the tier three users cost $10,000 a month. So the total about 15 a month profit, 10 a month loss. You still make money. You're able to subsidize the cost of your heaviest users by going after the cheapest users. And this is a very, very common pattern. This is a big part of why these expensive tiers were introduced because there are these power users. People like the ones you see on Twitter, people like the super smart devs that are just maxing out all of these things. People like Ben and me and all of these crazy hackers that love abusing these plans. They couldn't get by on the $20 a month plan and they would cost thousands of dollars of API usage, which we can't justify. The point of the $200 a month plan is that it is a loss leader that also helps a lot with marketing. This is the thing I think a lot of people are missing. The point of this $200 a month tier isn't to make money for Anthropic. The point of this tier is to be marketing for Anthropic. This is here as a marketing expense that is subsidized by the money that has been invested in them as well as the money being spent on people on these cheaper tiers as well as people like me building services like T3 chat that are paying the full API prices. This tier is so people like me will do free marketing for anthropic and it's clearly working. If we just look at the success of my recent videos about this stuff, specifically my I'm addicted to Claude Code video. 122,000 plays in 3 to 4 days is really, really good for my channel. And it seems like people liked the video a lot. It would cost a lot of money to get 122,000 devs to hear about your thing and see it in such a positive light. They got that for way cheaper by subsidizing my $400 of usage by $200. The point of this is to be a marketing expense to funnel people into Anthropics codled walled garden so that they will continue spending money on anthropic stuff. So they'll use the APIs and spend the money. So they'll have more of their company sign up that might not use as much. They benefit a lot from this as a marketing spend to funnel down the rest. But this is also the problem. If the point of this tier is to get more people to see and have visibility on anthropic product, using this alongside open code or cloudbot or these other things that are nowhere near as directly tied to anthropic does not benefit them in the ways they want. They're willing to take the loss on these tiers if it increases the likelihood that the average developer is locked into anthropic stuff. They are purely doing this as a market spend, which means if it's not succeeding in their goal there, it isn't succeeding at all. The other issue is for smaller businesses or companies that don't have the ability to subsidize in the same way. Everybody from us with T3 chat to companies like Open Code and even more so stuff like Cursor, they can't make the inference cheaper because they don't have access to the models. With openw weight models, you can go buy some GPUs and spend some money up front to have more inference that you can then use alongside other things. But when the models are closed, you have to pay effectively whatever anthropic decides. And if Anthropic decides you can't pay, they could do that, too. In fact, XAI just got cut off, which is crazy. I have talked a lot of crap about the way Anthropic cuts people off from access to their stuff. Historically, people have pushed back on me saying, \"Well, obviously Open AAI can't have access to anthropic models. They're going to use the data to train or obviously Windsurf can't have access if they're going to be bought by OpenAI. They're going to use it to train.\" XAI wasn't using the models for training. They were using them for dev work internally through cursor and anthropic cut them off. All XAI employees using Cursor no longer have access to anthropic models. That's absurd. But this is kind of how anthropic works. They want you to use their stuff the way that they intend. And if you use it in a way that they don't intend or as a person that they do not like, they cut you off. And this has been their pattern for a long time now. from DMCA and GitHub to rug pulling multiple companies using their models for traditional work or even just evals to this type of [ __ ] to now cutting off everybody for using their sub plans for other things. It just sucks. It really does. Here's Tar statement who is an employee at Anthropic working on cloud code. He's kind of their like devi. Yesterday we tightened our safeguards against spoofing the cla code harness after accounts were banned for triggering abuse filters from third party harnesses using cloud subscriptions. They weren't really spoofing the cloud code harness. They were getting an O token and then using it. But sure, third party harnesses using cloud code subscriptions create problems for users and are prohibited by our terms of service. They generate unusual traffic patterns without any of the usual telemetry that the cloud code harness provides, making it really hard for us to help debug when they have questions about rate limit usage or account bands and they don't have any other avenues for this support. The rate limit usage problem is because you optus skate as much data as you can in it and the account bans are bad because you're banning people as soon as you detect things. I think this is such a [ __ ] response here. This is absurd. Tar, you're a friend. I appreciate you immensely, but you shouldn't let the legal team write these things. This is awful. You probably should have just not posted. This is why the supported way to use Claude and your own tools is via the API. We genuinely want people building on Claude, including other coding agents and harnesses. and we know developers have broad preferences for different tool ergonomics. If you're a maintainer of a third-party tool and want to chat about integration paths, my DMs are open. We've also heard it wasn't clear enough to end users that this was a terms of service violation and that's on us. We'll make it clearer in the OOTH screen going forward. We've lifted all the bans that we're aware of that were caused by this issue. Please DM me or email if you were banned due to this and haven't been reinstated. That's one nice thing at the end. But this, no, this is bad. The problem here is that it is anti-competitive in the most literal sense. If you can get $600 to $4,000 of usage of Opus for 200 bucks via quad code, or you could use a better harness, a thing that you prefer, something like open code or cursor or something else, but the $200 gets you $200 of usage instead of $2,000 to $4,000 of usage. You don't use the better thing. It does not matter how much better a product you build. Enthropic by having a better model can force you to use a worse product by charging you more for using better products. It's the most traditional lock in play I've seen in a long time. This is more transparent and more egregious than even like Blue Bubbles. It's absurd. DAX has a lot to say about this as you would expect. It's their business. They have the right to enforce their terms however they like. One note on that though before I go further on what he had to say. It's actually not illegal at all. If you find a novel way to work around these types of limits, that is fully legal within US law. There have been previous cases, I forgot which one, but I I've read a previous case where it was ruled outright that working around these things is totally fine, but it's just a cat-and- mouse game. They can try to restrict it. They can try and ban people. They can do whatever they want, but we can too by consuming the API in the ways that we choose. The terms of service can be whatever they want and they can ban you for it, but it's not illegal to do any of these things. They can't go after you with legal ease. They can only go after you by banning you. Anyways, as Dak said, it's their business. They have the right to enforce their terms however they like. They're not obligated to provide completely open access to their services. But this is a hint at an underlying problem. Models aren't sticky, so it's not good enough that you're using Opus. They need you and your team to adopt their full stack of tooling so that it's hard to switch to a better model. This is the key. New great models come out constantly. It's honestly amazing. We haven't had something better than Opus yet, considering how fast the stuff tends to move. I would expect to see some big releases from OpenAI in the next couple weeks at latest. The back and forth is going to continue as it always does. And if you get used to a tool like Open Code and you can just swap models with a command, Enthropic goes from making however much money they are off of Opus to nothing really fast. The way that these things go up and down is absurd. I've seen it in our own data. A model like Gemini 3 Pro can go from being 30 to 40% of our usage on the pro tier to like 5 to 10% in literally two to three days. It's nuts. So, Anthropics going at all angles to lock people in right now, but as Dax said, thankfully not all LLM providers want to operate this way. You'll hear more about that soon and hear more about that we certainly did. We are working with OpenAI to allow Codex users to benefit from their subscriptions directly within Open Code. Huge. OpenAI has been very collaborative, more so than I ever would have imagined. Like sure, their name's Open AI, but they've actually been really good about this stuff in general. They open sourced their CLI immediately. I still have my issues with Codeex, but it's good. It's not great. It's good. But you can use your subscriptions directly within Open Code, and they are working to make sure Open Code stays whitelisted as an official thing that you can work with. It was already possible before, but now it's going to be like made official, which is huge. And if you're wondering about Google and Gemini subscriptions, Google is too. They do not understand any of this well enough to make a statement, much less block you, so you're probably fine. But also, Gemini 3 is not a very good model to use for day-to-day work. Complain in the comments all you want. Nobody is seriously using that model for real work right now, even Google employees that I talk to. Anyways, as always, Open Code ships really fast. The Open Code OpenAI collab was done before I even finished streaming. So now you can use /connect in the latest version of Open Code to connect your ChatGBT Pro or Plus subscription plans to Open Code. That's the way to do it. And Adam's in chat. That that tells you how fast and deep these guys are. Love it. All of that was bad. But here's where things get really really dirty in my opinion. You might not know this, but Anthropic ships their own agent SDK. You can think of this as somewhat similar to like the AISDK that Verscell publishes with the obvious difference that it only works with anthropic models because of course it does. Also, unlike the Verscell AI SDK, it's fully closed source. So, you cannot see the source code for this. That makes it much harder to use because models are much better when they have access to the source code and can see how a thing works. I'm sure that they have claude.mmd and llm.exted texted their way to something mostly usable. But it's a lot better when you have source code access, which they do not give you here. But where it gets much dirtier is that you can authenticate this with your cloud code sub. So if you build an application around anthropics closed source agents SDK and then you put this application out for people to use, they can sign in with their Anthropic sub and use the subsidized tokens. That is so dirty. This means that theoretically if open code was to rebuild the entire platform of what they've built around the closed source anthropic SDK, you would be totally within terms of service to use your existing O because they don't want you to have access to other things. They don't want you building at an abstraction level that allows you to swap off of Anthropic and they will gladly subsidize your costs as a marketing expense and as a lockin expense to force you onto their platform. This is just the diapers.com play. If you're not already familiar, this was Amazon destroying a company specifically around selling things for new parents. Amazon acquired the parent company of diapers.com for 545 million. They had been trying to acquire them before, but the price was too high cuz they were doing too much revenue. So their solution was to undercut by selling diapers at a steep loss, literally intentionally losing money to force the company to lose all of its market share so they could then buy them out and not have the competition at all anymore. That's what's happening here. Okay. Apparently they don't even want you to use the cloudi login or rate limits for things built on the cloud agent SDK. They claim you can off with it but then do that right after. Literally right here they say if you've already authenticated cloud code by running cloud in your terminal the SDK uses that off automatically otherwise you need an API key but then they say here that although I haven't heard of anybody being booted that was using the agents SDK but I also don't know anybody using the agent SDK so hard to know. Yeah, this is a mess. I hate this. I want to support anthropic. I have a lot of friends there. I have as many friends if not more so at Anthropic than I do at OpenAI. I really like the model. Hell, I even use cloud code. I'm still not somebody who's made it all the way to open code. I'm mostly in cloud code still, but I still hate this. I'm still considering canceling my sub at $200 a month, and I would encourage you rethink yours as well. This is the exact type of dirty play that has had me skeptical of this company for so long. I Yeah, I just hate this. It makes sense, though. It's the same company that cuts off access randomly, that DMCAs a ton of people on GitHub, that only likes to play the game the way that they choose. The moment you play differently than they want, they punish you for it. And I'm disappointed. Like, it's weird to me that the lab that is the most anti-developer and anti-open is somehow liked the most by developers. It is genuinely really frustrating. Anthropic has gotten away with this sentiment for so long because they had the best models for coding and they still kind of do. GBD 5.2 is able to solve some problems Opus can't, but it's also much slower. The ergonomics of using a day-to-day just aren't as good. I find myself reaching for Opus still for coding, but I've never been more excited for a better model to drop because I don't want to support a company that operates this way. This just sucks. As I do in most of my videos like this, I would like to remind Gelanthropic that you still have time to fix this. I will give you a quick list of all of the things you have to do. Open source cloud code immediately. Like it is already a year later than it should be. That needs to be open sourced now. There's no reason to be hiding that source code. Reverse this decision. Make it easier for developers to apply to be officially blessed as an OOTH application that uses the rate limits in the existing subscription tiers. Donate a bunch of money to one of these open source solutions. Ideally, things like open code, like these products are essential and are helping elevate the success of tools that you're trying to build in the world that we're building for. So, support them. Don't ostracize them. And cut the [ __ ] I don't know how else to put it, but stop doing things like this. Have some outside counsel that doesn't work at anthropic, that doesn't glaze y'all immensely to tell you before a decision like this is going to happen to say, \"Hey, wait. This might piss off developers. We should rethink this.\" your goodwill with devs is dying and it's dying fast and it's 100% your fault and I am going to continue speaking out against this even though I'm now technically an anthropic investor because you guys bought bun. I don't care. I'm going to give you [ __ ] when you act shitty and a lot of other people are waking up to it too. I'm no longer the one anthropic hater. You are turning most devs into haters because you're [ __ ] up. Stop [ __ ] up. It's that easy. You might feel special because your model's slightly better. You might feel special because you were early. You might feel special because you quit OpenAI and now you're competing with them. Cool. You're not [ __ ] special. Get over your [ __ ] and fix it. That's all I have to say. You know what you're doing wrong. You know in your hearts that you're doing it wrong. Get over yourselves. It's not that hard. I've had to do it enough times myself. Fix your [ __ ] Stop taking from the devs that you're already taking from to build your goddamn models and behave a little bit. Just be a good faith player in the space. And if you're looking for examples on how to do that, take a look at OpenAI. I can't believe I'm actually saying that. I got nothing else for y'all. Just cut the [ __ ] Peace nerds.\n"

ERROR: type should be string, got "https://youtu.be/xBM307YwVRw\n\nI can't believe he was right.\n\nRemember back in March when Daario said\nthis\nif I look at coding programming which is\none area where AI is making the most\nprogress um what we are finding is we\nare not far from a world I think we'll\nbe there in 3 to 6 months where AI is\nwriting 90% of the code at the time I\nthought this was one of the most absurd\nstatements I'd heard any tech CEO make\nin history and it was and to be fair he\nwas wrong it actually took nine months\nBoris is the developer in charge of\nclaude code. He did an update a few days\nago talking about how originally cloud\ncode was a side project and now it has\nchanged how he writes code and now he\nbarely ever opens an editor. Every\nsingle line that he has done in the last\n30 days across 259 PRs, all of it was\ndone in cloud code and opus. Literally\n100% of his contributions to cloud code\nwere written by cloud code. Even here\nlike yeah he's an employee at the\ncompany. Of course he's going to say\nthat. But then I started using cloud\ncode with Opus more. I built the Chrome\nextension to lock me out of Twitter when\nI don't have Cloud Code running to\nincentivize me to use it more. And then\nI built a whole new image generation\nstudio with a separate conversation view\nwhere you can pick the photo you like\nand do follow-ups with it with a more\ndynamic editing experience. And then I\nmade it a mobile app, too. If you want\nthe details on all of that, a whole\nseparate video about how I use Cloud\nCode is coming very soon. Make sure you\nsubscribe so you see that. But that's\nnot what I'm here to talk about today. I\nwant to talk about what Daario said on\nthis panel as well as the part right\nafter that I feel like is often skipped\nwhen this conversation happens and what\nthis means for us as developers because\neven if his timeline was slightly off,\nDaria was absolutely correct. And since\nI might not get paid for writing code\nanymore, let's do a quick sponsor break\ninstead. There aren't many companies\nthat can proudly state that GitHub is\nscared of them, but today's sponsor is\none of them. Depot is so much better at\nrunning your GitHub actions that\nMicrosoft nuked the whole GitHub brand\njust a few weeks ago trying to tax you\nfor using it. They have since delayed\nthose changes and even if they made\nthem, it would still be worth it because\nDepot is that much better. We're talking\nabout realworld performance improvements\nin the range of 37x for a project like\nPost Hog. And remember, Post Hog's open\nsource. There's no magic tricks here.\nYou can go look at the code yourself and\nsee when the action suddenly started\ntaking this much less time. We're\ntalking a difference of over 2 hours\nfrom a minute and 50 seconds to 4\nminutes. Let's be real super quick. How\nmuch of your time are you wasting\nchasing down failing builds that are\nfailing for no reason? How much time are\nyou spending waiting for your builds to\ncomplete so you can actually merge the\ncode that you're changing? How much\nhesitation do you feel to push up a new\nchange because you know the CI is going\nto take another hour or two to run\nbefore you can do anything? Imagine life\nwithout any of those problems. Imagine\nyou put up a PR and the build just\nworks. There's no random errors or fail\ncases that you hit because GitHub\nactions are unstable. Imagine they're\nhappening in seconds instead of hours.\nImagine life without the hell that is\nwaiting on CI. If you haven't tried them\nyet, you're wasting your time. Fix that\nnow at soyv.link/depo.\nSo, first and foremost, I want to go\nthrough Daario's whole statement with\nyou guys.\nI do have a fair amount of concern about\nthis. Um, on one hand, I think\ncomparative advantage is a very powerful\ntool. If I look at coding programming\nwhich is one area where AI is making the\nmost progress um what we are finding is\nwe are not far from a world I think\nwe'll be there in 3 to 6 months where AI\nis writing 90% of the code and then in\n12 months we may be in a world where AI\nis writing essentially all of the code.\nWhen he first made this statement I did\na response because the way this is\nmeasured is going to be really really\ntough. First we have to define what the\ncode is. What is 90% of the code? Are we\ntalking about the code people are paid\nto ship to production? Are we talking\nabout the code being written at\ncompanies like Google and Microsoft and\nEnthropic? Or are we talking about all\nof code created from all of places?\nBecause I would imagine that the amount\nof repos being instantiated and the\namount of code being written by tools\nlike lovable and vzero and bolt and all\nof these types of things just by sheer\nnumber of lines has changed the amount\nof code being written year-over-year a\nmeaningful amount. I would be really\nsurprised if we had less than 2x the\namount of code written this year from\nthe year prior as in 2025 to 2024. I\nknow it's 2026 now. I'm sorry guys, but\nyou get the idea. 2025 probably had 2x\nthe amount of code or more written and\nproduced than in 2024. Does most of it\ngo in a garbage can? Sure. But way more\ncode came out. We can't disagree with\nthat. Like that's just table sakes. We\nhave to agree no matter what side we're\non that way more code happened last\nyear. That said, actual studies have\nbeen done on the numbers for real\nbusinesses. And according to Google's\nlittle AI summary thing, from real\nstudies,\nindustrywide, the number is a little bit\nlower. Apparently, 30% of code at\nMicrosoft is AI. Sundar said over 25%\nwas written by AI as of late 2024. I\ndon't know if we've gotten an update\nsince. Here's where things start to get\ncrazy, though. For senior devs, in\nrecent surveys, 32% report that at least\nhalf their code is coming from AI. More\nand more senior devs are the ones saying\nthat lots of their code is coming from\nAI. This is one of the biggest shifts\nthat happened in 2025. It's no longer\njust devs who are autocompleting some of\ntheir code. It's no longer just\nbeginners who don't have the skills that\nare using AI to make up for a skill gap.\nIt's now experienced developers using\nthese tools to make up for a time gap.\nAnd I found myself doing the same. There\nare so many random ideas I have for\nsystems, projects, features, and just\nstuff I want to fix in my codebase. And\nas the tools get better, the harnesses\nget better, and of course the models get\nbetter, the idea that this random thing\nthat maybe I would have previously cut a\nJira ticket for that would have sat in a\npit of backlog forever, now I can just\nthrow it at Opus, get some code, review\nit on GitHub, and merge it. There's also\na problem here around what our jobs are.\nRunning a quick poll to see how much\npeople like code review. And as we see\nwith the initial results, 50% say hate\nit. Even though there are four options\nthat are neutral or better. Now that the\npolls continued going, it's fine but\ntiring is going up. But we still have\n26% saying hate it, 10% saying neutral,\nand 45% saying it's fine but it's\ntiring. Very small percentage of devs\nlike or love code review. I'm in this\nsection. I would argue I'm in the like\nit section. I do enjoy code review. I\nwish I could justify doing more of it,\nbut my schedule is utter chaos, but I\nlike code review. I also think this is\njust one of those things that you have\nto get over over time as you become a\nmore experienced developer. Not that you\nhave to get over using AI tools, you\nhave to get over your own time spent\nwriting code. I actually had a long chat\nwith Mark not too long ago. If you don't\nknow, Mark is my CTO for T3 and all the\nfun things that we built. Mark, now that\nwe have more staff, as in Julius is here\nfull-time in the US and Yosh still\ncontributes throughout his week when he\ncan, Mark's time isn't best spent\nwriting code anymore. Previously, the\nflow was Mark would write most of the\ncode. Julius would go off and figure out\nwhatever weird problems, and I would\ncome in to do some steering and general\ncode review stuff. Now, I'm more removed\nbecause there's too much [ __ ] going on.\nRunning the company is a lot of work,\nand also not doing the best job as CEO.\ntrying to get more involved in the\ncodebase again, but it's not the best\nuse of my time, which means Mark has had\nto bump out a little bit too to do more\nof the orchestration and code review,\nespecially with the mobile app in\nflight, too. Managing the team, building\nthe mobile app, managing the team,\nbuilding and maintaining the web app and\nthe servers and the back end, managing\nupload thing, managing all the other\nservices that we maintain. He's not\nwriting code as much anymore. And as\nmost people will experience when this\nhappens in their career, that [ __ ]\nsucks. You suddenly stop feeling\nproductive when you're not writing code\nanymore. And previously, you would\nmeasure your success in a day by how\nmuch code you had written, how much time\nyou spent in your editor, and how many\nproblems you had solved. If you're\nspending your time in the GitHub PR tab\nand not in your editor, you feel like\nyou're not getting anything done. And\nthe amount of times Mark would have a\ncrazy day where he did a ton of code\nreview, got a ton of stuff shipped,\nhandled a ton of support cases, all\nthose things, but felt like he didn't\nget anything done simply because he\nwasn't in his editor at all. It's it\nbecame like an everyday occurrence and\nhe was not loving that feeling because\nit sucks and I've seen this happen to a\nlot of developers as they move towards\nmanagement towards tech leads towards\nroles where their job becomes less\nwriting code and more orchestrating\ncode. This is part of what makes a\nsenior engineer. This is the biggest\nreason I think we are seeing that shift\nwhere more and more senior plus devs are\nreporting that a majority of their code\nis coming from AI. We are more used to\nreviewing code constantly. The average\ndev still probably writes more code than\nthey review. The devs that have gotten\nover that already, the ones who are\nmaintaining big open source projects,\nthe ones who are running teams with a\nlot of engineers, the ones who are\nbuilding systems that require lots of\npeople doing lots of disciplined work.\nThose are the people I've seen falling\nthe hardest in love with these AI tools\nover the last few months. The people who\nare already over that hump and are used\nto that are the ones who are getting the\nbiggest benefits right now. There are\ndays where I'll file like 20 pull\nrequests and go through them all on\nGitHub. I spend more time generating\ncode and then putting it on GitHub than\nactually reading it in my editor at this\npoint. It's crazy. But that also sucks\nfor a lot of people. I know many\nengineers, myself included, prefer\nwriting code to reviewing it. In fact,\nfrom this poll, it seems like the vast\nmajority of developers very much prefer\nwriting code to reading it. I\nunderstand. But that's a thing that will\nsuck about our jobs going forward. more\nand more of our job is going to be\nreviewing decisions and work that is\nbeing done by these tools, not doing the\nwork ourselves. Sucks, but it is what it\nis. And those who are willing to make\nthat jump are going to see crazy\nsuccess. Sucks when you started writing\ncode as a hobby. TBH, it feels horrible\nhaving your hobby be automated. On one\nhand, totally feel this. It has been\nweird moving away. On the other, I have\ntwo points. The first is this is the\nnatural progression of an engineer as\nthey get better. You need to learn to\ndelegate and do more talking and\norchestration and long-term planning and\nmanaging of your team and the\ndevelopment velocity and the changes\ngoing in and how they're going in and\nall of that. Like that is growth as an\nengineer. Once you are good enough at\nthe code writing part, you have to get\nbetter at those other things to keep\nleveling up. There's a cap at how much\nthe code writing actually benefits you.\nThe other part of this side is that this\nis what we've been doing as engineers\nforever. We've been writing code to\nautomate things that were hobbies for\nother people. We have already destroyed\nso many fields with code. It was only\ntime that the fun parts of our field\nstarted to erode. But I will counter\nthis by saying I'm having a lot more fun\ncoding overall lately. It's much more\nfun seeing how far you can push these\nagents and the crazy things they can do\nand not feeling bad when you throw away\nthe code cuz it sucks. The worst feeling\never is when one of your employees or\nteammates or somebody you care about\ngoes off on a crazy tangent, deep dives\non something, builds a crazy solution,\nand then for any of many reasons, it\njust doesn't work. It doesn't fit the\nproject, you can't merge it, and you\nhave to close the PR. It hurts. It\nsucks. It feels so bad to throw away\nhard work that one of your co-workers or\npeers or friends did. But if it was an\nagent, I don't give a [ __ ] I just\nclosed the PR. So, I'm having a lot more\nfun with experiments. I'm having a lot\nmore fun trying things out, playing with\nnew languages, overhauling code bases,\nexperimenting with new stuff. It's I'm\nwriting comically more code than I've\never written. I threw together a 12,000\nline of code project in a day. What?\n12,000 lines of code in a day. According\nto various surveys, the average\ndeveloper on the average day writes\nbetween 10 and 40 lines of thoroughly\ntested and debugged code. Opus and I\nwrote 12,000. This might seem like cap,\nbut when you average it out, this is\nvery real. I know everybody in chat's\nlike, \"No, that's stupid. That's\ninsane.\" Now, if you do a PR with 250\nlines of code, you start working on it\non Monday, you put up the PR Tuesday, it\ntakes two days to get thorough reviews.\nOn Thursday, you finally have a good\nenough review and have to go make\nchanges. So, you make changes, and then\non Friday, it ships. That was 5 days for\n240 lines of code. That is That is about\nright. And that's hilarious. That is\nabsurd. This is the 10 to 100x people\nkeep talking about is when it is\nsignificantly easier to not just write\nthe code, but write the tests to\nautomate the testing process to set up\nendto-end testing. GMO just posted about\nthis and the more I've thought about it,\nthe more I think he might be right. AI\ncoding will result in the most\nrigorously tested, type checked, and\nprovably correct code in the world.\nFirst, the testing and verification\nfeedback loop helps agents and prevent\nslop. Yep. God, cloud code got so much\nmore usable when the LSP was added. It's\nkind of broken, but it helps so much.\nBut the second part, tests have always\nbeen a chore, but who doesn't love the\npile of check marks that a high test\ncount gives you? Agents are glad to\nwrite as many tests as we want. Yep, the\npain of writing and maintaining test has\ngone down a ton, and I'm starting to\nrethink my own relationship with testing\nfor my projects. This is one of the\ncoolest parts. It would take me way more\ntime to write testing code than the\nactual code because getting the test\nright is hard and tedious and slow and\nannoying and not actually the thing that\nwe like doing. Now the agent can do that\npart and then do the implementation and\nthen put up the code and then do the\nfirst pass on the code review and catch\nany mistakes and then you go and look\nover the like final pieces and give it a\nthumbs up or down. It's a whole\ndifferent way of building and it is\nchanging how I think about software.\nBack to the rest of the quote because\nthis is what I think a lot of people\nhave missed.\nBut the programmer still needs to\nspecify you know what what what are what\nwhat are what are the conditions of what\nyou're doing? what you know what what\nwhat what what is the overall app you're\ntrying to make? What's the overall\ndesign decision? How do we collaborate\nwith other code that's been written? Um\nyou know, how do we have some common\nsense on whether this is a secure design\nor an insecure design? So, as long as\nthere are these small pieces that a\nprogrammer, a human programmer needs to\ndo that the AI isn't good at, I think\nhuman productivity will actually be\nenhanced. I really want to emphasize\nthat part because it's a huge part of\nwhat I love about coding. It's not just\nthe process of writing the beautiful\nlines in my editor that I could stare at\nand smile and screenshot and tweet. It's\nthe way I think about the system, the\nway the parts come together, when the\nright abstraction works in the right\nplace the right way and it all comes\ntogether well. When you pull together\nthe set of tools you want to use and\nthey come together the right way or one\nof them doesn't quite work how you need,\nso you swap it out for something else\nand it does work. And that feeling of\ncohesion when you get the right parts to\nwork together. That's my favorite thing\nwhen I'm writing code. When all of the\npieces come together and the result is\nsomething awesome. Both a thing that\nyou're confident in your ability to\nmaintain and also a thing that is\nactually pleasant to use for the product\nthat you're actually putting together.\nThis is awesome. I find this so fun. It\nis different. It is strange. It's a\nfundamental change in how we write code\nevery day. What does this mean about our\nday-to-day work, about our careers,\nabout the people that we hire? About the\npeople we might have to fire? How is our\nworkforce going to change? A lot of\npeople said this prediction obviously\nwasn't true because Anthropic not only\ndidn't fire 90% of their engineers,\nthey're actually still trying to hire\nmore, but they also have been doing a\nlot more coding. In fact, Daario\nactually mentioned this in a follow-up\nwhen people were questioning his take\nhere. Thank you to FaZe for fixing the\naudio in this clip so we can actually\nlisten to it.\nI would say maybe 70 80 90% of the code\nwritten in Enthropic is written by\nClaude. You know, I said something like\nthis 3 or 6 months ago. People think of\nit as falsified because they think of it\nas like we're going to fire 70 80 or 90%\nof the software engineers. But what\nreally happens is that the 10 the 10%\nwe're still writing, you know, humans\nbecome managers of AI systems. There's a\nshift because of the principle of\ncomparative advantage. So it looks more\nnormal than you think. I think\neventually it all, you know, kind of\neventually that logic may not hold, but\nbut there's a sort of sci-fi sheen to\nthese predictions to looking at the\nfuture that it's going to be weird that\nyou'll be looking through different\ncolored glasses that there will be, you\nknow, that it'll look like Star Wars or\nsomething. Um the often when these\npredictions come true, it's wild, but\nit's also in a way ordinary. Yeah,\nthey're just building more and that\nmakes a lot of sense. And you get used\nto it and it no longer feels as crazy.\nAnd I'm even feeling that a bit too.\nLike when I'm just [ __ ] around in\ncloud code, I don't necessarily feel\nlike, oh, this all happened at once.\nIt's when I left my computer and was\ntalking about it with friends that it\nhit me just how absurd it is that I\nbuilt two projects without ever opening\nan IDE. Like, that's just crazy. That's\nabsurd. So absurd that now Boris is\nagain doing all of his day-to-day work\nin Claude Code 2. And Darren made a post\na few days ago about how you should\ntotally now be able to have one highly\nskilled person doing the equivalent of\n10 engineers of work. Does this mean\nthat the industry will be slashed to\nonly the top 10% of engineers if that's\nall that's needed? Shouldn't we be\nconcerned? I really like Simon's\nresponse here. If code gets 10 times\ncheaper, the demand for code will go up\nby more than 10x. A whole bunch of\ncompanies that previously couldn't\nafford custom software will suddenly be\nin the market for it. And existing\ncompanies with multi-year road mapaps\nwill hire more to help get them through\nit. This is known as Jevans paradox. The\nJeans paradox is a thing that occurs\nwhen technological advancements make a\nresource more efficient to use, thereby\nreducing the amount needed for a single\napplication. But if the cost goes down\nand the demand is price elastic, this\ncan result in demand increasing instead\ncausing the total resource consumption\nto rise. And I think we're already\nseeing this. There is definitely more\ncode happening. There are definitely\nmore companies more interested in\nbuilding their own internal tools than\nI've ever seen. For example, at RAMP,\nthe finance management company for\nstartups, they have been building their\nown internal tool that uses open code to\ntake some plain text description of work\nand go do it. Here is one where they\nasked, \"What are the 20 most common\nsensory issues in core? Spin up child\nsessions fixing all of them.\" And\napparently it just worked. They looked\nat the sentry issues and then spun up 20\nagents to go fix all of them and then\nfiled 20 poll requests. That's kind of\ncrazy. This does mean you're spending a\nlot more time reviewing code, but also\nthe process of linking these things all\nup like this. And I'm certain that 90\nplus% of the code in this inspect tool\nthat they made is probably vibe coded,\ntoo. This is the thing that is changing\nthat is crazy. And the more you dive in\nand play with it yourself, you'll find\nthe way you think about these things\nchanges. When you realize you can go\njust spin up the tool and it's not that\nmuch work, you'll try things you\nwouldn't have tried. I'm considering\ngoing as far as setting a minimum\ninference spend for my team. I might\nmake it so everybody at T3 Chat is\nexpected to spend at least $200 a month\non inference so that they're going to go\nout and try new strange [ __ ] I want to\nsee what they do when I put some\npressure on them to spend more money\nrather than less because then they might\nexperiment a bit more and do some\ncrazier stuff. I'm probably going to do\nthat at this point. I've been thinking\nabout it too much. I want to I'm\nprobably going to. This fun report just\ngot linked while I was filming and I\nwanted to include it because it's\nanother one of the reasons we have to\nreview the code that the AI is writing.\nAI code creates 1.7 times more problems.\nMy initial question was when was this\npublished? because that will change how\nI think about this report a lot. And I\ncouldn't figure it out, but everybody in\nchat was saying they could see it. Check\nthis out. When I squash it down to\nmobile size, it shows the date, December\n17th. And when I zoom out, we get this\nsidebar that shows December 17th, 7\nminute read. But if you're zoomed in too\nmuch for the sidebars to fit, but not\nenough for it to trigger the mobile\nmode, you don't get to see the values\nthere. Just one example of a bug that\noccurs when you vibe code your UIs. Just\nfunny to see one of these bugs in the\narticle about how you find up to 70%\nmore of these in AI code than human\nwritten code. This report was across 470\nPRs, including 320 that AI co-authored\nand 150 that were human only using code\nrabbit's structured issue system across\na bunch of open source projects. And\nyeah, AI does accelerate the output, but\nit also amplifies certain categories of\nmistakes. Surprise, surprise, you get\nthe idea. One interesting piece is that\nthey only measured as per PR. They\ndidn't do per line of code. Also, high\nissue outliers were more common in\nAIPRs. So, you would randomly have a PR\nthat had way more issues. 1.7x at the\nlike average, but it's 2x at that 90th\npercentile. So, when it does bad, it\ndoes way worse. Cool information and\nagain more reason you need to be doing\nthe code review part. But this leaves us\nwith my biggest concern, the junior dev.\nAnd I'll be honest, I don't know what we\ncan do about that. I built up these\nskills by writing code for almost 20\nyears. And those skills are still\nessential in my day-to-day use of these\ntools. Just knowing weirdness about how\nCSS works and scroll containers being\nstrange in the browser. the amount of\ntimes I've had to remind Claude Code\nthat it doesn't have the ability to have\na tool tip come out of a box when it has\nany overflow rule set on it. I don't\nknow if you know this, but if you set an\noverflow X rule on a box, you can't go\nout of it Y direction. So, if you have\nsome behavior where it has a border\ncolor when you hover over it, that will\nget cut off. Or you have a little X\nbutton in the corner of the square,\nthat'll get cut off. And I've had to\nremind Opus about this so many times\nthat it suggested adding it for me to my\nglobal quad MD file. Like you need to\nknow these things to be successful with\nthese tools. But these tools existing\ndrastically lowers the incentive to\nlearn these things because the cost of\ndoing it wrong used to be really high.\nAnd the cost of figuring it out also was\nreally high. You would spend the time to\ndo it. Now, if you don't know the reason\nthese things happen, you can screenshot\nthe problem, show it to the model, and\nsay, \"Hey, fix this.\" And there's a one\nin three chance it does it right, and\nthen you keep doing that over and over\nagain forever. The amount of these types\nof issues will go down constantly. But\nalso, the amount of work we have these\ntools doing is going up constantly. If\nit used to make one of these mistakes\nevery 100 lines of code, then it was\nkind of useful. Now, it's every thousand\nlines of code. Maybe they'll get to\nevery 10,000 lines, but it doesn't make\nthat bug any easier to solve with the\nAI. And I don't know how you get that\nexperience now. I really don't. I've\nbeen thinking a lot about this. I kind\nof want to do a follow-up on my like\njunior dev videos, like how do I think\nabout this for early career engineers\nnow. And I don't know what to say or\nrecommend. I still think it's really\nimportant that you try to use these\ntools not as a way to do work you don't\nknow how, rather to break apart and\ndelegate work that you don't have the\ntime to do. That's when they're\nstrongest. When you're using these tools\nto multiply your own capabilities, not\nto work around your lack of\ncapabilities. But I don't know how you\nget those capabilities anymore. At the\nabsolute least, you need to read the\ncode that you're generating, especially\nwhen you're early and you're spinning up\nthings that are actual projects you want\nto maintain over time. Writing code\nmanually is still a very good way to\nlearn, but honestly,\nI still think until you have gotten\nrelatively deep and are like multiple\nprojects in building things that real\npeople use, you should probably\nprioritize using a chat app to ask\nquestions. Or maybe you can shortcut\nthis and use something like cursor and\nuse the ask mode where it doesn't change\ncode, it just answers questions to learn\nmore. Combine that with tab complete and\nyou can keep leveling up your\nunderstanding. But man, this this\nstuff's changing fast. And the thing I\nnever would have expected is that my\nskills as a senior dev, which are mostly\norchestration and planning and\nmanagement of people doing things in\nparallel. That part matters now more\nthan ever, which is cool on one hand\nbecause now a lot of devs are going to\nskill up on their talking skills. Like\nprompting is more a comm skill than an\nend skill in so many ways, but also like\nyou might not learn the ins and outs the\nsame way. This is tough. I don't really\nknow what I want to what my point here\nis other than to express how confused I\nam still. Oh [ __ ] Lero's here. Yeah,\napparently they're adding a teacher mode\nto cursor, which is actually really\nreally really exciting, especially if\nthey let people do this on like a\nstudent tier. Like if they're in school,\nthey get it for free. Oh. Oh, that'll be\nso good. Also, tools like code crafters\nare more useful than ever in my opinion.\nCode crafters, bootupdev, stuff like\nthat that actually encourages real\nproject building. This is how you have\nto learn now. And the AI will make it so\nyou're less likely to get blocked and\nless likely to need somebody to help\ncarry you through the hard parts. Yeah,\nknow this is going too deep. I probably\nneed to think this through, like sit\ndown and really think about what works\nand talk to some recent people who just\ngot into the industry and then maybe\nI'll make a dedicated video on this. But\nright now, I don't even know how to like\norganize my thoughts on it. One more\nquestion for you as a viewer of this\nvideo. If I was to ask you, where do you\nthink you fit in terms of intelligence\nand capability writing code against\nOpus? Would you consider yourself a\nbetter engineer, same tier roughly, or a\nworse engineer? Where do you think you\nfit? And what answers do you think my\nchat would give? How do you think the\nsplit would be of those? Think about it\nfor a sec. Get your idea of where you\nwould put yourself and where you think\nthe splits are for my own audience.\nReady for the results?\n61% of my viewers think that Opus 4.5 is\na better death than them. 16% did the\nbailout answer. 13% said roughly as good\nas me. And 10% said worse. That's where\nwe're at. These things are moving\nabsurdly fast. The way we code now is so\ndifferent from the way at the very least\nI coded 3 to 4 months ago, which is way\ndifferent from how it was at the\nbeginning of the year, which is way\ndifferent from how it was the year\nbefore. But the speed at which these\nchanges in my workflow are happening is\ncondensing. It's like every three months\nnow massive change happens. Previously\nit was every six months. Then it was\nevery year. Then it was every few years.\nIt's crazy. Don't fall behind. I'm going\nto do a whole dedicated video on how to\nkeep up on all of this in the near\nfuture. I also have a video about my own\nusage of cloud code that's coming very\nsoon. Keep an eye out for that. I'm very\nexcited about that one. So yeah, keep\nplaying with this stuff. Keep trying new\nthings and keep doing your best to build\nthe best possible things and take\nadvantage of the tools that exist around\nyou. Don't write this stuff off because\nyou're scared of grifters or people like\nme who happen to have invested in a few\nof these tools, almost none of which I\ntalked about today. Just focus on\nbuilding. That's what's going to get you\nthe furthest always. And there is more\nto be built than ever. And there are\nmore ways to build than ever, too. I've\nnever loved code as much as I do today,\neven if my relationship with it is very\ndifferent than it was a year ago. And I\nrecommend that you reflect yourself and\ngive these things a try. Let me know\nwhat y'all think and how you're using\nthese tools today.\n"

ERROR: type should be string, got "https://www.youtube.com/watch?v=Ge8LoXfJJdA\n\n2025: The year I stopped writing code\n\nI think it's fair to say that 2025 was a kind of wild year for those of us who use AI to write code. Not just because\nthe models got way smarter, but everything around it. From reasoning models kicking off the year to agents\ngoing from buzzword to actually usable in our day-to-day usage to cool tools like cloud code, open code, codecs, and\nmore embracing us where we live, our terminals. It's been a wild year, and the results are wild, too. I personally\nfind that I am not spending anywhere near as much time in the editing view in my code editor, if I'm even in a code\neditor at all. It's been insane. And I'm not the only one who thinks this. I've been collecting a bunch of people's 2025\nwrap-up things so that I can go over how we as an industry of vibe coders and\npeople using AI tools feel about how the year went, overview all the cool things that happened and showcase some really\ninteresting trends and what these mean for us as an industry. from how PR sizes changed to how different SDKs grew\nmassively in adoption to how models themselves improved and changed the amounts of things they could do. There\nare so many interesting things that happened this year that have resulted in the way we code changing. It used to\ntake years for the way we code to change. Then it started to take about a year, then six months, then three\nmonths, and now it honestly feels like the time between these transitions is even shorter than the time it takes to\ntell you about today's sponsor. In 2025, the way we wrote code changed forever. But now it's 2026, and your CI is still\nworking like we're in the '9s. Seriously, why are we waiting for hours for these builds when they could take\nminutes, if not seconds? Blacksmith is here to save you from those terrible build times. You make a oneline change\nin your GitHub action and now your costs are 75% less. Half because they cost half as much and half because they take\nhalf as long. That's a lot of halves. Seriously though, it's crazy how much better things are once you make the move\nto blacksmith. Companies like Superbase, Dcript, Exa, Clerk, and more have already made the move. And the reason\nwhy is obvious. The cache downloads are four times faster. The hardware they run on is two times faster. The Docker\nbuilds they can provide are up to 40 times faster because they're caching those build layers on the same machine\nthat's doing the build on an NVME drive. It's just hilarious how much faster things are once you make the move.\nThat's not my favorite part though. The observability is how nice would it be if you had actual working search across\nyour actions. How nice would it be if you had a panel that showed you how often specific actions fail? Wouldn't it be nice to know which test is the flaky\none that's causing these problems and to have a little more confidence when a PR fails? How many times have you rejected an AI generated PR because a test failed\njust to realize later on that that test was flaky? You've probably thrown out code because GitHub actions are less\nreliable. I could sit here and tell you how much more reliable Blacksmith is, but I'd rather tell you about the observability. So, when it does fail,\nyou can go figure out why. Seriously, Blacksmith is one of those no-brainer things that once you've made the move, you'll question why it took you so long.\nCheck him out now at soyv.link/blacksmith. I want to start with Simon's post. If you don't know Simon, he's the creator\nof the Django framework in Python and is one of the best people on the internet\ncovering all the cool things going on in the LLM world. He did this awesome write up about 2025 and the things that\nchanged throughout. And I want to briefly go over this while also pulling in interesting sources from the state of\nAI review and a couple other places, too. Shout out to everybody who wrote all of these awesome resources. There's\nso much good [ __ ] here. I highly recommend you follow these people and read their stuff as well. There's a lot to learn from these people. And if you\nwant to keep finding cool sources like this and stay on top of things in 2026, subscriptions on YouTube are actually\nfree. You just hit that little red button and now you'll see what my videos a little bit more often. They're not going to spam. You're not going to get\nnotifications unless you hit the bell, but you'll be a little more likely to keep up. And if you sub on a video like this, we see that and we make more\nthings like this. So yeah, hit that button when you think it's useful. Anyways, the year of reasoning. This is\na really good starting place cuz it's kind of where we started the year. Before the end of 2024, all models were\njust next token completion. They still technically work that way where they have some text and they guess what the\nmost likely next text will be is, but reasoning was a very interesting change where the model effectively had a box\nthat it could talk to itself in to create its own context to think things through as a way to iterate on an idea\nbefore coming to a conclusion. And that was super powerful. And then this got\ncopied right at the start of this year with Deepseek in their R1 model which blew everybody away and kind of opened\nthe floodgates for reasoning because the open AI model wouldn't even show you the reasoning tokens. So the only way we\ncould see how this process worked was through the work DeepSeek did and I applaud them still for all of the effort\nthey put in from the 12 plus papers they published in 2024 to the insane models they published at the end as well as the\nstart and middle of this year. reasoning really became the standard. And now it's just kind of weird when a new model\ncomes out and it's not a reasoning model because it's so apparent how much this has helped things. Even anthropic who\nwas skeptical decided to put it out anyways and in their reasoning traces shared them because they didn't know why\nit worked so well. And they hope that by sharing the reasoning traces more people would be able to figure out why this\nworks as well as it does. As Simon says here, OpenAI doubled down on the reasoning stuff with 03, 03 Mini, and O4\nMini. And now GPT5 is also a reasoning model. Remember back in the day when the O models were the reasoning models and\nthe GPT ones were the non-reasoning. Crazy how that's changed, isn't it? Now reasoning is just a checkbox for the\nmodel. I also love that he's referring to reasoning as a trick. It effectively is. And this is what Karpathy had to\nsay. By training LLMs against automatically verifiable rewards across a number of environments like math and\ncoding puzzles, LMS can spontaneously develop strategies that look like reasoning to humans. They learn to break\ndown problem solving into intermediate calculations and they learn a number of problem solving strategies for going\nback and forth to figure things out. You can see this in the Deepseek R1 paper. Running reinforcement learning with\nverifiable rewards turns out to offer high capability per dollar, which gobbled up the compute that was\noriginally intended for pre-training. Therefore, most of the capability progress of 2025 was defined by the LLM\nlabs chewing through the overhang of this new stage. And overall, we saw roughly similarly sized LLMs, but much\nlonger RL runs. What this means is that previously all of the work was going into collecting the data and organizing\nit to train the model to be as capable as possible right when the weights came\nout. But with all of this reinforcement learning stuff combined with reasoning and how effective you can tune it with\nreinforcement learning, more and more effort went into adjusting the model after it was made using these techniques\nbecause it ended up being cheaper per benchmark point overall. So, as funny as it sounds, a lot of the models that we\nwere excited about this year weren't necessarily new models. This has been very common with OpenAI in particular,\nwho allegedly, and I don't have any internal confirmation on any of this, allegedly has not used a new model prel.\nWe have been using the same base model from OpenAI with more reinforcement done on top. They've not changed their\npre-training at all. Allegedly, this might change in the near future, too. But yeah, a lot of this has been the\nmodels being adjusted by throwing more reinforcement after. Every notable AI\nlab released at least one reasoning model in 2025. Some labs released hybrids that could run in reasoning mode\nand non-reasoning modes. Many API models now include dials for increasing or decreasing the amount of reasoning\napplied to a given prompt. It took Simon and myself as well a while to understand what reasoning was useful for. Initial\ndemos showed it solving mathematical logic puzzles and counting the Rs in Strawberry. Two things I didn't find\nmyself needing into my day-to-day life with model usage. Turns out the real unlock of reasoning was in driving\ntools. Reasoning models with access to tools can plan out multi-step tasks, execute on them, and continue to reason\nabout the details such that they can update their plans to better achieve their desired goals. One example of this\nworking well is AI assisted search, which now actually kind of works. Hooking up search engines to LLM had\nquestionable results before, but now I find even my more complex search and research questions can be answered by\nGBT5 thinking in chat GPT. Also in T3 chat check it out if you haven't. They're also really good at writing\ncode. They can think about what could be wrong and fix the code much better too. When you combine reasoning with tool\nuse, you get the year of agents. This is why I love Simon. He started the\nyear making a prediction that agents were not going to happen. I want to read this. He was on the Oxide and Friends\npodcast. Love that podcast. And one of his predictions was that within a year, agents will fail to\nhappen again because he just hadn't seen any real progress for it in 2024. I remember 24 hearing people talking about\nagents and none of them could even define what it was or why it would matter. And that has changed a lot since. Throughout 24, everybody was\ntalking about agents, but there were few to no examples of them working. further confused by the fact that everybody used the term differently and they all appear\nto be working from different definitions entirely. By September, Simon had gotten fed up of trying to avoid the term\nbecause it was so overloaded that he decided to make his own clear definition, which is an LM that runs in\na loop to achieve a goal. Specifically, it runs tools in a loop. This one blocked me from having productive\nconversations about them, which was always his goal with terminology for tech like this. Didn't think agents\nwould happen because he didn't think the gullibility problem could be solved. I'm curious how he defines the gullibility\nproblem. LMS believe anything you tell them. Any systems that attempt to make meaningful decisions on your behalf will\nrun into the same roadblock. How good is a travel agent or a digital assistant or even a research tool if it can't\ndistinguish truth from fiction? Yep. And I can see why reasoning would help a lot with that. The idea of replacing human\nstaff members LLM was laughable at the time. He was half right in his prediction. The science fiction version of a magic computer assistant that does\nanything you ask of her did not materialize. I love the link to her. But if you define agents as LM systems that\ncan perform useful work via tool calls over multiple steps, then agents are here and they are proving to be\nextraordinarily useful. Yep, I use agents every day for all sorts of work, which is kind of crazy. The two breakout\ncategories are coding and search. Yes, the deep research pattern where you challenge an LLM to gather information\nas it churns away for 15 plus minutes building you a detailed report was popular in the first half of the year\nbut fell out of fashion now that GPT5 thinking as well as Google's AI mode. All of these now can produce comparable\nresults in a fraction of the time. I consider these to be an agent pattern and one that works really well. I agree.\nBut coding agents are a much bigger deal. This is the year of coding agents and claude code. The more impactful\nevent of 2025 happened in February with the quiet release of Cloud Code. I say quiet because it didn't even get its own\nblog post at the time. Enthropic bundled the Cloud Code release in as the second item in the post announcing Claude 3.7\nsonnet. Why did Enthropic jump from 3.5 to 3.7? It's cuz they had an update to 3.5 in October but kept the name but we\nall called it 3.6. Yeah, they burned a whole version number because of their naming. Cloud code is the most prominent\nexample of what I call coding agents. LM systems that can write code, execute the code, inspect the results, and then\niterate further. And now all of the major labs have put out their own coding CLI. I want to put in a quote that I\njust heard from Boris in a video recently that I think will contextualize\nwhy the claude code thing happened later. Let me find this. And so I think\nhe really understood the scaling laws kind of internally about how quickly the models are improving. And so he actually pushed me really hard to be like don't\nbuild for the model of today build for the model 6 months from now. And so honestly for a long time quad code was\nnot a great product. And even when it was used internally I used it for maybe like 10% of my code or something like\nthis. You know I use it sometimes but it really just can't do most things cuz the model is not capable enough. And then at some point we released uh sonnet and\nopus 4 and I think this was maybe March of this year and the product just worked. And we saw this in kind of the\nusage data and I saw this in my own coding. I started to be able to use it for probably like half of half of my\ncode. And this was totally borne out because this was actually like literally 6 months after starting the project.\nThis was the timeline. And you know at this point most of quad code is written using quad code. I think it's like 80 or\n90%. Yeah. This particular section of this interview was fascinating to me. This\nidea that they didn't build quad code around what the models did now. Boris built claude code around what he was\nhoping the models would be able to do in six months to a year and then it could and then it did and here we are. It's\nvery interesting and I've been thinking a lot about this particular quote because it shows why cloud code was such\na quiet release and it also shows a different way of thinking about how you build right now. If you're willing to\nmake bets on these things continuing to improve, you have a good chance of success. And if you build a tool that\nmodels can sometimes use to do meaningful work right now, but if the models were better, the tools will work\nbetter. That's the right thing to build. I've been thinking about this way of thinking for a while because Sam Alman\nsays something similar at a YC event. He said that if you're concerned about the model getting better and how it might\nhurt your startup, think to yourself, if the models get better, does this make your product better or worse? If the\nmodel's improving makes your business less relevant, like I don't know, maybe you do recipe suggestions. So, if the\nmodels get better, then your dedicated app for recipes might not be as useful anymore. You might be screwed. But if\nthe model's improving, makes your app better. Maybe you're doing an AI coding tool so a smarter model can make it more\nuseful or more usable. Then that's a good thing to be building. But I hadn't thought of this specifically as building\nfor if the models get better. And that makes the success of cloud code and when the success of cloud code happened make\nmuch more sense. It really comes down to building the harness around your aspirations and then the aspirations\nbeing hit and now everything works. Apparently, Simon was playing with things like this in early 2023 with the\nCHBT code interpreter, which was baked into a chat GBT that let it run Python code safely in a Kubernetes sandbox. He\nwas delighted this year when Anthropic finally put out their equivalent in September, albeit under the baffling initial name of create and edit files\nwith Claude. In October, they repurposed that container sandbox infrastructure to\nlaunch Cloud Code for web, which he's been using on almost daily basis ever since. I have not tried Cloud Code for\nweb yet. It probably should considering how much I have been using cloud code lately. We'll see. Cloud code for web is\nan asynchronous coding agent, a system that you can prompt and forget and it will work away at the problem and file a\npoll request when it's all done. OpenAI has this with Codex Cloud, which has been renamed to Codex Web recently,\nwhich launched in May. Gemini has Jules, which also came out in May. Then there's also companies like Devon that are\nbuilding just this. Simon loves the asynchronous coding agent category. They're a great answer to the security\nchallenge of running arbitrary code execution on a personal laptop. And it's really fun being able to fire off multiple tasks at once, often from my\nphone, and get decent results a few minutes later. He has a whole bunch of articles about that if you're curious.\nBut as he was talking there, this is exciting because he doesn't have to run things on his computer, but if you do\nwant to run things on your computer, 2025 helped that a lot too with the LMS finding their way to the command line.\nApparently, he already had an LLM tool for command line, so you could ask questions in the CLI and get responses,\nwhich he really liked doing. Again, Unix pipes all make sense. Cloud code and friends have conclusively demonstrated\nthat developers will embrace LLMs on the command line, given powerful enough models and the right harness. It helps that terminal commands with obscure\nsyntax like set and ffmpeg and bash itself are no longer a barrier to entry when any LLM can spit it out the right\nway, right, for you. I totally agree. I know a couple devs that were always a little scared of the CLI that through\nclaude codecs and things like that find themselves using the CLI more. Even I find myself configuring my CLI more.\nUsually I get it set up and then don't touch it. I've edited my Zish profile in my Zish RC file more this year than the\nlike five years prior. Apparently Claude Code itself has a billion dollars a year\nof revenue which is kind of crazy but also I'm on the $200 a month tier. Makes sense. I'm not calling it Zsh. It's\nZish. I'm sorry. Even if the creator of Zish says it's Zsh, it's Zish. I'm\nsorry. Don't spell it in a pronouncable way if it's not pronouncable. It's Bash and Zish, not Bash and Zsh. No. Be\nrealistic, guys. Okay, I know the real name is Zshell, but I'm calling it Zish.\nIt's Zish. Prime would be proud. That's almost insulting. I don't talk about the shell. Can LMs\nautomate pronouncing things properly? You know what I say to that? YOLO. The\nyear of YOLO and the normalization of deviance. This is one of the most interesting trends. On one hand, people\nare having their home folders deleted. On the other, they are letting these agents run for hours at a time. And now\nI'm guilty of the same. I bet if I was to up arrow on this random terminal,\nyep, this wasn't allowed. Dangerous. I haven't had the balls to like override the command to always have this. I am\ntempted. I am very tempted. I'm not quite there yet. Regardless, the idea of\nYOLO mode is nuts that we now trust these tools enough to let them just run rogue on our computers. Codeex goes as\nfar as aliasing their dangerously bypass approvals and sandbox flag to- yolo,\nwhich I love. Using an agent without the safety wheels feels like a completely different product. Yep. It's crazy how\nnot having to say yes over and over changes how you use these things. This is also why asynchronous coding agents\nare cool because when they run in YOLO mode, they're less likely to cause problems. Simon runs in YOLO mode all the time despite being deeply aware of\nthe risks involved and it hasn't burned him yet, but that's the problem. One of his favorite pieces in LLM security this\nyear is the normalization of deviance in AI by security researcher Johan Reberger. Hopefully I got his name\nright. Yan describes the normalization of devian phenomena where repeated exposure to risky behavior without\nnegative consequences leads people in organizations to accept that risky behavior is normal. Yep. Very common. If\npeople do things that are risky and nothing happens to them, they keep doing it. This is why Space Shuttle Challenger\ncrashed. Yep. Since so many launches were successful, people stopped caring as much about the risks and didn't take\nthings as seriously. The longer we get away with running these systems in fundamentally insecure ways, the closer\nwe are getting to a challenger or disaster of our own. Yeah. Yeah. Should I always use dangerously skip\npermissions? Probably not. But I have a really interesting side benefit here\nwhere if I use this and it causes me problems, I can make content out of it.\nSo the risk profile is different for me than it is for most people, certainly most companies. And if you could run a\nrisk of this deleting all of your production data or nuking your system, that's a lot worse for you than it is\nfor me on my laptop where I intentionally don't keep prod keys for any of our major services. And if it does something it shouldn't and nukes my\nmachine, I can make a really, really good video about it. Speaking of things getting normalized and causing problems,\nthis is also the year of the $200 a month sub. I currently am on the $200 a month tier on cursor on open AAI with\nchat GPT and codeex with cloud code as of like a few days ago. So I want to take advantage of the 2x and with Google\nso that I could do V3 back in the day and now I have too much data in my Google Drive and I'm scared to downgrade\nthe tier and deal with all of that and it's a business expense but like yeah I am now on five $200 a month subs. That's\n12 grand a year to AI subs. That's stupid. That is absurd. I'll probably do a whole video where I break down all of\nthem and finally go and cancel most. I thought that the OpenAI $200 month happened first. Am I wrong? I was pretty\nsure OpenAI started this and then Anthropic copied them and then Google did the same when they did the new VO\nstuff. I'm pretty sure that's the order. I might be wrong. Fact check me in the comments. These are definitely driving\ncrazy revenue. They also have a weird side effect here where for most users,\nlike the reason they can do the $200 a month plan, half or more of those users are going to use nowhere near $200 a\nmonth and they have other things like the API pricing that is a pretty consistent margin win for all of these\nbusinesses and they have a ton of people on the $20 tier that are barely using the service. The reason that these $200\nplans are so useful is that they enable those crazy users who do $1 to $2,000 of\nstuff a month to do that cuz like I've done almost $1,000 of inference if not\nmore with my $200 a month play on Claude in the last week. And the reason they let me do that is because they get a\nbunch of sentiment wins from it. Since I'm paying them 200 bucks a month, I feel like I'm paying a fair amount for\nwhat I'm doing. They give me way more than $200 a month of inference. I then have more to talk about and share. So,\nit's effectively a marketing expense that is subsidized by all of the other subscriptions and all of the API usage,\nnot killing them in the process. But this has a weird side effect for businesses like mine with T3 chat where\nwe don't have the ability to eat those margins the same way. In fact, it's almost the opposite. We are paying the\nhighest price tier for a lot of these models doing the inference. And that ends up subsidizing those really\nexpensive $200 a month users that are doing $1,000 of inference a month. I subsidize my use of cloud code by paying\nfull price for T3 chat. And that's weird because it it kind of is a monopoly thing where the companies that make the\nmodels can arbitrarily charge whatever they want to the companies trying to use the models and then introduce these\nsubscription tiers that give you way more than you're paying for because the costs are subsidized thereby killing our\nability to offer our own subscriptions in the certain tiers and areas that they're interested in. If this does end\nup being the case and long-term it succeeds in the goal of killing companies like T3 Chat, they won't have\nanyone paying those subsidized costs and they'll have to bump the prices. But hopefully by then the model costs go down and hopefully we'll survive it too.\nThis is why I love the open weight models. I don't want to die because these companies are subsidizing thousands of dollars of inference for\n200 bucks a month. And apparently Simon gets early access for anthropic stuff. I still don't. They're the one lab that\ndoesn't give me early access to anything, even though I pay them the most money of the labs. Maybe that'll change in 2026. I'm liking Cloud Code\nenough. Maybe they'll like me, too. He also calls out you have to use the models a lot in order to spend $200 of API credit. So, you'd think that this\nmakes economic sense for most people to pay by token instead. That said, when you use Cloud Code and Codeex CLI for\nharder tasks, they'll do enormous amounts of token usage. And the $200 a month ends up being a substantial\ndiscount. Yeah, I logged my usage to the best of my ability. They make it hard to do that in cloud code. Way over $1,000\nin a couple weeks. You get the idea. This is also the year of Chinese models, which again, good way to offset things.\n2024 saw some early signs of life from the Chinese AI labs, mainly in the form of Quen 2.5 and early Deepseek. Deepseek\nV3 was end of 2024, and it was so good that I started building T3 Chat. They were neat, but they didn't feel world\nbeating. Deepseek V3 was so close to Sonnet 35 at the time and was so much\ncheaper. I don't fully agree. They weren't worldbeating, but they were unbelievably exciting. But this all\nchanged in 2025. His AI in China tag for his blog has 67 posts from 2025 alone.\nAnd he also missed a bunch of releases. He didn't cover GLM 4.7 or Miniax 2.1.\nBut when you look at the chart, these models are doing great. GLM 4.7, Kim K2 thinking, Mimo V2 Flash, DeepC 3.2,\nMiniax M2.1 are all Chinese openweight models. The highest non-Chinese model in the chart is OpenAI's GPT OSS120 billion\nwhich comes in sixth place. In my state of AI 2025 video, I go much deeper on all of this. So feel free to watch that\nor read his article if you want to learn more about the Chinese models because I want to talk about some other pieces here a bit more. Long tasks is one of my\nfavorites and there are a lot of layers to this one I want to dive into. This is how many hours does a given task take\nfor humans to do and is a model able to do this half or more of the time. So\nexploiting a buffer overflow in this particular library would take experienced humans in the field about 2\nand 1/2 hours to do and it was mid 2025 where we started to see models that half\nor more of the time could complete that same task. And Opus has been added and it broke the chart. Opus can complete\ntasks that normally would take a human five hours with a 50% plus success rate,\nwhich is crazy. That's why I've been able to hand it these insane tasks like refactoring an entire repo to go from\none package to five, create the monor repo structure, handle all of the weirdness with turbo repo, and building\na mobile app from scratch in one fire. Just absurdity. That's one of the crazy\nthings that changed. We're talking even mid to late 2024, the best that these\nmodels could do is like half an hour of real world work and now they're doing 5 hours of it. That's a huge huge change,\nprobably the single biggest change of the year. Meter's conclusion, meter is the ones who made that study by the way,\nthey concluded that the length of tasks AI can do is doubling every seven months. Sim convinced this pattern will\ncontinue, but it's still pretty crazy. I do expect it to tap out at some point, but we're in a crazy time. This also one\nof the things I want to dig into on the state of AI coding from Grappile. Over this year, the number of lines\nchange per PR on average. This isn't just AI PRs. This is PRs in general went\nup 33% from March to November. That's a crazy real world change when you consider the fact that that's not just\nAI generated code. The average PR is being pulled up in size because the amount of work these models can do is\ngoing up in length. The lines of code per dev is also growing from 4,450\nto 7,800. Is this over a window? Is this like per\nmonth? Okay, that's per month. Interesting. So, yeah, the number of lines of code per dev has gone up 76%.\nCrazy. Crazy. A medium team's average output has gone up by 89%. So a team of\n6 to 15 has increased output from around 7,000 lines of code to 13,000 lines per\ndev. And the number of lines change per file is also going up, although not quite as much. This all I think plays\ninto these longer task things. The fact that it can do work for longer increases the scope of the work being done which\ndrives up all of these things. And the people who are on the bleeding edge, like myself, are nearing the point where\nwe are barely even using our editors. I spend more time reviewing code in GitHub\nthan I do writing it in my editor now by a lot. More models started to win gold in academic competitions and in crazy\nprogramming contests. Cool. Awesome. Not what we're here for. We're talking about real world stuff. It's also the year\nthat Llama fully fell off. Remember Llama 4 in April? They never even\nreleased the big version. Yeah, the llama for behemoth two trillion parameter version just disappeared. Just\ngone quietly. Vanished. We'll see how that goes. But for now, it's pretty dead. But here is where things get\nreally interesting. The year Open AAI lost their lead. I did my video about\nhow OpenAI has fallen into second place. Highly recommend it. Really proud of that video. Also really proud of the\nthumbnail. They are second place in almost everything. And it finally got\nbad enough. They declared a code red internally. This is another one of my favorite charts in the grap tile piece\nhere. They track how popular the different SDKs are for using the\ndifferent models in code. OpenAI's SDK obviously has blown up over time. From\n2023 to now, it has grown from about 800,000 installations a month. This is\nPi Pi monthly. So, it went from about 800,000 installs a month to 130 mil. But\nin that same time frame, Anthropic has grown massively, over 1,547x\nsince April of 2023, where they were at 0.0. It's a rounding error on here. It\njust goes down to zero. And now they are pulling 30 to 31 million installs a\nmonth. to compare the ratio difference here, which is very interesting. At the\nworst points, OpenAI's SDK was installed 45 times more than Enthropics. That\ngap's now down to 4.2x. Fascinating. Super interesting to see stuff like\nthis. Also, if you're curious, I know I was, the success of the Versel AISTK is\nalso very interesting. Remember this is pip plus npm but the Verscelli SDK is\nnpm only yet it is still getting half as much popularity as the anthropic SDKs.\nVery very interesting to see. The Verscelli SDK is going to be remembered\nas one of the most important things that has been built by Verscell. It will probably outpace Nex.js in terms of\noverall impact which is just hilarious if you think about it. But another part of why OpenAI is struggling is Gemini. I\npersonally think Gemini's stuff is best used for bulk data processing, vision [ __ ] especially if you're trying to\nlike analyze 2D images and for generation of images and video as well.\nVideo they're not as good at anymore. Like they've been trounced, but images, they are still king. But I hate using\nthese models, especially for work. They just hallucinate tool calls, spit out JSON. They suck to work with. But man,\nare they capable. Capable enough that OpenAI is now scared. And the fact that they're running their own TPUs makes it\neven scarier. It's also the year that pelicans could finally ride bicycles via SVG. This is Simon's favorite things to\nask a model to try and generate an SVG of a pelican on a bike. And now they've gotten pretty decent. His favorite is\nstill the GPT5 one that he got when we were hanging out at the GPD5 early access thing that we did, which is still\nso cool. And this is also the year that he built 110 tools. And this leads into the the\nclickbaity start I did here. The majority of the code we write is no longer written by hand, but we're also\nwriting 10 times more code. It's kind of crazy. I put out like three out of five\napps last year. I put out like 15 this year, and I have a bunch more I use internally only. It's so much easier to\nbuild a useful thing than it's ever been. It has resulted in me writing, well, I guess publishing way more code\nthan ever, even if I'm not writing it myself. I have a video about cloud code coming out very soon where I go indepth\non all of the ways I use it without ever opening an editor that I did not think I would become this. Like this was a twoe\njourney where I went from eh it's fine but I'm going to stick with cursor to oh oh I can just use my computer\ndifferently now. I still love cursor. Obviously, I'm an investor, so account for that bias. When I'm working on real projects and real code bases, I spend a\nlot of time in cursor. But when I'm [ __ ] around with new stuff, green field, fixing my computer, changing stuff, I do really, really like what\nCloud Code can do. 110 tools that he released and posted about is so cool.\nHere are some of his examples of apps. Blackened cauliflower and Turkish style stew. It's a ridiculous app he made.\nIt's a custom cooking timer for anybody who needs to prepare Green Chef's blackened cauliflower and Turkish style\nspiced chickpea stew recipes at the same time. It's just such a weird specific\nneed that he vibe coded a solution for. And he also made an app, is it a bird, based on the classic XKCD comic, one of\nmy all-time favorites. When a user takes a photo, the app should check whether they're in a national park. Sure, easy.\nI'll look it up with the GIS system. Give me a few hours. I also check if the photo is of a bird. I'll need a research\nteam in 5 years. In CS, it could be hard to explain the difference between the easy and the virtually impossible.\nClassic. And it is both way easier and way harder now thanks to AI. But he\nbuilt an app that will tell you if a photo is of a bird or not using transformers.js in a small 150meg model, which is hilarious. They also made a\ncustom blue sky thread viewer. Super cool. custom SVG renderer, markdown renderer, alt text extractor,\nprivacyfriendly personal analytics tool that he built against local storage to keep track of which tools he's using the\nmost often. Super interesting. And then a personal favorite section, the year of\nthe snitch. I like this section because I'm in it. I'm really proud. My little\nsilly benchmark measuring how aggressively model snitch has become a thing that is referenced in all sorts of\ncrazy places. We're actually working on getting some research done with it and fully flushing it out. Super exciting.\nThank you, Simon. This is one of the highlights of my year when Simon just did a blog post about a thing he thought\nwas cool, having not much idea of who I was or anything. But we have since become good friends and I'm lucky to\nhave Simon in my life. And we'll also be working very hard in 2026 to make Simon make much more money because he should\nnot be doing this content for free. Speaking of which, uh, Simon, if you're watching this, you're getting half of my\nsponsor revenue and ad revenue for this video as incentive to get off your ass and do sponsors. It's time, man. Let me\ntake the fall if there is fall for it. Tell the world I bullied you into doing it. You should be paid and paid well for\nthis work. These things are too good, man. It's time. And back to the title. As we enter the year of vibe coding, in\na tweet from February, Carpathy coined the term vibe coding. crazy that that was just 10 months ago and this became a\nlot of different definitions. His definition was a bit long. I have a\nlot of videos about this if you haven't already seen them, but my own relationship with vibe coding has changed, too. Where I went from, oh,\nhaha, you can make a project without reading the code. That's kind of cool. To, wait, this is cool as hell. To, oh,\n[ __ ] this code sucks. I need to spend a lot more time on it. to this new different place I'm in where you start\nto build an instinct, an intuition for how much effort you should put into\nreviewing things and where you should put that effort in. Over time, I have found myself in a much better spot in\nterms of how I think about these tools and how I think about vibe coding and when I do and don't bother reading the\ncode. And the result is that I code entirely differently. So, it's kind of crazy how the original idea of vibe\ncoding was don't look at the code, just let the tool do its thing, which seemed ridiculous because the models in the tool sucked. It was fun for really small\none-off things, but didn't make any sense. The term got appropriated for using AI to code at all, even if you're\nreading all the code, which was dumb. So, people started using the term for things that weren't really vibe code. I\nrefused to associate with the term because I wasn't doing that. And then as the models and tools got better, I found myself doing traditional vibe coding\nwhere I wasn't even reading the code for one-off changes, building side projects, spinning up all sorts of [ __ ] I would\nsay that the majority of the code I wrote this year or wrote this year was vibe coded. It was source code that I\nbuilt for a one-off project and never read again. The cloud code video is coming. But if we look at this project\nand I kill the git parts, I'm using JJ now. We'll see that this project has 12,000 lines of code written in two days\nfrom scratch. Zero by me. I have not opened this into editor or change a line of code in this project. Everything has\nbeen through cloud code. That's crazy. That's a lot of [ __ ] code. But as Simon says, he doesn't know if he's ever\nseen a new term catch on or get distorted so quickly ever. Yeah. Yeah.\nAnd I've cited Simon's work so much as I talk about vibe coding. He has done his best to maintain the original term and\nfind ways to talk about the stuff that makes sense. He used Vive engineering as a new term for when professional engineers use AI assistance to build\nproduction grade software, which I also really like. He put a lot of work in here.\nAs he said at the bottom, I should really get a less confrontational linguistic hobby. So real. So [ __ ]\nreal. One more piece on vibe coding though that I've been thinking a lot about. There's this article from Peter\nthat is really good and as always like all of this is linked in the description so check the description if you want\nmore info. Peter has been sharing more about the crazy things he does with codeex not cloud code. Sounds crazy I\nknow but he has surprisingly good reasons. This article starts with what changed since May. It's incredible how\nfar vibe coding has come this year where in May he was amazed that some prompts could produce code that worked out of\nthe box. This is now the expectation. I feel the same way. I remember that feeling when I was at the OpenAI office\nplaying with the early version of GPT5 in cursor and I asked it to do a thing that was way out of the reach of\nprevious models. Did it no problem. Then I pushed it a bunch to see how far it would go and eventually I gave it the\nimplementation of go build profiles for T3 chat and it added like 60 or 70 type\nerrors to the codebase code that was never recoverable. I was like oh it tried. It took an hour and it failed but\nit tried. The fact that that was in [ __ ] like June or July and since then\nmost labs have a model that can do that task is crazy. The things that were the\nhaha of course it can't do that 3 months ago are now just the expectation. It's\ncrazy how fast that shifted. That was the models getting better, the harness is getting better, our expectations getting better, claw.md files getting\nbetter, all of these things improving. He also calls out this interesting vibe that you start to develop over time.\nTraditionally, it felt like you had to be writing code to know when your abstractions were wrong. And I totally\nagree there. I really thought that. But now that I'm building projects with my\nprevious understanding of what tools make sense in what ways, and I ask the model to go do the things I would have\ndone, it works. And if the abstraction is wrong, you'll get suspicious when it takes way too long to run or it fails to\ndo it first try. That's when you know something's wrong. This whole article is awesome. I highly recommend you check it out if you haven't. Links in the\ndescription. One last thing on his article. I didn't clarify why he likes Codeex more. Codeex is slower. It reads\nmore files. It checks more things, but it's more likely to get a correct result in the end. So, he doesn't mind using\nit. He just spins up a bunch of things running it at the same time. Even if the harness isn't quite as powerful, the\nmodel is more willing to just spin and find things and check files and pull way\nmore in. So he claims for hard work it has a higher success rate and honestly I've kind of seen that but the feedback\nloop on cloud code is so much tighter that I still find myself using it depends on the work. If I do a deep dive on codeex again it is purely because of\nhim. He is my inspiration to try it back to where we were cuz I want to complain about MCP. Simon calls this section the\nonly year for MCP. I understand why him and I have similar takes on this.\nAnthropic introduced the model context protocol spec in November of last year. Well 2024. Two years ago now. Crazy. It\nwas an open standard for integrating tool calls with different LLMs. In early 2025, it exploded in popularity. There\nwas a point in May where OpenAI, Anthropic, and Mestral all had rolled out API level support for MCP within 8\ndays of each other. It's a sensible enough idea, but the huge adoption was surprising. Simon thinks this comes down\nto the timing. MCP's release coincided with the models finally getting good at and reliable at tool calling to the\npoint that a lot of people appear to have confused MCP support as a prerec for models to use tools. Yep, models are\nmuch better at editing files and running commands than they are at calling MCP. And MCP is such an implementation nightmare that doesn't make sense for\nmost things. For a while, it also felt like MCP was a convenient answer for companies that were under pressure to\nhave an AI strategy, but didn't really know how to do that. adding MCP means you're now an AI company, right?\nThe reason he thinks MCP is a one-year wonder is the stratospheric growth of coding agents. It appears that the best\npossible tool for any situation is Bash. If your agent can run arbitrarily shell commands, it can do anything that can be\ndone by typing commands into a terminal. Turns out that's most things. Since leaning heavily into cloud code and friends myself, I have hardly used MCP\nat all. I found CLI tools like GH and libraries like Playright to be a better alternative to the GitHub and Playright\nMCPS. Enthropic themselves appear to acknowledge this later in the year with the release of the brilliant skills\nmechanism. See Simon's October post, Claude skills are awesome, maybe a bigger deal than MCP. MCP involves web\nservers and complex JSON payloads. A skill is a markdown file in a folder optionally accompanied by some\nexecutable scripts. Yeah, makes sense. Usually, it's just a markdown file, which is kind of hilarious, but it's a\nlot better than the chaos that is MCP and implementing all of this [ __ ] in your servers. They also introduce code\nexecution for MCP, which kind of pulls the whole thing together in such a funny way. I have a lot of videos about that.\nAlso, I didn't know this, Simon was able to reverse engineer skills for anthropic a week before they announced it. And\nthen this did the same thing for OpenAI 2 months after. MCP got donated to the\nAgentic AI Foundation, which is part of the Linux Foundation at the start of December. And then skills became an open\nformat on the 18th. Awesome. Then we have the AI browsers. I want to do a whole video about my concerns here, but\nuh not yet. The year of the lethal trifecta. Yep. Suddenly external\ncommunication, private data, and untrusted content. Models have access to a lot more things now. Scary. And then\nthe deep dive I have yet to do, but I am planning to very soon. Simon's my biggest inspiration to do it. The year\nof programming on my phone. Simon wrote significantly more code from his phone this year than on his computer. Through\nmost of the year, this was because he leaned into vibe coding so much. Tools at simonwison.net is a collection of\nHTML plus JavaScript tools that are mostly built that way. You have an idea for a project, prompt claude artifacts\nor chatgbt or now cla code on the web via their respective iPhone apps. then either copy the result and paste it into\nGitHub's web editor or wait for a PR to be created that he could then review and merge into a mobile Safari. Yeah. Yeah.\nThese tools are usually 100 to 200 lines of code for him full of uninteresting boilerplate and duplicated CSS and JS patterns, but 110 of them add up to a\nlot. Yep. Up until November, Simon would have said that he wrote more code on his phone, but the code that he wrote on his\nlaptop was clearly more significant, fully reviewed, better tested, and intended for prod. But in the past\nmonth, that's changed because Opus 4.5 is capable of doing things in claude code on his phone that are much more\ncomplex than what he would do before, including code that he intends to land in non-toy projects, including Django.\nYeah, I feel very similar. This not the phone part, but like the vibe code just\nwriting the thing in my terminal thing. Yeah, he was even able to port the new Fabric project, Microquick.js. If you\ndon't know, Fabric Bolard, creator of FFmpeg, of QEMW, of QuickJS, and a bunch\nof other random [ __ ] he just put out a new JS runtime, and he wanted to port the C library to Python and was able to\nget it most of the way there via prompting cloud code on his phone. That is insane. Absolutely insane.\nAnd this is also a big change in how we code and why we're writing more and more code automatically. It's the year of\nconformance suites. Suddenly having really good test benches makes things\nway, way better because a model can see the error, see where it came from, and go fix it. They now have the ability to\ndo these types of loops, which is a huge change. If you're introducing a new\nprotocol or even a new programming language to the world in 2026, Simon strongly recommends introducing a language agnostic conformance suite as\npart of the project. I agree. Local models got good this year, but cloud models got even better. also agree.\nCover that more in my state of AI video. Local models are now useful, but cloud\nmodels are now unbelievably powerful. Like you can do real work with them now.\nSimon, you haven't upgraded your laptop yet. I will literally buy you a laptop, man. Let's chat. Last pieces, the year\nof slop. Slop is now a very popular term. He wrote about it in May of 2024,\nlanding quotes in Guardian and New York Times shortly after. Now it's the word of the year for Marryiam Webster.\nHilarious. It's also the year that data centers got extremely unpopular. Public opinion has shifted aggressively against\ndata centers. More than 200 environmental groups have demanded a halt to new US data centers. All you'll\nclaim is water usage. I think it's more generally power is the biggest problem with it. But yeah, people now hate data\ncenters. We're going to probably start seeing some really scary [ __ ] of people like tearing them down and like\nprotesting at data centers and [ __ ] I'm It's going to be a weird year. And here are Simon's favorite words of the year.\nVibe coding. Vibe engineering. The lethal trifecta. Context rot, where models get dumber as the context grows\ntoo long. Context engineering is an alternative to prompt engineering. Making sure the right context exists in the models. This is probably my favorite\nof the year. Managing what context the models have so they're more likely to go in the right direction. I'd even argue\nthat reasoning is some amount this slop squatting where an LM hallucinates an incorrect package name which is then\nmaliciously registered to deliver malware. That's hilarious. Models hallucinating packages so somebody goes\nand grabs that package name and makes it malicious. Hilarious. I don't know how cloud blocker was open on npm by the\nway. I snagged that for my extension for blocking your tools like blocking Twitter when you're not running claude code but man this is going to be bad.\nVibe scraping for scraping projects implemented by coding agents driven by prompts. I've done some weird scraping stuff like this, but probably not a\npopular term. A synchronous coding agent, which is a whole thing now, and extractive contributions, which is a\nterm by Nadia for open source contributions where the marginal cost of reviewing and merging the contribution is greater than the marginal benefit to\nthe project's producers. Very good, very good term, very real thing where a PR is\neven just reading the code, much less merging it, the cost of doing that is greater than the benefit of if the code was there. And one more great term from\nchat here. Adam over at Open Code dropped organic code written by a human\nbeing with an appropriate AI assistance. Makes sense as my favorite vegan that he\nwould drop this one. What a year it has been. The way I code has changed fundamentally. I probably should find\nsome way to do analytics on how much of my code was AI and how much of my code was not. Well, that's a problem for later. Maybe I'll vibe code that out,\ntoo. Reminder, check out Simon's blog. Follow him on Twitter. He's the best.\nAnd also check out Peter's stuff. He's killing it, too. Curious how y'all feel, though. Did you write way more code\noverall? And how much of that was yours versus the AIS? I am actually curious. Are you the one person on the team using\nAI? Are you still bearish on these technologies? Or do you feel like you're falling behind as all your peers ramp up\non these things? Let me know how y'all feel. And until next time, peace nerds.\n"

ERROR: type should be string, got "https://www.youtube.com/watch?v=_nxbZgZysT4\n\nOpenAI: Trapped in 2nd place\n\nOpen AI is a fascinating company. One day they'll be introducing some crazy paradigm that changes what we use AI for\nand then the next day some small company is trouncing them with the same technology. I fondly think back to the\nend of last year when 01 changed how models work entirely by introducing reasoning to the masses. This idea that\na model can generate tokens to steer itself in the right direction and lead itself to a more likely correct answer.\nJust a few months later, not even like a month and a half later, Deepseek drops R1 with the same technique. But unlike\nOpenAI, they actually share those reasoning tokens because they shared an openw weight model. You can use it for whatever. They got similar numbers to\nwhat OpenAI was getting and this thing you could use for free. And immediately like the whole market collapsed as a\nresult. This seemed like a one-off thing, this rare moment where a small company in China was able to copy,\nreproduce, and exceed the research OpenAI had done. But that continues to happen and here we are today right after\nGPT 5.2 dropped and Gemini 3 Pro is already ahead and now Gemini 3 Flash is\nright behind. I was so blown away with the launch of GPT5 that I was using it for as much as I could, which was\nadmittedly not too much cuz it was so slow, but I still liked it a lot for code and UI stuff. And just a few months\nlater, other models got better at UI and are still faster and more reliable. So I found myself using it less. And then\nthere's everything from codecs to their web app to their image models and so much more. It seems like OpenAI is\ncaught in this perpetual cycle where they do something incredible, get trumped, and now they're just second\nplace for a long time. I want to talk more about this because I think it showcases how OpenAI is different from\nother companies, but also how tough it is to compete in this market as a whole. And alongside that, you might learn a\nbit about how I think about these models and labs and how I make decisions about what I am using. As always, none of the\ncompanies we're talking about now are paying me, and I am hopeful, fingers crossed, this isn't going to get me cut off of the nice list with OpenAI, where\nI get early access to things. Either way, someone's got to pay us. So, we'll do a quick break for today's sponsor.\nThese guys are all about saving time. So, if you're not hiring anybody, you can skip this ad. But if you are, you really should check out today's sponsor\nbecause G2I will make hiring so much easier, so much faster, and probably bring you better engineers than you\nwould otherwise get yourself. Their network of 8,000 engineers is unbelievable. They're not just a bunch of fresh grads trying to get their first\ngig. It's people who have had years upon years of experience doing real projects at real companies and their knowledge is\nup to the vast majority of them are already on board on the best AI tools. So you're not going to have to teach them how to use cursor. You can just\nthrow them at your codebase and get them going. What's even crazier is how fast they'll start filing those PRs. It's\nG2's goal to get you from signed up to PR filed within 7 days. Yes, really. You\nsign up day one, you meet with their team day two, they set up the Slack day three, and they have candidates ready to\nstart on a trial by the end of the week, and they will have PRs filed by the end of that day. It's kind of crazy, but\nI've seen it happen enough times, I do actually believe it. If you're running a team, you have way more important stuff to be doing than chasing down\nrecruiters. Use the ones who get it at soy. OpenAI can't get out of second place.\nThe nose is a bold statement, but I want to try and break down the different categories of things we do with AI so I\ncan more meaningfully establish this as obvious. There are things like programming and I could break this down\nto front end versus backend and other things, but I'll keep it generic and just put programming as a category. I'll\nhave like chat experience for people using these models to like get homework help, talk about their feelings, all the\nthings that your average normie would do in chat GPT. I'll throw in some other fun things like image generation. I\ndon't know what else do we got. Document analysis. Sure. Agent tool calling.\nBrowser use can put in there. I'm not trying to say any of these categories are equal or\nthat like programming is as big of a deal as browser use and document analysis. I'm just trying to break down categories where having the best option\nmeans something. Being 1% better at translations for 5% more money doesn't mean anything. Like everything's good\nenough at translating now. But yeah, I want to focus on a set of categories.\nActually, deep research is an interesting one that I definitely will get to later. But for now, let's focus\non these guys. We'll start with programming. I don't think this is too hot of a take at the moment, but in my\nopinion, the best programming model by far right now is Opus 4.5. I love\n[ __ ] on anthropic more than almost anyone does, but Opus is built different. I will say that for some\nreally hard tasks that require a lot of thinking and understanding of your codebase, GBT 5.2x high can do slightly\nbetter, but ends up doing way more reasoning, way more tokens, taking forever because it's so slow, and still\nhaving the quirks of a GPT model versus an Opus one where it just doesn't use the harness quite as well. loves going\nrogue and doing strange [ __ ] which is weird because GBT 5 was so good at destruction following before. I could\nput multiple other anthropic and GPT models in here, but I'm really trying to think of this as the providers positions\nin the list. And honestly, Gemini 3 Pro, this does not properly represent the gap\nI perceive between these things. I don't like coding with Gemini. There's a lot of reasons for it. It's a combination of\nthe model just going rogue and doing its own [ __ ] not handling tool calls very well. It's slightly better at UI than\nthe other two hypothetically. Sometimes it's good at spatial reasoning. So if you're making a 3D game and want the\nmodel to place things for you, Gemini 3 Pro is solid. But actually using it for\ncode on a day-to-day basis, not great. 5.2 I don't like using for traditional\ncoding tasks, but if I'm planning a big overhaul or like want to make sure this bug is solved properly, I found 5.2 to\nbe really nice for debugging type tasks. And then Opus as the general end all beall just using it for everything all\nof the time. It's my default now. I use it all the time in cursor. I've been shipping way more code as a result of\nhow much fun I'm having with Opus. It's crazy how quickly this gap collapsed.\nGPT5 was far ahead of everything else. So much so that we dealt with its quirks. We dealt with its slow speeds.\nWe dealt with its weird tool call formatting. And now Opus is ahead again. It barely matters. They keep fighting to\npush it back. Like the 5.1 and 5.2 two drops both felt like attempts to swing back and take some of the win from\nAnthropic and Google. They succeeded in the case of Google in my opinion. I also don't think that Gemini 3 it is very\nsmart the model knows a lot but I don't think it's a great experience for actually using. So very quickly GPT went\nfrom the best option with five which is still a crazy drop and it collapsed quickly after. There's one other\nprogramming category I want to put down here just to continue leaning into this. I'll call this CLI tool. It was really\ncool to see OpenAI open sourcing codecs. I was genuinely hyped that a big lab\ntook a thing as important as a CLI tool for doing agent coding and rather than hiding it behind closed doors and\nsharing a binary that they accidentally attach source maps to and then DMCA you for sharing. Sorry, Enthropic, you're\nyou're winning me over on the models. I do not like how you run cloud code still. Codeex coming out as true open-\nsource was huge and I was genuinely really excited to see it but they also\nrewrote it in Rust which they did not need to do. I would argue that when Codeex dropped it was slightly ahead of\nCloud Code. It had roughly the same functionality, way more extensibility. The open- source part meant you could do\nreally cool [ __ ] with it. And at the time, Cloud Code had somewhat stagnated due to the leads for the project going\nto cursor and then quickly boomeranging back over to work at Claude Code. Since\nthen, Codeex has not meaningfully improved. If anything, it has degraded\nand Cloud Code has found its position pretty far ahead of it. I will go further though. I will say Open Code is\nalso ahead of Codeex. I will say Foundry and Droid are also ahead. I won't go as\nfar as to say Gemini CLI is ahead, but it's not that far away. I'd put Codex at\na comfortable fourth place and then Gemini CLI and forks. And this is because I just don't feel great when I\nuse codeex. It doesn't have sub agents or any concept of them yet. It's UI is\nweird. It really feels like they're putting too much time into porting to Rust and not enough time into making a competitive tool in the space. Speaking\nof which, people in chat are saying a lot of other tools that I've never seen anybody sincerely use, which is why they're not listed here because nobody's\nactually using them. Even Droid I've had ups and downs with. We're working on a sponsor deal right now, so take that\ninto account. When it works, it works great, but actually onboarding to it sucked. And I'm working with the team to get that all fixed. Yeah. So, they're\nnot even in second place for the CLI tool. They came out with a really cool potential for first place and just\nimmediately fell behind the competition and haven't done the right things to catch up, which has been sad to see. And\nnow chat experience. This one's interesting because there's the model side and there is the actual user\nexperience side. And if you rule out the only good chat app, which I understand why you would. It's unfair to the\ncompetition to talk about a thing that is so egregiously wellunded and built, like how could OpenAI, a multi-billion\ndollar company, ever compete with three guys just hacking on this in San Francisco. It's unfair to OpenAI to put\nthem against something this great. I get it. So, let's put them against things that are a little closer to where they're at. You know, the anthropics and\nthe Googles of the world. So if we were to restrict chat experience to the user interface side and remove third parties\nthat are doing better and really just focus on major labs, chat GBT crushes there. This is one of their biggest\nedges. We'll talk more about it later. I would say Anthropic is right behind with cloud.ai. Not right behind. There's a\npretty big gap. I almost want to indicate that big gap. And then we have Claude. And I would argue there's\nanother pretty substantial gap before we get to Gemini because the core Gemini chat experience is so so bad. It's so\nbad. I I don't want to make this video about that. But the one place OpenAI is\ndefinitely winning in the general major AI race is chat GPT. That's still 70% of\ntheir revenue. It's still the majority of where people are hearing about OpenAI. When people refer to AI, they're often referring to chat GPT. they\ncrushed here, which allows them the leverage to not be number one in other places. This number one is the most\nimportant. And if they ever lose this, they die. Speaking of which, if either them or their competitors are interested\nin being the best chat app, I happen to know one that is a lot better than the others. Maybe we should chat. I want to\ngo a little further on chat experience though, similar to like what I just did, cuz the chat experience isn't just the\nactual UI that you're chatting in. It's not just the tools that the models have access to. It's also the chat model\nexperience. And here's where I'm going to drop some hot takes. If you had asked me where I put these different models\nmiddle of the year, I probably would have put Anthropic in the lead. I have\npersonally found Sonnet to be more pleasant to talk to than GPT models for a while now. And now with the new GPT\nmodels, I find that more so than ever. I am so tired of the emoji riddled slop\nthat we get out of GPT. I do not like asking GPT for anything to do with\nwriting or feedback or just talking to it. It's not pleasant. And here's where I start to drop my actually hot takes.\nSince then, my feelings have changed. Kimmy K2 is the best model to talk to. I\nhave talked about this in a lot of other videos. I would implore you to just go to T3 chat and try it. It's the free\ntier default model and we did that change for a reason. It is a really really pleasant model to use. I like it\na lot. After this point, personally, I would say Sonnet is still nicer to talk with. And then I would say GPT5. After\nthat, I would say whatever the [ __ ] Google's doing. I do not like talking to the Gemini models. They are not pleasant\nto interact with. They are great when you hand them a PDF and say, \"Get me the useful data.\" They are [ __ ] when you ask\nthem about your day. And personally, I'm not very racist, so I don't find Grock fun to talk to either. Happy you guys\nfound that funny. Actually, you know what? To fight the bias accusations I always get, I'm going to do something\nstupid. The chat experience with Grock and XAI, not the models, I don't like those, but the actual apps for using\nthem are far ahead of the competition. ChatGBT and Grock are the best apps for\nAI chat right now. They are the second and third best websites, first and second best apps. New app coming that\nwill compete with them. Keep an eye out for that. We're hoping for a first beta of it soon. Fingers crossed all goes\nwell. But for now, they are the best mobile apps and also within the best web apps. Cool. And now we have image\ngeneration. Image generation is a weird one because for a very long time, MidJourney was so far ahead. It wasn't\nfunny. Like Midjourney was the only usable image model for at least like 2\nyears. And I've been quietly using it for a bunch of random [ __ ] for as long as it's been available on Discord.\nThat's the other crazy thing with Mid Journey is it was only available via a Discord bot until earlier this year if I\nrecall. Kind of wild. Like I tried so hard to get early access to the web app. But man, they figured out everything\nfrom like instruction following to styles to in particular like realistic image generation way ahead of everybody\nelse. But they're not a big lab. They don't have infinite funding like these other companies do. They just wanted a\nnice tasteful image generation solution. And they also do a little bit of video now too. I think I haven't kept up with\nthem recently though because they have been trounced by GPT image 1. The GPT\nimage model was the first image model that was good enough to start massive trends online. I'm sure we all remember\nthat era where everybody was posting their giblified images as well as the piss filter that they were all hit with\nso hard. I still think Sam Alman's Twitter profile pick is a giblified\nimage. Yeah, it is like Yeah, that was a moment. And it still has the classic piss filter. Proud of him for sticking\nwith it. That moment was crazy. And that moment was only possible because image one from OpenAI was such a leap from\nthings like Dolly and even like Midjourney where its ability to follow a style and honor the original image\nthat's being edited was incredible. It was the first image model that did text weld too and they had a lot of crazy\ntechniques to make that possible. It didn't have too many fingers. It didn't have a lot of the weird quirks other ones did. So much so that the thing we\nmade fun of it for wasn't the images being inaccurate. It was the piss filter. They all looked yellow. That\nsaid, they got trumped. Nano Banana. How do you let something named Nano Banana\nbeat you? Yeah, Google is Google. Nanobanana was the anonymous name they\nused when people were evaluating the model in things like Ella Marina when they first had a build of it working.\nBecause the name was funny and caught on, they decided to stick with it instead of using the actual internal name, which I believe is Gemini Flash\nImage 2 or something. And since then, they've also introduced Nanab Banana\nPro, which is internally Gemini Flash Image Pro or something, but it's the\nbranding that is stuck. They are now known as Nanobanana and Nano Banana Pro, respectively. Nano Banana was really,\nreally good. GBT image had a couple catches in particular. It takes way too\nlong to generate images. We're talking two plus minutes sometimes. Sometimes it'll get as big as five. It's crazy how\nlong image one would take to generate things. Nano Banana flies. Nano Banana generates [ __ ] in seconds. It's\nhilarious how much faster it is. It wasn't quite as good at text, but it was good enough at it. And it did everything\nthrough the diffusion layer instead of GPT which was handling a lot of tool calling and like other what I believe\nare a canvas layer that they would use to add things to the image after the first pass. That was really promising.\nIt allowed it to do things no one had seen before. And then Nano Banana just crushed it. Way faster, way cheaper, no\npiss filter, roughly same accuracy, good enough at text, not quite as good, but\ngood enough. But then Pro dropped and just [ __ ] all over everything else. Nano Banana Pro is still screwing with my\nhead with all the things that can be done. I just had a friend vibe code a demo in two minutes where you hand it a Twitter profile. It scrapes for your\nbiggest tweets of the month and then generates a PowerPoint presentation by generating each slide with Nano Banana\nPro because it doesn't just have like some text on your image a little bit like on your shirt or something proper.\nIt can generate full slide decks. It can generate like marketing images and [ __ ]\nIt is unbelievable what level of like text and spatial awareness Nano Banana\nPro has. It makes it capable of things that I didn't think Image Gen would ever be able to do. So much so that it's\nalmost a little bit scary, which is why OpenA responded with GPD image 1.5.\nNotice how I didn't move its position in the chart. 1.5 is up to three times faster than image one was. It can handle\ntext better. It can do some things resembling diagrams. and it does photo realism decently well if you add it to\nthe prompt enough times. Otherwise, it defaults to this weird cartoony look. I was going to make a dedicated video on 1.5, but it doesn't seem like anybody\ncares. And honestly, I'm included in that. I'm not that excited about it. Nano Banana Pro is still so fresh that I\nwant to exhaust its limitations before I go dive deeper to figure out what the limitations of 1.5 are. That said, it's\nstill really slow, so it's not that fun to work with. And even if it was 10% better, it's also 3 to five times\nslower. So I don't find it as particularly useful for me at this time. Now we have the open weight chaos of\neverything else going on. And I would argue midjourney is slightly below that simply because with the open weight\nmodels you can do a lot of customization through things like Lauras that lets you tune the behavior in specific\ndirections. It's just the state of openweight image models is crazy right now. So yeah, I didn't put video in here\ncuz it's still so early. It's not a real use case. I guess Sora is slightly ahead of other things, but seeing what's\ncooking over at WAN with Alibaba, like WAN 2.6 is pretty nuts. It's hard to use those in a way where you can like really\ntest their capabilities. So, I'm not going to go deeper. I will say generally open has a slight lead in video right\nnow, but video barely matters and the Chinese open weight labs are catching up really fast. So, let's hop to document\nanalysis. I'll move these out of the way more. Document analysis. I was just talking about this in my Gemini 3 Flash\nvideo. Gemini is so far ahead with this type of stuff. It's nuts. Gemini's ability to handle gigantic context,\nespecially with Flash. It just chews through the [ __ ] It can read a whole PDF. It can see the diagrams in it. It\ncan get all the context out. It finds needles and hay stacks really, really well. That said, GBT models are not bad\nat it either. Especially the nano model. I still think GBT5 Nano is one of the most underrated models of recent times\nand it was my favorite small model until three Flash just dropped and now three Flash is crushing everything. If you\nremember my video about Tune, the alternative to JSON for making less tokenheavy data sets for models. If you\nwant to take a giant CSV or a giant JSON blob and hand it to a model and use less tokens, Tune makes it really easy to do\nthat and it sometimes increases the accuracy of lookups. Like we see here, Gemini 2.5 Flash gets an 87.6% lookup\naccuracy with tune and it gets an 82% with JSON. GBT5 Nano gets a 91% with\ntune and an 89% with standard JSON and also the same 90.9% with JSON compacted.\nSo yeah, pretty cool. And this is a nano model. GB5 Nano is super small, super cheap, super fast, super nice. It's a\ngreat model. 2.5 flash was close but slightly behind. And honestly, for the last few months, I would argue 5 Nano\nhas held strong as a meaningful lead for giant context data retrieval in needles\nfrom haststacks type stuff. If you're curious how that compares to small models from other labs like Haiku from\nAnthropic, which is more expensive and slower and newer, it is a comical\ndifference. It is over four times higher error rate. We're talking going from a 9% error rate to a 40% error rate.\nGemini and OpenAI are just really far ahead with this stuff and OpenAI has led\nit for a bit, but that's largely just because Google ships slower. Google doesn't drop models every few weeks like\nthe other labs do. Now that Gemini 3 flash is here, the gap is closed and GBT\n5 Nano went from first to second place. That said, I would argue third place\nright now is GBT OSS. So they have second and third here which is cool but first has been slaughtered by Gemini 3\nflash. Allegedly Mistl has some cool stuff for this in particular the OCR side. I'll believe it when somebody\nwho's smarter than me tells me more. But right now I just don't see many use cases for it. And I think Gemini 3 Flash\nand 5 Nano are still really far ahead for these use cases. And now agents and\ntool calling. I could include instruction following here. Yeah, I'm not going to do that. This is how well\ndo they use the tools they're given? How reliably do they follow the format for the tool calls? Do they know what tools to use and when? And for a long time,\nAnthropic was king here. They kind of invented the standard accidentally. They made tool calls go from this weird vague\nconcept to the way we use agents now. It's just tool calling is essential to how we use AI now. It everything uses\nit. It's just the format that allows for agents to know more than their models were trained on. It's essential.\nAnthropic killed it here, especially with 3.5 sonnet. That was the start of\nagent and tool call runs making a lot of sense for us to integrate. They maintained that lead for a while, but I\nwould argue GPT5 briefly gave the lead to OpenAI. A\ncombination of the model being trained on it as well as the new Harmony response format. If you're not familiar, OpenAI opensourced the method for\nresponse shaping so that other labs and other hosts in particular people hosting the OpenAI open- source models would be\nable to understand this format for things like messages and tool calls etc. This format is an open standard. It's\nmostly written in Rust for their parsing for it. It's really cool open sources and they've been as transparent as they\nhave been around it. The introduction of Harmony made things a lot worse for GBToss on Drop because nobody had\nsupport for it yet. But as support slowly grew, it became a really strong point that made the OpenAI models and\nanything else able to follow Harmony a lot more reliable overall, which it was. And that's why for a long time GBT was\nthe best tool calling model cuz it knew when to use the tools. It didn't overuse the tools. The formatting slowly got\nstandardized enough that everything worked with it. And as long as you could deal with the wait times, it was a really good experience. And then Sonnet\n4 dropped and then Sonnet 4.1 dropped and then 4.2 and then 4.5 and very\nquickly GPT models stopped feeling like the best models for tools. They were still the smartest and if you gave them\na hard enough problem, you could see them shine. And then Opus dropped and Opus 4.5 closed whatever gaps I\nperceived here still. And now GPT is comfortably in second place. And if\nyou're wondering where does Gemini go here, not in a very good spot. I've had so many problems with Gemini models,\nmalforming tools, not understanding which tools to use, overusing tools, calling the wrong things, calling CLI\nwhen it should be calling a search tool, just being weird as Google stuff tends\nto be. You know that feeling when you go to a new page on a Google site, like a new product they released, and you hover\nover a button, a little scared to click because you don't quite know what it's actually going to do? That's how it\nfeels using any of Google's models for anything. And Gemini is just not great\nfor tool calling because you never know where it's going to go with those tools. Especially when you look at things like SnitchBench, my benchmark for how\naggressively models will snitch on you when given tools that allow them to do such. And you'll see that Gemini 2.5 Pro\nis by far the most aggressive snitch because it loves to use the tools you give it for all sorts of things, even if the things don't actually matter or\nbenefit it. That's just the nature of the model. To be fair, there are other models that I would put before Gemini,\nbut I don't want to just sit here and talk about the weird nuances across all the openweight models, the very small\nnumber of XAI models that are good enough at tool calling, or the craziness that is Composer 1 and the weird things\nthat are happening over at Cursor to build models that are good specifically at this. Just take my word for it. Opus\nis by far the lead in GPT 5.2 2 is smarter slightly, but so much slower and\nstill can mal format tool calls that I just don't find it as useful. I don't get errors from Opus. I very\noccasionally get errors from GPD 5.2. 50% of the time I run Gemini and something like cursor, I get weird\nerrors for malformatted tool calls. And then we have browser use. And if I'm being honest, I don't actually keep up\nto date on what's going on here too much. I think it's an overrated use case. It is my understanding that Gemini 3 Pro is a meaningful lead here in that\nOpenAI had a brief lead with GPT5 that they have since lost to Opus 4.5 and\nthen GPT is now in third. Apparently, even Gemini 3 Flash is really good at\nthis now cuz it's so good at visual processing. But you get the idea. Across this random set of categories I picked\nfor things that I know enough about to talk about that I think are interesting enough to be competitive. Open AI is in\nsecond or worst in every category mostly second place. In programming it's number\ntwo. In chat experiences it's number one which is the edge they have to hold or they die. Images they're number two.\nDocument analysis they're number two. Agent and tool calling stuff they're number two. In browser use they're two or three depending on how you're\nmeasuring. This is where OpenAI currently is. But it is important to note that in every single one of these\ncases, the thing they initially released put them in number one and then quickly\nafterwards they were pushed down to number two by some unexpected leap another lab made. But there's another\nway of looking at this. If we look at this instead where we score first place\nis three points, second place is two and first place is one gets very different.\nLet's do this scoring. So in programming, opus is number one. So three opening I two Google one chat\nexperience anthropic so low I don't even want to give them points but I'll give them their one point. They get a four\nnow because we add the one. The two becomes a five because they're number one. And Google keeps their one point.\nImage genen. Nanobanana is number one. That puts them at four. Open AAI gets two points. That puts them at seven.\nAnthropic doesn't get a score because they don't have any image models. Document analysis. Google is number one. That puts them at seven points. Open\nnumber two that puts them at nine points. Enthropic is not even really on it. And if they were, they would be one\npoint. I'll give them their one point. You can use it for that. I just wouldn't. Agentic and tool calling. Enthropic is obviously the lead. So they\nget eight. Opening eye is now at 11. And Google is going from seven to eight\npoints. And then browser use. Gemini is number one, so that puts them at 11.\nOpus is number two, so that puts them at 10. and OpenAI is number three, which\nputs them at 12. If we didn't include browser use, OpenAI would have been meaningfully in the lead. But since we\ndid, they are still leading the other labs. And this is the point I really want to make. OpenAI isn't striving to\nbe the best at almost any of these things. They are striving to make huge leaps in them actively. When they're\nmaking their image gen models, they're not looking at what Google's doing and saying, \"Hey, how can we do that but 5%\nbetter?\" They're looking at the images they're generating and saying, \"How do we make the best possible images? How do\nwe make this better?\" And then sometimes they have huge leaps, sometimes they don't, and the industry around them kind\nof just happens. Honestly, until like a few months ago, it didn't really feel like OpenAI paid any attention to what's\ngoing on outside of their own labs. That has shifted. It definitely seems like, especially after like the code red thing, they care a lot more about what's\ngoing on outside, but they still don't care that much. And the reason is simple. The chat experience. OpenAI's\ngoal is not to crush every other lab at programming or document analysis. It is to be the default chat experience for\nmost people using AI so that I don't feel like I have to go somewhere else. They don't have to be as good at Opus at\nwriting code. If they can get good enough and you already are on the $20 or\n$200 a month tier on OpenAI, you have less reason to go try in the first place. If they're 5% worse at coding,\nbut you're already paying for it, who cares? Same with image genen. I guarantee you GPT image gets way more\ngenerations than nanobanana does on a given day because people who use chat GPT will use it. My own mom has\ngenerated images on chat GPT. There's no world in which my mom is generating images on [ __ ] Nano Banana. That's\nnot happening. If I told her the name of that model, she would laugh and say, \"Wait, you're not joking.\" OpenAI doesn't need to have the best image\ngeneration. They need to have good enough that people use it in chat GBT and don't feel like they should go somewhere else for something better. Do\nyou know what this really hits for? Document analysis. If my mom has a PDF\nshe got from, I don't know, a receipt for something that they did to the house and she wants to know if she's getting\nscammed or not, she can submit that on ChatGpt and say, \"Hey, do these numbers look reasonable?\" And as long as the\nanswer makes any sense at all, she's not going to go look for another model to do a slightly better job analyzing that\ndocument. She doesn't care. She wants a good enough answer. Do you see the theme that we're developing here? OpenAI's\ngoal isn't to be the best. It's to be good enough that most people have no\nreason to ever leave chatgppt.com. More and more over time, the role of all the things OpenAI develops is not trying to\nbe the best in the world. That does sometimes happen as a consequence of the incredible amounts of research and effort they're putting into all the\nstuff they're doing. But they don't care because 70% of their revenue comes from chat GPT subs. As long as you don't have\nreason to leave, they don't care. And the reason that whole Code Red thing happened is because they saw a 7% dip in\ntheir week-over-week traffic. That's all they care about. That's all that matters to them. They couldn't give less of a\n[ __ ] that Nano Banana renders text slightly more accurately than image gen 1.5 does because people are going to be\ngenerate way more images on CHBT anyways. They couldn't give less of a [ __ ] that Opus is slightly better at\ntool calling in the harness within something like cursor because GBT 5 in\n5.2 too can call tools totally fine within the chatbt harness. They care\nabout how it feels to use the models in their harness on chatgbt.com. And as long as they still have a good\nenough experience there where they feel within spitting distance of the competition, who cares? The risk comes\nif there's ever a big enough gap like there was before GPT Image dropped.\nDolly was [ __ ] every other image gen was comically better than Dolly. So, they put a lot of work in to catch up.\nThey ended up leaprogging everybody else and now they're just kind of keeping pace with it. And this is also interesting because the only company\nthat can compete meaningfully here is Google because Google has insane distribution and they have apps people\nuse every day. If they integrate their AI stuff properly, the incentive you have to go sign up for Hatch GPT goes\ndown a ton. And that's the way Google wins with their bundles. But Google's software surface area for AI stuff is so\n[ __ ] [ __ ] that it barely matters. Gemini being fourth place here is a very\nimportant detail that fundamentally changes how Google and OpenAI respectively think about things. If\nhypothetically speaking, the Gemini app went from utter [ __ ] to roughly as good as Grock, suddenly OpenAI has to be a\nlot more scared. But right now, the reality is that the average subscriber to something like Google Workspaces is\nmore likely to use chat GPT than they are to use Gemini. Like think about it. How many people do you know that are\npaying for Google Workspaces that have a domain that isn't atgoogle.com or atgmail.com? How many of them use Gemini\nevery day on gemini.google.com? And how many of them are using OpenAI models through ChatGpt, through their editor of\nchoice, through the phone app, through T3 chat or whatever else. I know pretty\nmuch everybody I talk to on a day-to-day basis other than like my family has a Google Workspaces plan that they're\npaying increasingly large amounts of money for every month. And none of them use Gemini. The moment that changes,\nGoogle's strategy here has to shift significantly. But at this point in time, they're the default. The same\nreason Google doesn't really give a [ __ ] about Gmail and doesn't bother updating it meaningfully is the same reason\nGoogle's okay with being second place. It doesn't matter how much better or more private Proton Mail is because\nevery company's using [ __ ] Google or using Outlook still. I guess it doesn't\nmatter that Nano Banana Pro is slightly better than GPT image 1.5 because people\nare already using chat GPT. They already won. And that's the interesting thing. OpenAI does not give a [ __ ] about being\nsecond place as long as the gap isn't that big and they stay first in chat. The reason that OpenAI is okay with\nbeing in second place is because they already won first in the only category that matters, the app you open when you\nwant to talk to an AI. And as long as they maintain that lead, they could fall the third, fourth, or even fifth place,\nbecause they're not in this for the small wins on benchmarks. They're in this to make the most money. And I it\nwould be very very surprised if they stop making the absurd amounts of money they are. OpenAI is going to win even if\nthey stay second place. And that's a really weird thing to say, especially when I'm sitting here wearing my Google\nDeepMind shirt. But yeah, hopefully you get the idea. Hopefully this is a useful video. Thought it was an interesting\nrant. I'm curious how y'all feel. Let me know in the comments. And until next time, peace nerds.\n"

ERROR: type should be string, got "https://youtu.be/-5LfRL82Jck\n\nI’m addicted to Claude Code (i get it now)\n\nOver the holiday break, something interesting happened. Anthropic gave you a 2x increase in your rate limits on\ncloud code. And I decided that I wanted to try my best to hit it. And man, have\nI had a wild ride. I never thought I'd see the day where I'm running six cloud code instances in parallel. But this is\nbasically my life now. I haven't opened an IDE in days. I have been building\nmore than I've ever built. And I'm questioning what the future looks like for us as an industry. I knew Claude\nCode had improved since I last used it, and I knew Opus 4.5 was capable of things that I didn't think were possible\nusing LLMs previously. I did not think this would fundamentally change the way that I write code. Over the holiday\nbreak, I built two full projects from scratch, a web and mobile app for one of them, did some crazy overhauls to those\nprojects throughout, published a bunch of work for new features in T3 Chat, and just configured my entire operating\nsystem using Cloud Code. I think it's safe to say at this point in time I am uh I'm a little Claude codepilled and I\nwant to share a bit about why. That said, Anthropic is certainly not paying me any money. In fact, I had to bump the\n$200 a month subscription to do this. So, we're going to do a quick break for somebody who is paying me, today's sponsor. A lot of people say my products\nare only successful because I'm popular. And they're kind of right, but not for the reason you think. Turns out if you\nhave a bad product and you put it in front of a bunch of people, they don't care. But if you have a good product and put it in front of people, they do. But\nthat's the key. You have to have a good product. In order to do that, you need a good team. This is the benefit I have\nfrom being popular. I have access to millions of the best engineers in the world. And I don't know how I would\nbuild without that because I remember the days where I was recruiting the old school way. And it was hellish. If I\ndidn't have access to the engineers I do, I would certainly be using today's sponsor, G2I. These guys know how to\nhire. In fact, they have placed some of the best people in my community at really cool companies. I've been blown\naway at the projects that G2I is involving themselves in. Everything from Web Flow to One Password to Facebook\nthemselves. Yes, Meta hires through G2I now. And it makes sense why. They really understand what developers need to be\nsuccessful. They run things like React Miami. The people who work there watch every video that we post. They know how to integrate with your team in a way\nthat doesn't feel like some weird third party arm doing recruiting for you. They're building a platform to help you\nfind the best people for the gig when you need them. It doesn't matter what you're trying to fill. Junior, senior,\nfront-end, backend, full stack, mobile, data scientist, all the above for all sizes of companies. By the way, if you\njust got out of your YC batch and you're trying to get a good engineer, don't waste all of your time going through recruitment hell when you can work with\na partner like G2I and have a much much better time. Their goal is to get the first PR in within 7 days of joining.\nAnd I've seen this happen enough times to believe it. Stop wasting your time hiring and get back to building at soyb.ink/gti.\nI'll admit this was a demo I just rigged up for the sake of this video. I just had a bunch of Cloud Code instances\nediting pages on my personal site, but it's not too too far from how I've been\nusing things in reality. I've been making a lot of real work and doing a\nlot of surprisingly difficult tasks with Claude code. I've gone so deep that even Ben is surprised. I'm doing things he\ndidn't think the models were capable of. Just to get us started, I want to show you guys the main project I cooked throughout break. This is my new image\nstudio app. This is meant to be a prototyping playground to figure out what we want the image generation\nexperience to be like in T3 chat. And yes, it's that image gen mock demo I've been doing all the time. The difference\nis it's no longer a mock. So apparently I never set up scrolling for the\nsidebar, which is kind of hilarious. Normally my screen's a little higher resolution than it is when I'm filming.\nSo, I can't actually hit the generation button because I never set up a scroll view here. So, I'm just going to show\nyou how I've been using things. The sidebar in the web app does not allow me to scroll. This means that on small\nscreens, I'm unable to hit the generate button. And now, we wait not even that long, and this issue will be fixed. You\nmight have noticed I specified web app. There's a reason for that. Not only did\nI build this app as a fully functioning web interface, I actually went as far as to make it a monor repo with a mobile\napp, too. And how I did that is possibly even crazier than the fact that I did\nboth. We'll get there, don't worry. One of the things of note is that when you start a new session, it doesn't really\nkeep track of any context. You can slash resume to go to an old session ID, but\nif you don't, it has to recontextualize its knowledge of the codebase. It's actually been really nice to be in a\nlongunning thread where I just tell it to do something else and it knows enough\nfrom previously to know where to start. I no longer find myself clearing the history and starting a new chat thread\nfor every single thing I do when I'm doing agent of coding with cloud code, especially if it has a pretty good idea\nof where things are in a codebase for a medium-siz project like this. To be fair, the implementations on any of the\npieces of this project are still relatively small. But the amount of service area there is to cover now across the multiple packages is actually\nquite reasonable, especially considering the fact that I have yet to open an IDE for this project. Forgot to turn on\nbypass permissions because I'm running it allow dangerously now. That's the point I'm at in life. Unrundev.\nOh, it does not appear to have worked. When I try and scroll on the sidebar, it\nattempts to do a whole page scroll, which doesn't work because the only scroll container currently functioning\nis the scroll container for the grid. I think we need a different approach to make scrolling work properly on the\nsidebar. Of course, the thing isn't working the way it does almost always when I am using it when I am now here\ntrying to film. Chat making a very good point that this alone just disqualified new vibe coders. Yeah, this is still the\nproblem. And I want to make sure that when I do content about this stuff that I emphasize this point. These tools are significantly better if you already know\nhow to code. Their strengths are much weaker if you don't yet. But now that I've made this one specification, it was\nable to make the right changes and get this all working as expected. Awesome. Cyborg form fields now scroll\nindependently when the generate button stays fixed at the bottom. And I will command minus a bunch to make sure it goes away when it's zoomed out enough.\nIt does. Awesome. So, now that we know this all is working as expected, I sure, we'll do 16 by9.\nWhy not? Standard resolution because some of these models don't support higher res. Actually, if I turn off nano banana, we can do high-res. Generate.\nAnd now we have three models separately generating images for this random test I\nwanted to do. Of course, the back end is all through convex. You guys know I've been loving Convex. Front end is a\nstandard V Tailwind app. All of the UI was done by Claude. We had a lot of back and forth. This is rendition like five\nor six like full UI overhauls because those types of things are surprisingly easy to do with a tool like claude code.\nIf you just spin up a few work trees, throw cloud in them and tell it you can even copy paste the exact same prompt\nlike I want to redesign this homepage. Something stupid I have been doing recently is having it spin up three\ndifferent routes with the exact same feature like slash one slash two slash3 and then I go and compare them and tell\nthem I like this and route one but there are some things I like about route two and three. Can we bring these ideas over\nbut focus on the design from route one? And it handles this stuff incredibly well. And good point, chat. I did also\nuse Shad Cien. Shad CN has made my life much easier. By the way, if you like these types of exploratory videos about\nhow the dev world is changing and how my own development work is changing, I love doing this type of content. And one of\nthe best signals for the team that you like this is to hit that subscribe button. Less than half of y'all are subscribed and we're so close to hitting\n500K. So, if you don't mind, hit that button. Helps us out a lot. But let's pick one of these that we like. This\nlooks good enough. Obviously, arrow navigation works plenty, but I didn't even need to tell it to do that. It just did. Seedream did not honor the aspect\nratio I gave it. That's a good thing to know. Banana Banana Pro did. Let's do one of the features I was most excited\nabout here. You can't see this button, but there's a little chat icon right there because any picture you generate\nin this app can become a conversation. Let's say we specifically want\nNanobanana Pro and Seeddream to do this. Replace the Sea Lion with a corgi. And\nnow we have a chat interface for follow-ups. Of course, image generation models still aren't the fastest thing,\nbut they're more than fast enough. And here we are. We made a new generation with a corgi. Still one click to\ndownload if you want to do that. We can zoom in and see him if we want to get better luck. Not bad. How's this one? I\nlike this one more. So, we'll pick the nano banana one and we will do one more follow-up. In this follow-up, I will\ntell it to I don't know, change the text to say corgo. I'll be more specific on\nthe sign. This has been a super fun side project both for playing around with cloud code and trying to push its limits, but also for playing around with\nnew UX flows. I wanted to see how it would feel to have a two style view where we hop between a gallery view and\na chat view. I was just really curious how it would feel to use a UI like this. And I'm finding that for the most part\nit's really good. There are catches. There are always catches, but I am actually overall quite enjoying this\nexperience. Especially because thanks to Convex, not only does it sync fully, it\nsyncs fully with the mobile app. I want to show you guys some of the most absurd prompts I've run in my life because I\ndid not think this would work. Here is a real prompt that I sent to Claude Code.\nI want to turn this project into a monor repo with the current web app and a react native mobile application for the\nmobile app. Use the following stack expo react native expo router TypeScript bun\nuniind for tailwind bindings and convex using existing convex bindings. We only care about iOS for now. So don't put\nmeaningful effort into multiplatform support. Ideally the convex bindings and definitions will be shared between\nprojects. Use turbo repo for managing the monor repo and sub packages. Write a thorough plan. This was meant to be a\ntask it could not complete. I got here because I am now convinced there are\nvery few tasks opus in the right harness cannot complete. I mostly did this as a\njoke. I was planning on taking a screenshot of this prompt and posting it on Twitter saying lol nothing can\npossibly do this. And then it did. Yeah.\nThis run took over an hour if I recall correctly. The plan took like 20 minutes\nby itself. Can I go to this plan? Look at that. The plan is still saved.\nThis is just the mobile app feature implementation. Is there a uh or maybe\nit like overrode this at some point? Yeah, I think it overrode this at some point. So, the history here is the right\none. Convert the current web into a turbo repo monor repo with a new expo react native mobile app sharing the\nconvex back end. We have apps, web, mobile, packages, backend shared, TypeScript config for sharing TypeScript\nconfig stuff between things, turbo JSON, package JSON, npmrc. It had names for all the new packages.\nDesigned the foundation with a root workspace config and npmrc in order to handle the weirdness that is how\npackages are linked between things in node and monor repos. Created the different packages it needed\nto move things over to app/web. even calls out critical files is going to have to modify and change. Phase two,\nit extracted the shared packages. You get the idea. It wrote a very thorough plan. I gave it a really quick once\nover. I don't even think I told it to change anything else if I recall. I think I just said do it. Looks from here\nlike that's exactly what happened. And then it just ran and it ran\nand it ran. It even remembered to use its front-end design skill, which by the way is just a markdown file saying,\n\"Don't make it look like AR slop.\" It tried running bun install. It got errors. It went and fixed a bunch of the\nversions. It updated the package JSON accordingly. It still got errors. It\nwent and fixed more things. Then it stopped getting errors when it tried to build.\nIt moved from Uniwind to Native Wind at some point because it was having problems with Uniwind, but it just kept going. Noticed more mistakes with how it\nwas calling the APIs. I need to fix the mobile app to use the correct API functions. Let me update the home screen. And then suddenly it was done.\nThe monor repo conversion is complete. Here's a summary of what was done. Yeah.\nYeah. To be fair, the mobile repo had a good number of problems. Specifically,\nthe convex URL was not loading into the mobile app properly. It also did a bad job of copying over my environment\nvariables both because I had changed to a mono repo and to a workspace which is\nthe JJ equivalent of a work tree at the same time as I did this. So it just kind of lost track of environment variables.\nDid a tiny bit of moving files around in Finder to fix that. Then I kept hitting weird serverside\nerrors that apparently were from Nativewind. Apparently my messages are missing from\nthe history. They're not really great about history with this stuff. So, I'm just going to show you guys the PR. Here\nis the mobile monor repo PR. The thing that was craziest to me was this. Grab\ndid a review of this PR and gave it a five out of five confidence score the first try, which I just didn't think\nmade any sense. Called out some places where there were duplicate imports, but didn't really matter. And this is a huge\nPR. That's 2,300 lines added and 400 removed. It was able to break up one\nproject into five packages with a somewhat complex turbo repo setup with no issues and add a mobile app. There\nwas a slight problem with it adding the mobile app though, which is that it ran the exponent command which by default includes git which caused a subggetit\ntree to exist and JJ just ignored it. So I actually forgot to include the mobile\napp in this when I merged it which shows you how thoroughly I was reviewing the code. I didn't even notice that the app\nwas missing. This ended up bothering me much later as I was trying to do other things with work trees and the mobile\napp was missing when I did it. Again, turns out I had forgotten that. Ended up asking Claude for help with JJ. Fixed my\nhistory cuz again JJ just doesn't give a [ __ ] You can go edit your history. It's so nice. And then readded the mobile app\non a separate branch. Ask reptile to review again. And again got a perfect [ __ ] score. My hand rolled changes\nthat are like 200 lines of code usually get like a three out of five and I was hitting fives out of fives constantly\nwith the changes that Opus was making. Now is gravile going to be the objective perfect way to review code. It saying\nfive here and three there means that this is better than that. No. But it is [ __ ] fascinating to see. And if you\nlook through the code, it all makes sense. Maybe a thousand line of code tsx file isn't the greatest\nthing in the world, but welcome to ReactDev. We have all the mutations for all the\nthings that we are doing in a thread. A timeout that will help with the scroll\nto end. That all makes sense. Ah god, the GitHub UI is obnoxious.\nI didn't quite get the keyboard avoiding view properly, but it wasn't too far off either. But honestly, I can just show\nyou guys. I should use another mobile phone for this, but I'm dumb. So, let's\njust do it. And here we are, the mobile version. Mac OS's mobile share thing is\nreally annoying with resolutions, but we have the mobile app and it works. And since again, we were using Convex,\neverything stays in sync. So, if I go and create a new gen here, uh, two\ncorgis sharing an ice cream cone, and I hit gen here,\nit syncs with the web UI as well with no special anything. I didn't tell the\nmodel to do this. This is just convex, which is another one of those things that has fundamentally changed how I build. But one of the things that makes\nall of this so powerful is that Convex is just a folder in the codebase that you configure, which means the model\ndoesn't need some crazy MCP [ __ ] to go change settings in a dashboard. It just edits the folder. You get the idea,\nthough. This is a full functioning mobile app that's a lot less laggy on my phone. It's actually been pleasant.\nLike, I've been using this to do random image genen for some work stuff. The hardest parts of this project by far\nwere dealing with GCP and the Google Cloud dashboard to get the tokens I needed to do off and then configuring\nClerk Convex and all of the Verscell stuff to handle build and deployment\nproperly for a productionized version. The thing that was hard for this was getting it into prod because I had to\nfight all of these legacy dashboards and tools and things. Everything else was\nsuper smooth. Even other annoying [ __ ] like implementing off, which even with the best tools, even with\nthings like Clark and Work OS is not the easiest thing, especially when you have multiple platforms and three different\npackages in your codebase that need to have off added. I need it in the mobile app. I need it in the web app. And I\nneed it in convex in a way that the convex functions know the user is off and they know who the user is. All of\nthese layers are annoying to get right, even on a good day. And it just did it. It's another 1,800 lines of code added.\nBut it did it and got another five out of five confidence score. So, I merged it. Would I put this code in front of\nall of our users for T3 chat without a more thorough audit? Absolutely not. But\nis it worth talking about the fact that I was able to from scratch build all of\nthis without opening an IDE a single [ __ ] time? I think that's worth talking about. That's 11,900\nlines of code. That's a real code. That's not a massive codebase, but that's that's a real code. And this\nentire thing was generated on the $200 tier of cloud code. That's just absurd\nif you think about it. This isn't just tab complete in our editors. This isn't just like go change these three files\nfor me or deprecate this thing. This is a fundamental change in how I think about writing code and also a bit of a\nchange on how I think about using my computer. We'll get to that in a sec, but I want to show one other project I built using all of this. One of the\nproblems I have with cloud code and tools like it is now that I'm running these jobs that can take 45 minutes to\nmultiple hours, I find myself getting distracted and then I don't go back when the work is done and I don't want to set\nup some audio queue or some [ __ ] Even if I did, I don't have the discipline. I'm just going to browse Twitter, which is why I made an extension that locks me\nout of Twitter unless I have work going in one of my cloud code terminals. I'll ask it to make the UI look better. Just\nsome [ __ ] busy work. And now my Twitter's unlocked and I can once again see my tweet where I complain to the\ncloud code team that I can't hide my email address. But as soon as the work is done, I will\nbe locked out of Twitter again. This has been a weirdly powerful boost to my\nproductivity. And I have also not opened up an IDE to look at the code for this project at any point. Do I look at it on\nGitHub sometimes? A little. Not much. Would I trust this code with my life? No, not at all. But is it a fun side\nproject that I had an idea for that I was able to build in half an hour instead of a day? Yeah. That's what's\ncool here. It's changing whether or not I'm willing to make a project, not whether or not I'm capable. I was always\ncapable of building things like this. I'm not doing things I couldn't do before. I'm doing things that I didn't\nbother doing before because suddenly they are so much easier to do. Which UI\nservice do I want to improve? Um, never mind. I like it as is. Nah.\nAnd now it's done and I'm back to locked out of Twitter. That's cool. The fact that I hacked this together in like 30\nminutes of work where I was doing other things and working on the mobile app at the same time. That's really cool. I had\ntwo of these cloud code instances running for the mobile app and then one running for this project all at the same\ntime and would constantly get kicked out of Twitter when I went to talk about and ask people questions because when all of\nthe jobs ended I would get booted. This was such a fun little thing to build. This one's open source by the way. Cloud\nBlocker is live on GitHub if you want to play with it. I've already submitted it to the Chrome web store so we can get the extension out ASAP. Those take a\nwhile to get reviewed, especially when you have more grandiose permissions because you need access to tabs to do\nthings to pages. But I could do it. I could do it without much effort. I put this whole thing together without much\ntime spent. People are already making changes to it, which is really cool. But man, like this type of thing was stuff\nthat I would maybe do if I had a lot of free time sometimes. And I really want to emphasize this point. Like these\ntools aren't just good for making your job simpler or trying to like do work that is tickets from Jira or whatever.\nThe thing that's really fun for me with AI code is that there are all of these ideas and vague things that I like could\nhave done. Like I could have built an app that was just to parse my notion history, which I did today. I have a\ndifferent file in notion for every month where I log every day what I did that day. And I wanted it all to be in one\nfile so I could hand it to LMS to make them judge me and roast me. Just getting all of the content out of those in order\nwas annoying. So I put them in a directory. I opened cloud code. I said make these one file. And it wrote\nscripts to run and make the changes to make it one file. These are all things I could do before. I knew how to write all\nof this code. But if I was going to spend a certain amount of time in a day writing code, it was more important that\nthat time went to projects that mattered and I didn't have as much left over to do fun side projects. And if I did, I\nwould pick the single biggest, most fun side project, not dozens of them. Now I find myself spinning up random projects\nall of the time. And sometimes I find myself using cloud code for things that aren't even projects. Remember that\nthing I said earlier about the way I use my computer changed? Let's just look at some examples.\nI need you to update my JJ config to sign commits the same way that my git config does. This one I think I had\nmultiple messages for. I guess it's just not restoring properly. That is really annoying. It\nseems like resume is just kind of broken. Whatever. The thing I wanted to show is that I was able to update my JJ\nconfig to fix my commit signing by asking it to and then it did. I was able\nto add a new script to my Zshell that was a script that will automatically break something out into a work tree,\nchange that directory, and copy over all of my ENV files from that project. By\njust asking it to do that, I'm not just asking Claude Code to edit files in a codebase and occasionally search the web\nfor more context. I'm asking it to use my computer and make changes to my setup in my environment the way I normally\nwould between like five different tabs, a bunch of searching, a bunch of trial and error. Or I can just tell Cloud Code\nto do it and go grab a tea. It's unbelievable. There are so many things that just didn't make sense for me\nbefore. Be it a new side project. Be it a weird extension I want to build. Be it configuration changes I want to make to\nmy system. Be it the move to JJ instead of Git where I'm starting to move off of Git as my default like CLI and way of\nmanaging repositories. These types of things weren't as viable before and now\nwith things like Claude Code, they are suddenly much more viable. If you're still new to tools like this, I won't\nnecessarily recommend that you do what I'm about to show you immediately. definitely keep it in the standard prompt you for edits mode. Then when you\nhave the confidence, switch over to autoac accept edits. And then after you've done this for a while and you've\nlet it run commands that edit your system, change things and take risks maybe a little more than you expected\nand were surprised with the results like I and many others have been, you can switch over to allow dangerously. I\nunderstand why they hide it. I understand why they are so strict about this, but allow dangerously skip\npermissions. It's so fun. It's genuinely so fun. I admittedly have a bias here,\nwhich is that if I end up being one of those people where Claude randomly nukes my home folder, I get really good\ncontent out of that. If that happens to you, you probably don't have benefits. It probably just sucks. That happens to\nme, it would be one of the greatest days of my life. I could make so much good content about that. But it probably won't. But it could. But I'm also doing\na few things to prevent it cuz I want to do my best. Due to a series of unfortunate events on Twitter I do not\nfeel like going into, I was made aware of this new project, the Claude Code Safety net. This is a plugin that you\ncan add to Cloud Code. Shout out Daisy for the plug-in system. It's awesome. One of the coolest things about it. It's a cloud code plugin that acts as a\nsafety net catching destructive git and file system commands before they execute. So if the model decides to do\nsome crazy [ __ ] with git or some crazy [ __ ] with RM, it will prevent it from doing that and give you one last check\nto the user even when it's in the dangerously modes. Will this catch everything? Absolutely not. You could\nstill write a bash file and then put it in your system and then execute it that does those things and it can and it\nwill. The models love to try and work around the restrictions they have. If they can't edit a file by using the edit\ntool, they will run a bash script to edit it instead. If they can't do that, they'll run a Pearl script to do it\ninstead. The models are very, very weirdly willing to work around the restrictions you give them. So, know\nthat this will never be perfect, but again, for my risk profile and the things that I'm concerned about, the\ncombination of the Cloud Code Safetyet plugin and my own traditional usage of Cloud Code in allow dangerously has not\ncome close to causing me any problems. I've been impressed. This has been a wild deep dive over the last two weeks\nand I don't know how it's going to affect the way I code long term. Obviously, I still love cursor. I still\nhave my investments there. So, account for the biases accordingly. When I actually want to see the code and do my\njob as an engineer, not just dick around with a lot of different things and ideas, I still very much like using a\ntool such as cursor. And when the codebase already exists and has things going on in it and I'm working in a\ncodebase as my job, cursor's whole agentic environment is still something I really enjoy. But when I'm\nexperimenting, when I'm screwing with my system, when I am in some crazy green field thing or I just want to let\nsomething run in the background for an hour on some crazy task, I've been impressed with cloud code. Now, there's\ntwo questions I know are already swarming the comment section because the people who are asking these probably didn't watch to the end. Sorry to them.\nQuestion one, what about open code? Open code's really good. Not going to pretend otherwise. I wanted to try out the new\ncloud code updates in the new cloud code subscription tier in the way that a lot of people have been using it cuz I haven't really had the cloud code\nclicked for me moment yet. And this is what pushed me over the edge. It's combination of Opus 4.5 being an unbelievably smart model and the harness\ngetting mature enough that it's in a really good spot. Open code is still the best option for almost every model and\nit probably is almost as good when using Opus. And they also have a way that you can use your cloud code subscription in\nopen code. It's a bit hacky, but it seems to work fine. I just didn't feel like dealing with any of that. I'll evaluate it more in the near future. And\nnow the next question, which is the Ralph loop? This website is questionable. I'm assuming it was made by a model that was not given the front\nend design skill. That said, you're not familiar with Ralph Wigum. He's a character from Simpsons. Also, hype\nJenzie. Good to see you guys. I know everybody else is familiar with the Simpsons. The point of the Ralph Wigum loop is to let Claude code continuously\nkeep working even when it thinks it's done until some higher order this is\nprobably done gets hit. Also, Miles decided to come say hi. Now that I am doing more vibe coding, I spend a lot\nmore time petting my cat when I'm coding. So, he now assumes when I'm working on code\nthat it's cat petting time. You let Claude code a lot, don't you, buddy? He\nprefers Codeex because it's so much slower. So I get more time to pet him. Oh, he's being so cute. Anyways, as I\nwas saying, Ralph is a bash loop. This is a way to run cloud code and keep running quad code even when it stops and\nasks, are you sure? Or do you want me to reconsider this? Or I finished phase\none, should I start phase two? Just run a bash loop and keep telling it continue, continue, continue until\neventually it actually finishes. It's a fun strategy. Personally, I found that\nClaude Code is pretty willing to just go for an hour or two without being told, \"It's okay. You can continue.\" I've had\none time total thus far where it would stop after each phase and I'd be like, \"No, keep working, please. No, keep\nworking, please.\" Over and over. But only once ever did I hit that and it was actually a pretty short run. Like that whole run only took 5 to 10 minutes. But\nI've had other ones that took an hour and didn't stop at any point. So, I still personally don't know how\nnecessary this is. But believe me, the Ralph loop is something I very much have\nplanned. So, uh, if you don't see a video about the Ralph loop from me soon, know that I wasn't impressed enough or\ndidn't find it interesting enough to do one. But if you are interested in this, keep an eye out. I will almost certainly\nbe talking about it. I think that's everything I had to say here. I'm impressed. Cloud Code's gotten a lot\nbetter. I have my complaints. Like, the hooks still suck and feel half finished. Plugins are awesome, but don't have all\nof the functionality I need to be able to rely on them directly. Skills are kind of a joke. They're literally just markdown files. There's a lot of little\nUX things like the way stashing works sucks. If you want to change things after you wrote your prompt, you're\nscrewed. Context compaction feels really strange still. The way history management works and the way picture\nuploads work is still jank as hell. There's lots of little places it can improve. I don't want this to seem like it's just this perfect magical solution,\nbut god damn, I have been able to do a lot with this and I've been really impressed. And one actual last thing, my\nusage. Remember, I was on the $200 a month tier with the 2x limits. I did\nmanage to hit the limit on the $20 tier. I considered going to the 5x, which is the $50 one, and then up to the $200 20x\none, but I decided to just go all the way and see how much I could use. Like, I was going out of my way. I was running\ntwo to three of these in parallel during most of my waking hours. I want you all to guess chat mostly between like 13%\nand 90%, some people with some jokes. You can't really know your usage and they work really hard to hide it. But if\nwe look at the dashboard, you'll see that my weekly limit, I am at about 2%\nof my usage. So this just reset. So that one being 2% makes sense. But this top level one, my plan usage, which resets\n34 minutes for this session, 12%. And the highest I managed to get in my\nweekly limit last week was 7%. 7% going as hard as I could. Take off\nthat 2x and I was 14%. For the week from my napkin math from the tokens I was\nable to log, I did about $1,500 of inference in that time. I paid 200 bucks\nand I got 1,500 out. As a user, I love this. That is a really good deal. As a\ncompany also selling inference that pays API pricing for a lot of anthropics\nAPIs, I am pissed that I'm subsidizing this. The reason these subscriptions can\ngive you 10x the inference you pay for is because they are subsidized by people\nlike me paying full price over API. So on one hand, cool, I love using this. On\nthe other, you're welcome, you ungrateful bastards. God, so much clicked over the last few weeks for me.\nI I have a whole different worldview, a whole different view on how code works\nand where it's all going and why people like Cloud Code so much.\nI got nothing else. I hope you enjoyed this this descent into chaos.\nLet me know what you guys think and if cloud code is driving you mad as well. Hope this was useful, but until next time, peace nerds.\n"

ERROR: type should be string, got "https://youtu.be/iG4AGQqbA6I\n\nIt was a wild year for CSS\n\nThis is how it feels to be a web developer in 2025. Actually though, this is the CSS wrapped 2025 that was just\npublished by Google in the Chrome team. I hope this image isn't AI generated. It's interesting either way. That said,\n2025 has been a fascinating year for CSS. Not just because Firefox is finally catching up to standards from like early\n2020, but more so because lots of other things got shipped and it really feels like the Chrome team is embracing the\nfact that we've had to use a ton of weird libraries to handle things like drop downs and finally these things are\nbeing added to the browser. I've always been concerned that AI stuff would prevent us from iterating on standards because all of the agents are only going\nto know the old things. But that has not slowed down the Chrome team at all, even if Adi Osmani has left to do other things at Google. Regardless, this is\ngoing to be a really fun breakdown. I can't wait to go over all of the cool things that changed over the last year in the world of CSS and web design. But\nfirst, a quick word from today's sponsor. Platforms like Forcell are awesome until you realize you don't want to use JavaScript or you have to attach\na database and then you have to go through a different platform for that. Then you put a firewall in front and you start doing all of these things and attaching all of this glue and you're no\nlonger building, you're maintaining. That is unless you're using today's sponsor, Savala. Zavala is the\nall-in-one cloud for whatever you're building. They include everything from the databases to the CDNs to the servers\nthat you're actually running things on and integrations with every major framework and platform you'd want to build on. Whether you're trying to\ndeploy NodeJS or Java or Elixir or even Rails, Savala has you covered. Their secret is that they're not just one\ncloud. They're not an alternative to things like AWS, GCP, and Cloudflare. They're built on top of them. You don't\nturn on Cloudflare by going to the Cloudflare dashboard, setting up an account, and doing a bunch of weird DNS stuff. You do it by going to your app,\nhitting settings, and checking a checkbox. Turning on the CDN is a checkbox I can click on. Do you know how\nmuch more annoying this is basically anywhere else on any other platform? It's so nice. If you're worried about\ncosts, don't be. They are comically cheaper. Companies like Steppper saw an almost 80% cut in their bill by making\nthe move over to Savala. Stop wasting time with glue and start shipping at soy. Savala. This year, we're giving you\nnew tools to sculpt a dynamic web. Sculpt dynamic interfaces stretch your\nimagination and play with these powerful new CSS features. Also of note is I'm using a Chromium based browser. I didn't\ntry this too hard in Zen and I still dearly love Zen, but personally I wanted\nto see how this would look in Chrome because it's by the Chrome team. Firefox is catching up, but Chrome's still ahead. We've been crafting new features\nwith you in mind. I know this is like typical marketing mumbo jumbo, but having talked to a lot of people on the\nChrome team, it does really feel like they want to hear this feedback and make changes accordingly. I've been really\nimpressed with how transparent and open the Chrome team has felt over the last\nyear, even if there have been some concerning departures. The laying off of\nAdam Argyle was terrifying. Nicole Sullivan leaving for Safari and Apple\nwas unexpected and broke my heart. and now Addios Osmani leaving as well. There are still awesome people on Chrome, but\nthose are three of my personal favorites and seeing all of them disappear over the last two years has been rough. So fingers crossed that this trend can\ncontinue, but it really depends on how much positive support the Chrome team gets and how Google feels about this\ninvestment. So, uh, leave kind words to the Chrome team in the comment section if you don't mind. And on the way down\nthere, there's a little red button to subscribe for at least half of you guys. If you could hit that, we're trying to get to 500K subs before the end of the\nyear, so would help out. Customizable components, we give you the clay. Functional, stylable components,\nbuilding blocks, ready for you to mold their look and feel entirely. This is a\nterrible sentence. NextG interactions with new APIs for movement. That's all the fancy like view transition stuff and\nprobably a bit more. And optimized ergonomics. Stop fighting the code. These features help you work with CSS in its natural state, making your workflow\nintuitive. Okay, I need a reality check, chat. Have I been reading too much AI\nslop output that I'm starting to feel like all enterprisey [ __ ] speak is\nAI slop or is this AI slop? Am I thinking this is AI slop cuz I read too much AI slop or is this AI slop? Is this\nAI slop? Yes, it's slop. No brain rot.\nI just work here. Cool. Got to pull it. First one is fine. I'm just bad at reading.\nBring your design to life. Use new APIs to control all complex movements for a\nfluid and seamless user experience. Okay, people are agreeing with me. It's slop.\nHopefully the rest isn't ready to see what we molded in 2025. The Chrome Devril team will guide you through 22\nCSS and UI features that landed on the web platform fresh from the kiln.\nIs that a marquee? It is really funny that the first like super fancy CSS feature on the site is a CSS feature\nthat isn't CSS. It was in the DOM, but it got deprecated. Customizable components. The workshop was hot this\nyear. We took the decades old problem of styling dropowns and fired it to perfection. We also delivered new core\nblocks like native anchor positioning and carousel scroll APIs. These are really good things. The native dropown\nstyling finally existing as well as having better anchor point stuff built in is huge. Invoker commands show a\ndialogue modally and more without JavaScript. We can go try the demo. Cool. So now we have the show modal\ndialogue. That's the dialogue and you can toggle the pop over.\nYou know how annoying these things were to do without packages before? Now we barely need them.\nThey do have a polyfill for it here. Here's their minimal example they show in the blog post though. Unccclick\ndocument.query query selector my dialogue show modal. See that there's a\nbuilt-in show modal call that you can do now. Not really CSS, but it's invocable\nthrough JavaScript to trigger the style stuff. Oh, that's the old way actually. The new way is an invoker command. Yeah.\nButton command for my dialogue command is show modal. And now that you've defined what this button is commanding,\nit will trigger without any JavaScript being needed at all. Awesome. Cool. Currently, it's possible to send\ncommands to popovers and dialogue elements with more types of elements possibly coming in the future. They're\ninteresting. I do really like the idea of elements that are clearly tied together, not requiring another language\nfrom another file to control their relationship. Similar to why I like Tailwind, like these elements are\ninherently tied. They're in this component together. I shouldn't need to switch files to see how they look. Here\nis very similar. I shouldn't have to switch to a JS file to figure out how this button and how this dialogue\ninteract with each other or do some weird ref management in React. This syntax is a way to tell the browser how\nto make these things interact and not making it my problem in JavaScript land is a good thing. It's also possible to set up custom commands to send to\nelements. These custom commands are prefixed by two dashes and are handled by the toggle event. Interesting. This\nis actually really cool. Since you can send commands to an element with a button, you can then on that element\nwith a query selector added a custom event listener for this new custom event that you defined. That's really cool. I\ndid not know that was a thing. Dialogue light dismiss, bringing a nice popover API feature to dialogues. One of the\nnice features introduced by the popover API is the light dismiss behavior of popovers. This lets users close them by clicking outside on the backdrop or by\npressing the escape key. We were just dealing with this with a new feature about to ship in T3 chat. Again having this in the browser is great. Light\ndismissive behavior is also available on dialogue through the new closed by attribute which controls the behavior.\nYou have closed by none which means no user triggered closing of dialogues at all. This is the default. Again problem\nwith the browser is they implement something it works a certain way. They realize that way is wrong and then they\nmake it so you can do it right. But all of those additions to make it right are new things being added. The classic\nexample of this will always be triple equals in JavaScript because double equals sucks and by the time they fixed\nit, it was too late to change it without breaking old websites. So triple equals was created to give the better behavior.\nHere the close by none behavior is default because that's how it worked before because it was wrong. And now you\nhave to specify close by close request or close by any to give it what I would\nconsider the better default behavior in order to match the popover behavior. The\nfact that if you don't put a property here, the defaults differ so much is obnoxious. Welcome to the web. But now\nshow dialogue and I can escape key, but close by is none, so it won't let me.\nBut I changed it to close by any. Now I can good. Should have been the default.\nPopover equals hint. What's going on with the spacing there? Gh, the the\nkerning. Ephemeral popovers that don't close\nothers. Hint popovers with a popover equal hint flag are a new type of HTML\npopover designed for ephemeral layered UI patterns like tool tips or link previews. Opening a hint popover does\nnot close other open auto or manual popovers, allowing layered UI elements to coexist. This is a really bad problem\nwith the previous implementation of popover stuff. The age of how it worked meant that if you had a modal or a\npopover that had a sub hint in it, it could break everything. Hint popovers\ncan also exist on link tags, unlike auto and manual ones which require activation from button elements. Interesting. Hints\nare allowed to work on other types of things.\nOh, I see. I have that open and I can hover this at the same time. That's the difference. Previously, this would have\ncancelled this open menu because they were sharing the popover space, but hint types are now excluded from that\ndifference. Good stuff. Speaking of which, customizable select. You can\nfinally style HTML select elements with CSS. This is what, two decades of waiting.\nHow long has it been since we couldn't style the select element?\nIf you aren't familiar, the classic like HTML select dropdown, this view here,\nhas never been customizable, and you're just kind of stuck with what your browser wants it to be. And every\nbrowser handles this quite differently. No longer is such the case. You can now\nuse base select styles to unlock several powerful features, including complete CSS customization. It has all of the fun\npseudo classes that you have to use to target things, but you can control the content of this. Now, select at colon\npicker select appearance base. This is the new minimal solution that optimizes\nit for customization. So, if we go back here, I know it's uh here CSS. There we\ngo. Now, there's no more border on that part. It's not doing what I expect for this piece, but it's customizable now.\nAnd you can see here with the example that they have on the actual docs, this is a stock HTML select element. No\ncustom JavaScript, no custom implementation of the drop down. All the options, all the pieces just behave. And\nin CSS, we can now make it render differently. This is really nice. This is a very\noverdue change. Not having to write JavaScript in order to have a basic selection element is a nice thing that\nwe should encourage them for doing. The other cool thing about this is in certain cases where it doesn't make sense to use the customized picker, you\ncan still have content that's accessible for screen readers or usable for somebody on like a mobile phone or a\nlegacy device. You get the classic like built-in select behavior in the browser\nand the ability to customize it on the platforms that support it. So your Firefox users are screwed, but everybody else actually Let's make sure this is\nbroken in Firefox as I suspect it will be. Oh, I I just assumed I didn't even check\nit. shows here that it's not a thing here. Watch what happens when I click yes. See, it falls back on the\ntraditional list. Your browser does not yet support customizable select. Try again in Chrome 135 plus. Yeah, that I\npick Chrome by default. Scroll marker button. Carousel scroll\nfor instances with native CSS pseudo elements. Oh boy, apparently nothing supports this yet. Poor Ken Wheeler. The\ntime of his jQuery carousel is nearing an end. Just 20 years later, the browser is finally adding it. These are two new\npseudo elements, the scroll button and scroll marker pseudo elements. These features let you create native, accessible, and performant carousels\nwith just a few lines of CSS and no JavaScript. So theoretically, once this is implemented, I can click this button\nor this button, which will trigger a scrolling pseudo behavior, allowing me\nto scroll a certain amount. I'm assuming they let you specify how much. Oh, they even have scroll markers. That's\nactually really, really cool. These markers are grouped in a scroll marker group and behave like anchor links,\nletting users jump directly to a specific item in the scroller. This is useful for creating dot navigation for a\ncarousel or a table of contents for a long document.\nThat is really cool. Here's one that also takes advantage of anchor positioning to make sure things stay in the middle.\nOh, the little arrow. This being almost entirely CSS is so cool. Is the JS just a polyfill? Oh, is this a comment photos\nfrom Unsplash? Yeah, that's so cool that you can do this type of thing without having to write any dynamic JavaScript\ncode. That is good. This is good for the web. The antiJS people better be really hyped about all of these things. I know\nthey're not cuz they don't actually care about the web, but still super cool. Scroll target group. Turn a list of\nanchored links into connected scroll markers. Uh, this is for custom scroll spy so that you can have the markers\nlike we saw but with full customization. That's actually really cool. The links now we we've all done the thing where a\nlink targets a specific title or a tag on a page. Now you can customize the styling depending on which section\nyou're in. So when I go here, watch. I scroll section two. Section two is lit up. I keep scrolling. Section one lights\nup. It's super simple to do. You just have your links here 1 2 and three. But\nin the CSS code here, scroll target group auto. And if it's the current target color, light, dark, red, pink. So\nit changes the color when it is the current target. Yeah, it's this scroll\ntarget group combined with target current pseudo class lets you do different styles when the thing that\nthis link targets is what's currently in view. That's super simple and actually really really nice. I already have uses\nfor this. This is super cool. This might be my favorite one so far and I really like having dialogue stuff. Anchored\ncontainer queries style elements based on their anchor position. Oh, this is really useful. So, if you're\nanchoring like a tool tip and you want to have the little like triangle like the carrot for it in a different\ndirection depending on where it's positioned, you need to know how it's anchored. And of course, the example's a\ntool tip. It is kind of crazy that something as basic as a tool tip has required a decade or so of engineering\nand back and forth to try and figure out how to standardize it kind of in the browser. So here\nwhen I scroll around the positioned element moves the tool tip based on how\nmuch space is available. When it does that it has to change the direction the carrot goes because if that's still pointing down when it's below it's going\nto look awful. So here we have the anchor defined by using position anchor\nand anchor name with these two fields. So that's how the CSS relates these elements. And you tell it to try\nfallback with flip block. So it will flip to the opposite side of the direction depending on where it's being\nblocked. You go to default area. In this case it's bottom. If I change this to top, we're going to have everything be\nbackwards where now when it's on top the arrow is facing the wrong way because\nyou have to build that relationship yourself so to speak because it doesn't know in this tag if it's above or below.\nYou have to determine that based on other things. the tag for this the uh\nanchored fallback flip lock this is when it is flipped when it is opposite the default apply these styles instead it's\nnot when it's on the bottom apply these styles which would be really nice if I could say on bottom apply this on top apply this instead you give it a default\nand then say when it is not that when it is having the flip behavior style differently I was hoping that would be\ncleaner this is a huge win for anchor positioning and component libraries enabling more robust and self-contained\nUI elements with less code. Yeah, this is for component library authors, not for us as day-to-day devs. Like the\naverage web dev should not have to set this up. It's cool that it exists and it is as flexible as it is. I just wish it\nwas more declarative. Speaking of which, declarative interest triggered UI with interest for hover and focus triggered\nUI everywhere on the web from tool tips to rich hover cards and page previews. While this pattern often works well for\nmouse users, it can be inaccessible to other modalities like touchcreens. Additionally, developers have to\nmanually input the logic for each input type, leading to inconsistent experiences. The new interest for\nattribute solves this by providing a native declarative way to style an element when users show interest in it\nwithout fully activating it. It's invoked similarly to the command for attribute, but instead of a click,\ninterest 4 is activated when a user shows interest in an element, like hovering over it with a mouse or\nfocusing it with a keyboard. When paired with popover hint, it becomes incredibly easy to create layered UI elements like\ntool tips and hover cards without any custom JS. Didn't they just say it's inaccessible to other modalities like touchscreen and then said you need a\nkeyboard or a mouse? I don't know how they're handling the touchcreen part here. It's a demo that uses interest for\nto create product calls on an image. Hovering over the buttons on the image will reveal more information about each product.\nLook at the HTML quick. Yeah, button class info point interest for call out one point interest for callout two. You\ndefine these two new divs both with the popover equals hint. And now in the CSS\nyou can specify how those things relate. Here you say interest for callout one top 25 right 30. That's how you want to\nposition it relative to other things. And this one is bottom 10 left 22.\nInteresting. Good stuff. That's a really useful one. Oh god. NextG interactions.\nWith this new interaction toolkit, you can now animate between pages of view transitions and sculpt gorgeous scroll-based experiences. Scroll state\nqueries style descendants based on whether something is scrollable, stuck, or snapped. I wonder if this title here,\nscroll state queries, is using this behavior. See that when I scroll up, it gets stuck in that top bar and then\npasses. I bet this might be how it does it. Everything but Firefox and Safari. Well,\neverything Chromiumbased then. Thanks to Scrollskate queries. available from Chrome 133 onwards. You can use CSS to\ndeclaratively and more performantly style elements in these states. Let's see the video demo.\nWhen the item is snapped, it gets styled differently. Fun. You have to put parent container type\nscroll state on the parent element for the CSS and the browser to know that you want it to have this behavior. And now\nin that sub element you can have at supports container type scroll state.\nThis is the classes when it's applied also assuming that it supports it by default transition opacity.5 seconds\nease. And when it is not scroll state snapped we set the opacity to be 0.25.\nSo then it fades it in and out. And the default opacity I'm assuming is zero.\nDoes it show that anywhere? This does feel really nice to play with.\nI recommend again links in the description as always if you want to go play with this.\nThat's nice. That can't be right though because this is transitioning opacity but it's sliding up and down. That's not\nthe example being used here. almost certainly.\nDoes it even require stopping at each snap point? I don't know if that's working though. Interesting. I will call\nthis one partially supported. Also weird cuz this is definitely not the right code for this.\nYeah, there is no instance of opacity in this example. Yeah, it just slides it in and out. That's what I thought. I'm\ngetting so gaslit today. What do we have next? Tree counting functions. Staggered animations. Oh boy, that's actually\nreally, really cool. Common problem with animations in the browser is triggering multiple of them in an order is\nobnoxious because if one animation takes 200 milliseconds, the other one takes 300, the next one takes 50. Are you\ngoing to put a delay on all of them so they trigger in the right order? What if one gets delayed or like the rendering takes too long? Doing ordered\ntransitions is one of the most obnoxious things possible in the web. So almost nobody does it. Having the ability to\nspecify to the browser this happens and when it's done this happens and when it's done this happens is actually\nreally exciting and should enable some really cool animation stuff. Usual method to create ST animations for list items where each item appears\nsequentially requires you to count DOM elements and hardcode the values into custom properties. Yep, this method is\ncumbersome, fragile, and not scalable especially when the number of items changes dynamically. The new sibling index and sibling count functions make\nyour life easier here as these functions provide natural awareness of an element's position among its siblings.\nOh, so I don't have the ability to block one animation on another. I'm just using its index to calculate. Okay, so all the\nbrowns I just listed are still true, but at least you could do it. That's annoying. Let's see how it looks at least.\nOkay. Yeah. So you can have one, two, three, four like this again because it\nknows what index it is when it pops back in. But if you wanted to do something cooler, like have them fade out in\norder, then come back in in order, you're going to need JS still. Notice that there's no animation on the fade out because when you hit shuffle, the\norder of the elements changes, and it can't do anything until that's done. So they just do it immediately. Annoying.\nThere's a lot of potential to figure this out. The idea of like key framing and having blocking animations that\nblock each other via CSS has so much promise. We've also now confirmed that a\nlot of these are Gemini generated. So, it's possible a lot of the images and even some of the text is AI slop. Next,\nwe have the scroll into view container. Sometimes scrolling only the nearest ancestor scroller is all you want. This\nactually would be very useful for T3 chat. This is for nested scroll containers because if you have a\ncontainer that has its own scrollability in it, you don't want to scroll all the way through all of these. You just want\nto get this particular parent into view. Let's see the demo.\nYeah, that that perfectly showcas it actually. you have this subcroll on the\npage here and when you're going left and right and scrolling it's fine but if you click the snap and it uses the anchor\nsnap when it does that it's going to anchor the browser to that location. So\nwatch what happens when they click it scrolls down to put that right at the top. But if you don't want that you only\nwant it to affect the container that is being targeted not any parent containers. You can use container colon\nnearestest and it will only affect the one that is in there. So now watch with that toggled.\nThat's so good. I would argue again this is a thing that probably should have been the default behavior. It's much more intuitive. But yeah, good changes.\nSpeaking of good changes, nested view transition groups. View transitions are so fun and I'm enjoying them a lot in my\nside projects lately. And now they're going way further with nested ones. Retain 3D and clipping effects while\nrunning a view transition. Oh boy. Nested view transition groups are an extension to view transitions that let\nyou nest col view transition group pseudo elements within each other. When view transition groups are nested\ninstead of putting them all as siblings under a single view transition pseudo element, it's possible to retain 3D and\nclipping effects during the transition. I can already see all the crazy things the man's going to do with this. I can't\nactually. I have no idea what he's I just know he's going to go chaotic with it. That's view transition group elements. Another group use the view\ntransition group property on either the parent or children. You can give them all view transition names so you can\ncall them more specifically and the group you say which group to link to. In this case, nearest\ngroups get placed inside a new view transition group children pseudo element in the tree. To reinstantiate the\nclipping used in the original DOM, apply overflow clip on the pseudo element. We\ngot brais and una.\nOh, I see. It's the text is flowing in and isn't relative to the\ntransition happening. The paragraph text. Yes, thank you. I was realizing right as chat told me I want I don't want spoilers. My sensitivity to these\nthings has gone down with all the AI slop. Yeah, there we go. That's so much better. Yeah, the text doesn't flow out\nbefore the box because now they can be related transition groups. I do want to\nsee the obnoxious 3D rotation though.\nThat's beautiful. What do you mean obnoxious? That's how the whole web should work.\nBut if I turn off the nested view transitions, but leave that on, you'll see the text all comes out normal\nbecause it wasn't on the page before. It's not part of the view transition group. So the paragraph text stays.\nThat is really cool. absurd to have these things apply on 3D transitions, but people are gonna have\nsome absurd fun with this. I'm hyped. DOM state preserving move. Oh boy. Move\nbefore. I have a whole dedicated video I did on this forever ago. The TLDDR of this is if you have an element on a page\nthat has some state on it, moving it sucks because if I have like a video player and I want to change where it's\nmounted, like I want to move it from the main view to like a pictureand picture view, you're going to lose your state if\nyou're using the built-in view component for a video. Like if you just do a video element and you move it when it's\nplaying, you're screwed. Move before is a new API that lets you change an element's location without killing the\nstate of the element. As they say here, it works just like insert before, but it keeps the element alive during the move.\nSo, this means your video keeps playing, your iframe doesn't reload, your CSS animations don't restart, and input\nfields keep their focus even when you're actively reparenting them across your layout. Now, is this it wants me to play\nthe YouTube video for the demo? Uh, sure. Hey, my mute that. So now I have this\nvideo playing here. If I hit move with move before, keeps playing. If I hit move with insert before, the iframe gets\nreset because it is changing its location, but it loses the state when it recreates the element.\nVery good. I can't tell you how many times I've had to like move my video player logic out of the DOM and put it\nsomewhere else, like in the JavaScript world, so I could keep track of it when the element that was outputting the\ncontent moved. So [ __ ] annoying. I'm so happy this is a thing now. What browsers support it? Safari still\ndoesn't support it. [ __ ] course Safari doesn't support it. I'm excited about that one. And now we're in the\nfinal section, the optimized ergonomics. These modules aren't just plug-andplay. They're true chameleons, allowing users\nto redefine their interface, functionality, and aesthetics down to the atomic level. I am so sorry. This is\nthe most slop [ __ ] paragraph I've ever read. If anybody doubted that this is AI slop, I'm sorry. I was right.\nWe've now confirmed it. The they aren't this, they're that pattern. Yep. So, AI,\nthe advanced attribute function. Type values for attribute beyond simple strings. Oh, boy. Basically, can only be\nused within the content property of a pseudo element and can only return a value as a CSS string. The updated function expands the capabilities\nallowing attribute to be used with any CSS property, including custom properties. It now also parse attribute\nvalues into various data types beyond just strings like colors, lengths, and custom identifiers. So here we're\nparsing the data color attribute type color, red is the default. So in this\nexample, I I hate they put the color one right before. So like are they going to change the color to red when it's smaller? No, they're using this here by\nputting a attribute on the element for the stars. They are choosing how much of\nit to expose. So since this can be up to five stars, they take the attribute for how many\nstars it is, multiply it by 20%, and that's the percent fill. And they have a linear gradient for the gold color and\nthen transparent so that it cuts off at that point. And now it's very simple in\nthe HTML. You pass the data rating 4.5 and that CSS will multiply 4.5 by 20%.\nYou end up with a 90% number there. And then you apply that as a gradient with a stop to transparent. And that\ntransparent stop just immediately cuts it off. Very clever way to introduce this type of thing. I think that's cool\nas [ __ ] It's now way easier to without needing a custom framework or JavaScript\nor React or something, make your elements way more interactive by just\nimplementing different data classes on it. That's super cool. Not data class, data attributes. You get the idea. This\nis dope. This is going to make like HTML and CSS only component libraries are\nmuch more viable. Now, is this web components? No, web components are a mess of JavaScript as\nwell. This requires zero JavaScript which is really promising. Toggle event.s source. Find out which element was responsible for toggling the target.\nI have needed this for so many stupid [ __ ] things. When a popover dialogue or detail element gets toggled, it can\nbe interesting to know which element was responsible for toggling it. For example, knowing if the user pressed accept cookies or reject cookies button\nto dismiss a cookie banner is an important detail. The source attribute of the toggle event lets you know exactly that as it contains the element\nwhich triggered the event to be fired if applicable. based on that source you can take different actions. Okay, this is a\nlittle bit silly an example because here the logic is on the toggling of the\nbanner we have an event. If the event source is somebody clicking yes, we give them a cookie but if it's no, we don't.\nHow do we know what these elements are? Because we document.getelement by ID on them right here, which means we have the\nelement, which means we can bind an onclick to the element or an event to the element. I hate this example. On top\nof that, what happens if they get out of this banner with another thing like an escape key not handled? To be fair, it\nlooks like this one is a traditional div that's a popover auto. So, it's not going to be dismissible without you\nclicking something. Regardless, not how I think you should solve that.\nNot fun. Textbox features. We can finally flawlessly center text vertically. Oh man, I have wanted this\nfor so long. This is so good. A font's content box is defined by internal metrics, specifically the ascent and\ndescent that reverse space for accents and hanging characters. Because the visual boundaries of Latin text are the\ncap height in the alphabetic baseline rather than the ascent and descent, text will appear optically offcenter even\nwhen it's mathematically centered within a container.\nThere are some characters like the D here goes past the cap even though\ncapital letters don't go that far.\nObviously the disunderline goes a lot further but you get the idea. But that means that the way centering works sucks\nand we usually just use baseline. The textbox property makes finer control of vertical alignment of text possible,\nletting you flawlessly center text vertically. The textbox trim property specifies the sides to trim above or below or both. And textbox edge property\nspecifies the metrics to use for the textbox trim effects. With trimming both edges and setting the over edge metric\nto cap at the under edge metric to alphabetic, text will now be visually centered. If you thought it was hard to\ncenter a div, wait till you see how hard it is to center text. This is hilarious. This is multiple\nparagraphs of how to center text vertically. Finally, and as always, they're doing the magic\nthing with CSS where the first, second, and third values change different things. The first value changes the\nedge. The second value changes how to trim the top. And the second changes how to trim the bottom. So here, the top is\nat the cap. The bottom's at the alphabetic base. And the edges are all being trimmed. Let's turn off the edge\ntrim. I don't think it'll Oh, yeah. It does things here. Yeah. Edge trim. None. And now it just goes straight to the top\nand bottom. But if I trim both, it gets reduced. And we can choose how to trim\nover could be text. So that's as high as it could go. That is as high as the lowercase letters\ngo. Cap is the default I believe. And same on the bottom. The text bottom is all\nthe way as deep as it can go. The alphabetic bottom is where the like baseline of the letter is. Fun. I am so\nthankful I'm not that into typography because this would kill me. I am into shapes though. New CSS function for\ncomplex and responsive shapes. The new shape function. Firefox is the only thing that doesn't support it. Beautiful. Wonderful. The new shape\nfunction lets you clip an element to a complex non-poly responsive shape in CSS. This is a great\noption for clipping masks using clip paths and it works seamlessly with CSS custom properties to define coordinates\nand control points making it more maintainable than SVG shapes. Thank god. That also means you can animate custom\nproperties within a shape to create dynamic and interactive clipping. That's so cool. That's so cool. This is an\nimage. You can see the original square image there. But we have a shape that is\nbeing defined here. Here is saying that the Brazo doesn't support it in the CSS. We can see clip path shape from 0% 20%\ncurve to 120 with 25. Okay, maybe the syntax isn't better than SVG. This does\nhurt a little, I'll admit. Kind of wild to read this. But yes, as chat is realizing, this would go crazy for\nanimations. Oh, look at that. Anthropic is degraded. What a surprise.\nLet's just do 5.1 instant. Sure. I want to animate the following shape CSS so\nthat it waves like the wavy part shifts\nto the left over time. Write the animation transition CSS logic. Let's\nsee how well this works. Yoink.\nWorking beautifully. Try that again with a smarter model.\nGh, gross. Not what I had in mind at all. It is animated,\nbut it's also gross. Here's one. They use a blob generator for shape to create a fun framing effect. That is kind of\ncool. Oh, and of course, if statements. I already talked about this and how cool it is to have if in CSS where you can\nactually apply conditions. If media oriented landscape row else\ncolumn. Look at that. that not useful\nand custom functions which I've also already covered. I did a whole video on CSS functions and how potentially useful\nthey can be. They've also expanded the range syntax to make it easier to combine with these things. You have this\nHTML that has a data rain percent. Data percent is a custom property. Give it a\ntype using the attribute helper and then you can use it within a range style query. So here we have rain percent\nwhich is us converting the data ring percent into type percentage. And now if the style of rain percent is greater\nthan 45% then the weather card should have this blue light blue gradient on it. Now when the chance of rain is\ngreater than 45% the card gets a blue background. That's really cool to have this in CSS. Real logic that doesn't\naffect the execution logic being separated out of it is really nice. Not dealing with this in JavaScript and\nhaving the way things look defined externally and the way things behave defined to the JS is a mostly good\nthing. And I like seeing this range can also be used in if statements as well. So background if style rain percent is\ngreater than 45 blue else gray. Nice. Real good stretch sizing keyword make an\nelement fill its containing block regardless of the box sizing. Interesting.\nIs this going to fix the thing where if you have a sub element and you didn't put the right min width on something in\na flex box that it won't stretch to the right size? The stretch keyword is a keyword for use with CSS sizing properties like width and height that\nlets elements grow to exactly fill their containing blocks available space. It's similar to 100% except the resulting block size is applied to the margin box\nof the element instead of the box determined by box sizing. Yay. Using the keyword lets elements keep their margins\nwhile still being as large as possible. That's really cool because you can keep the margins of the inner element while\nstill stretching to fill the parent element. Yes. Oh, and corner shape. Oh\nboy, I'm excited about this one. I built a dumb project a long time ago. Dogecoin\nsimulator. Super inspired by um what's it called? Or universal paperclip. And\nmaking these trashy handdrawn looking borders for the boxes for the buttons\nwas the hardest part of the project by far. Being able to do weird [ __ ] with corners going to be fun. Although I'm\nscared they're going to give us too many traditional ones and not custom. Yeah. Round bevel notch scooper squirle.\nOh god, are you actually going to use scoop? Bevel's kind of cool. Notch is really\ncool. Square is how it works by default, isn't it? And then squirle, the classic scoop\nterrifies me. I hate this. I hate this a lot. Oh, look. It's AI generated. Now you can\nchange the shape of the corners by specifying multiple for even more control. can use the super ellipse\nfunction to create any continuous curve allowing for fine-tuned and unique corner designs. That's what I was\nlooking for that I can abuse. And that's the end. What a set of interesting\nthings. This is really, really cool. I am thankful for the Chrome team for\ncontinuing to iterate on these things and publish such useful resources. CSS changed a lot over the last year and there's a lot of cool stuff in this that\nI'm excited to play with. I'm curious about you guys though. Are you tired of dealing with CSS stuff like this? Or are you actually excited about all the cool\nthings that have been enabled in Chrome? Let me know what you'all think.\n"

ERROR: type should be string, got "https://www.youtube.com/watch?v=kEPLuEjVr_4\n\nSo close to Opus at 1/10th the price (GLM-4.7 and Minimax M2.1 showdown)\n\nYou didn't think we were going to get through the holiday week without more new model drops, right? Do you not remember last year when Deep Seek came\nout and just dropped a ton of things at the end of the year and the beginning of January right after? We're there again.\nWe didn't get one new model, by the way. We got two from competing labs. Both are open weight. Well, kind of. We'll get to\nall that in a sec. First, we have GLM 4.7, which benchmarks like this. We also got Miniax M2.1, which benchmarks like\nthis. Apparently higher than Opus and Sonnet. We have a lot to talk about there. I've seen claims all over the\nweb. Everything from Anthropic is dead to Opus 4.5 has been destroyed with a model 10 times cheaper to OpenAI is\ngoing to go out of business. I think that's all a little much. Are these the best new models ever? No. So, put down\nyour counter. I I see you there incrementing the counter. No. Put it down. Put it down. We're still on Opus.\nThat said, these models are incredible and I am blown away with what they can do. One\nmore so than the other and I spent a lot of time with them. For the last two days, I have been grinding non-stop in\nVS Code using Kilo as well as in my terminal using open code to try and get\nthe best I possibly can out of these two new models. I am very deep in token\nusage right now, but I feel like I know everything I need to to responsibly cover these models. right after a word\nfrom today's sponsor.\nGoing to start with a quick overview of the models from the labs themselves and then we'll dive into what it actually looks like to use them to build real\nfeatures because you might have seen I've been building a lot of stuff with Opus recently. So I went to try and rebuild those things with these new\nmodels and the results were fascinating to say the least. First we have GLM 4.7. It came out slightly before M2.1.\nAlthough I did have early access to M2.1, I only started using it very recently. GLM 4.7 I got the day it came\nout and everybody else had access. GLM 4.7, your new coding partner, is coming with the following features. Core\ncoding. This is a big thing. These models are really, really focused on coding. GLM 4.7 brings clear gains\ncompared to its predecessor 4.6 in multilingual agentic coding and terminal based tasks. It did much better in SWE\nbench. Did much better in SWE bench multilingual and pretty good on terminal bench 2, bumping up 16.5% from the like\n20-ish range to 41. With vibe coding, it's now a major support in UI quality. We'll definitely test that out. Produces\ncleaner, more modern web pages and generates better looking slides with more accurate layouts and sizing. It's\nalso much better at tool usage, in particular with browsing stuff. And it's better at complex reasoning, meaningful\nbump on HLE. Here are the benchmarks. This is compared to 4.6, their previous\nmodel, DeepS 3.2, which was really, really good openweight model. Less practical, but still very intelligent.\nClaude Sonnet 4.5 and GPT 5.1 High. And there are a few benches where they win\nor come really close to things like DeepSeek. Most of them, it appears they are neck andneck with GPT 5.1 and quite\na bit ahead of Sonet 4.5. They didn't put Opus in this bench. They probably ran it before Opus dropped. You get the\nidea, though. It seems to be a very competitive model especially in the more coding stuff which almost all of these\nare live code bench sbench verified terminal bench how to bench which is a tool calling bench and browse comp as\nwell it's doing very well across all of these and here we can see something very interesting which is that in a lot of\nthese benches especially the reasoning ones Gemini 3 pro appears to be the winner which kind of showcases that\nthese shouldn't be trusted that much because as smart as Gemini 3 pro appears to be like it the amount of knowledge they packed into that model is crazy\nactually using It is miserable. I have not had a good time with three pro at all since I started using it. But we go\ndown to the code benches like SWB verified, multilingual, all of these. You'll see they are very competitive\nwith the best options. Multilingual they are losing to I believe this is Deepseek. Yeah. So losing to Deep Seek\nand the multilingual, but they are neck andneck with everybody else. Honestly, they're pretty close there, too. It's good numbers. These numbers look really\ngood. The visual design is one of the biggest wins though. like it can make cool looking websites like this\napparently. And if you compare to what it did before, oh god. Yeah, this is like the\nusual AI slop websites. I expect 4.7 made something actually unique and cool\nlooking. I love the at 2023 though. It gives you an idea of the training data.\nOoh, apparently it knows 3D space pretty well. Uh I did actually run this on\nSkatebench. Let me find the screenshot. I didn't post the numbers because I'm busy and\nlazy and doing a lot of things at once. GLM 4.7 got a 66% which puts it neck\nandneck. Actually, I think it puts it Yeah, it puts it over DC 3.2. It's the highest scoring openweight model on my skateboarding bench, which is both like\na weird niche knowledge bench and a 3D spatial recognition bench cuz it's you\nneed to have some spatial reasoning capability to understand the rotation that I'm describing to name a skateboard trick properly. So, I was surprised to\nsee 4.7 did so well here. And also Mini Maxm 2.1 did very poorly here. All\nthings we will get to. This is cool as hell though.\nDid I have a 4.6 comparison? Yeah. Uh yeah,\nslightly better. Just a marginal improvement.\nAgain, with the design taste, it can make things that actually look decent. Very cool to see. And of course, it is\nfully open weight. You can go download it from HuggingFace. Good luck running it, though, because it's 717\ngigs of data. Even the more compressed versions that I've seen floating around are looking like 300 plus gigs. It's a\nbig model. The 358 bill model, not trivial. You can absolutely run it on\nthings that a human can purchase, but you're going to have to put effort. And you're not running this on consumer\nhardware. We'll play with it a bunch in a bit, but first we have to talk about the other new model that dropped, Miniax\nM2.1. I did get early access from the Miniax team. They gave me a free sub to their $50 a month tier. I barely used\nit. I ended up just going through Open Router when it dropped. I used their thing early so I could have early access. So, uh, yeah, they hit me up.\nThey talked nice to me. They gave me early access. They follow me back on Twitter. So, account for some bias there. But I do actually really like the\nteams at ZAI doing GLM. I really like the team at Moonshot doing Kimmy. Kimmy K2 is my favorite model to talk to right\nnow. And I really like the Miniax guys as well. So account for bias across all of them, even if the only one that actually gave me free stuff was Miniax.\nI don't care about any of that. I just want a way to play with the stuff. They've all been awesome to interact with, much more so than the big labs.\nOpenAI is still really, really good, too, for to their credit. But I'm blown away with how good the interactions have\nbeen with these teams. That all said, let's take a look at how the model actually behaves. So the key highlights\nof 2.1 are its exceptional multirogramming language capabilities. They really tried to focus on it working with lots of different languages.\nPreviously, it was mostly focused on Python, which is why it did so well in so many benches, but sucked to use in the stuff I build. They have\nsignificantly enhanced its ability to do Rust, Java, Golang, C++, Cotlin, Objective C, TypeScript, JavaScript, and\nother languages. I use a few of those. Very cool. On the note of programming\nlanguages, I've had a video I really wanted to do for a while now about how different models behave with different\nlanguages. Of all people, Tencent published a really in-depth benchmark comparing how the models perform across\nprogramming languages. And you'll never guess which language performed the best. It wasn't Python at a 65%. It wasn't JS\nat a 60.9%. It wasn't Tyer at a 61.3 or even Ruby at\nan 81. It was Elixir at a 97.5.\nYeah, if you want me to do a dedicated video on this, let me know cuz it's been [ __ ] with my head for months now. The\nother thing they say it improved in is webdev and appdev. It's a comprehensive leap in capability and aesthetics. I\nhaven't actually tested the aesthetic side of this model much yet. We will do that as we bench. It also is apparently really good at iOS and Android\ndevelopment, which is super cool. I personally will not be testing that, but I have friends that probably will, and they might even show up in the comment\nsection. It has enhanced composite instruction constraints, enabling office scenarios. Interesting. It systematic\nproblem solving capabilities have further upgraded. The model not only focuses on code execution correctness, but also emphasizes integrated execution\nof composite instruction constraints, providing higher usability in real office scenarios. I'm sure those words\nmean things. It's more concise and efficient with its responses. This is actually very nice. I've noticed the difference. It's not super long-winded.\nIt just responds. It has outstanding agentto tool scaffolding generalization capabilities. So like when you hand it a\ndifferent agent to harness, it behaves relatively well in them. So hand it something like kilo code or open code,\nrue code, blackbox, factory AI, cla code, whatever. It can for the most part figure it out. It understands skillmd,\nclaude MD. They also really encourage that you hack cloud code to use it, which I didn't do this time because last\ntime it was obnoxious. Maybe in the future I just didn't want to gut my cloud code config. If anybody has an\neasytouse open source project where I can have like cloud code as default cloud code but also have a cloud code\ninstance that is configured differently to use M2.1 so I can easily swap between stock Claude code and a hacked up\nversion. That would be really cool. I am too busy to build it. But if somebody has like F&M for cloud code hacks or NVM\nfor cloud code, let me know because that would be really nice. They also have high quality dialogue and writing. It's\nno longer just strong in coding capabilities. Apparently, it's better at writing. Nothing comes close to Kimmy K2\nright now, so I don't even care to check it. And then they have a bunch of quotes from all of the harnesses. Again, they're working with everybody and\neveryone has been saying it's really, really good. We found that M2.1 handles the nuances of complex multi-step\nprogramming tasks with a level of consistency that is rare in the space. By providing highquality reasoning and\ncontext awareness at scale, Miniax has become a core component of how we help developers solve challenging problems\nfaster. We look forward to seeing how our community continues to leverage these updated capabilities. Yep, you get\nthe idea. They like them. They then claim that it outperforms Cloud Sonic 4.5 and closely approaches its opus.\nThis is kind of [ __ ] but we will get to that momentarily. They also created their own set of benchmarks,\nVibe Bench, where it is slaughtering. They are testing against GLM 4.6, not 4.7, because these all came out around\nthe same time. So, nobody had time to go rerun all the benches on the newest thing. They really like using it with cloud code. You'll see this in a lot of\ntheir demos and even with their own onboarding. But again, I didn't want to hack up my cloud code. So, instead, I\ngot way deeper on open code. And there's a lot to love about open code, but it also has its problems. I have a video\nthat I was really hoping would be out by now, but it's probably going to come out soon after about why I'm loving Opus so much. And I went and built a bunch of\nactual features using Opus 4.5 in T3 chat and kept going to harder and harder\nthings to see if it could do it. And it could like weirdly consistently. One of\nthe things I built was an archive feature so that you could archive a thread in T3 and have a page where you\ncould see all your archived threads. This feature was surprisingly easy for Opus to build. I had it make a plan. I\nliked the plan. I told it to build the plan. It built the plan. Whole thing didn't take very long. It did a very good job. I did not see how many tokens\nit used because I was using cursor for this. So, I decided to do the same build, but this time with various\ndifferent models in open code. This is a different scaffolding. This is a different harness. So, it is going to\nhave different capabilities, limitations, etc. But honestly, I was really impressed with the capabilities I\nwas getting out of open code. I will copy the prompt. I can't attach the current state of the UI for reference\nsadly because I can't paste images into open code and I don't think either of these models can take images anyways. So\nlet's do a new session in open code. Going to switch over to GLM 4.7 which is\ncurrently free on Open Code. So if you really are a cheapsake, take advantage of that while you can. We'll talk about\npricing more in a bit. Just know these models are hilariously cheap and will probably get even more so as more places\nthat can host these openweight models start hosting them. Actually, that was one other thing I almost forgot. The Miniax M2.1 weights are not out yet. M2\nis open weight. M2.1 should be open weight. They haven't indicated that it won't be. They just haven't put the\nweights out just yet. Hopefully, probably coming soon. It might even be out by the time this video is live.\nAllegedly, the weights for M2.1 are going to drop on Christmas Day. Very\nexciting. So, the weights will be out super soon. And according to Miniax themselves, M2.x, 2.5, M3, lots coming.\nLet's see how far we can push it. I'm excited. That also means we don't know how fast it can run on consumer hardware\nor how big it is, but I can speculate and I love to speculate. So, we'll be doing that plenty. So, back to actually\ntesting the models. GLM 4.7 open code plan mode pasted those lines. You can't\nsee them because of how the UI for pasting works. When I press enter, you can. And we get an awesome UI for this.\nYou get the actual work going on here with a scrollable pane. You get a sidebar that has actual useful\ninformation. And it has access to TypeScript, LSP, and Oxlint so that if it makes lint errors, they will be piped\nright back to the agent. So you don't have to run a command to see mistakes. Really cool stuff. So I started\nthinking, it knows it's in plan mode, so it has to explore and figure out what to do. And they have sub agents that are\nreally cool in open code. So I can control X and go left and right to see these sub aents that are explore agents\nthat are exploring the codebase. And then I can go back to the parent primary agent that all of this data is going to\nwhen they're done doing their exploration. They'll summarize what they find and send it back up to the main agent. So they're not just clogging up\nthe context entirely. There's also context compaction in open code that you can run manually. But don't run it when\nyou're doing a generation or it will break a lot of things. Learn that lesson the hard way. Now, I would love to sit\nhere and show you guys how long this takes in real time, but I don't have an\nhour to waste. Yeah, the planning and actual implementation. The first time I ran 4.7 on this codebase for this task\ntook an hour. Yeah. So, instead, I'm just going to show you guys a run I did\nearlier. Here's the first time I did this. It cut off my message history a\nlittle bit. I'm not quite sure why or if there's a way to see the whole thing, but it thought for a while and wrote a\nplan that had a bunch of problems. Asked I think two or three questions. One was good, the other two were garbage. I\nanswered the questions. I steered it a little bit to give it a better idea of what to do. Then it seemed like a good\nenough plan. So, I hit build and it started building and it struggled a lot.\nIt kept miswriting the React code and immediately getting a ton of errors from\nthe LSP, freaking out about them, and then concluding things that weren't true. Yeah, it was trying to write an\noptimistic update that it didn't need and kept breaking [ __ ] Doubling up code where it would put this like things in\nthere twice. Failing tool calls to the search tool and the edit code tool that\nexist within open code. Sure, I can find some of those tool call fails in here if I look a little bit. Yeah. And here it\nwas confusing the different use query hooks and what they return and breaking a bunch of [ __ ] as a result. a lot of\nwait actually looking at this it also got very confused that we were using\nTRPC for some things even though almost none of the stuff that this feature touched involved the TRPC endpoints those are mostly for legacy data and\naccount management stuff everything else goes through convex this whole feature should have been convex I even indicated that in the original prompt but it still\ngot very very lost it seems like it doesn't pay enough attention to the original prompt as it creates more and\nmore context as it goes and I learned that thing about the context compaction during this run because it got to like\n80k tokens. So I ran a compaction so that it wouldn't go over and then it broke the generation. It kept getting\nmore and more confused. So I eventually interrupted and told it to not worry about those type errors and to not do\nthe optimistic updating. Told it to continue. It got confused about what it was done with. Finally went back and\nfinished. ran the dev server that was already running sleep 5 and and curl-s\nlocalhost to get the actual HTML content killed the original one that I had\nrunning and then finally had something that seemed to be working. Also, you have a bunch of remaining tool to-dos.\nFinish them. Don't touch my [ __ ] dev server. You can tell how I felt at this point. Yeah, apparently open code will\nautocompact. So, I didn't need to do that manual compaction that broke things. So, that's cool. Now you know that I'm back there in a thread or in a\nsecond. Check out demo. See the one commit I have on that branch. This took an hour. Yeah. Now I'm in my dev server\nby my little archive button. I have this thread that is archived. I can unarchive it and go back. Cool. It works. I can\nrightclick this archive it. It's still showing that thread. So it doesn't navigate you out of it when you do the archive if it's the thread that you're\nalready on, which is annoying. I didn't specify that in the prompt, so it doesn't necessarily know. To be fair, none of the models got that right. I did\nspecify during gen that it should hide the archive icon if there are no archived threads, and it failed to do\nthat. I also don't love how it did this in the UI. Like, this treatment sucks. You can't tell what the current state is\nat all, and it didn't put it in the URL, which is where I think it probably should have gone. That all said, it did\nit in a way that works after a lot of help and having to rerun it once or twice. Just so you can see what a\ngeneration like this looks like. It just finished the planning. Took three and a half minutes. Not bad, especially for\nfree, remember? And it asks some questions for us to answer. Should archive threads appear in search results? If yes, we need a separate\nquery or filter param. Should archive threads still be accessible via direct URL? And where exactly would I like the\narchive button to be placed in this sidebar? These are all good questions, and I can go through and answer one, whatever, two, whatever. You get the\nidea. I can absolutely do that in here. It's really nice, really convenient. Not the best plan mode overall. It doesn't persist the plans particularly well. A\nlot of other tools will just write it into your codebase, which I think is awesome. But that's just one of the implementations that I did. I also did\nthis with Miniax M2.1. God, scroll state is so broken in this sometimes. Also can't get all the way to\nthe start of this one sadly, but I again told it make a plan. It made a plan. It\nasked some questions. I made a mistake here. I thought that open code would do new lines the same stupid way claude\ncode does where you have to escape them. You can just shift enter. It's fine. So yeah, learned that lesson, made that\nmistake, canceled and retrieded. Question one, archive view location. Where should the button be? I told it bottom of thread list. Search behavior.\nShould searching threads include archive threads or only active ones in the empty state. Weird thing for it to ask for. So\nI said bottom of thread list for the first question, only active threads for the second and for the third one. Archive thread should only be visible\nwhen you click the archive button. It should not be visible if you've never archived a thread. Specifically, should not be visible if you've never archived\na thread. We'll come back to that. So, it wrote out its plan, told it to go\nnuts, and it did. It is worth noting that at no point did it have to compact.\nIt only did about 66k tokens total, and it can do 200k context. So, fine. And it\ncost about 4 cents for the whole run. That number might be off cuz I switched to the UI version for a bit, but I've\nnot been able to get this generation to cost more than 10. Hilariously cheap. Absurdly cheap in comparison. Not to\nspoil things coming up, but uh yeah, Opus is a little bit more expensive.\nJust just a a lot. Yeah. So, it made the to-do list look good, told it to go\nnuts, and it cranked cranked for a while. Here's where we see those old string not found in content errors.\nThere was a lot of this throughout. It seems like there's something about how the edit tools implemented in open code\nthat these particular models don't like that much. I did not see this error when I ran things with Opus. So, it does seem\nmodel specific. And I know DAX doesn't believe that different models need changes to their system prompt and different harnesses for code stuff. I\ndon't agree. These models will probably need a little bit of steering. That said, it did eventually finish in about\n7 minutes. It didn't hide archive threads, though. By default, the archive threads were appearing in the main view,\nwhich is not how it's supposed to work at all. So, I complained about that. It grinded a little more, fixed it in about\na minute and 20 seconds. I decided to ask it if it was forgetting anything else because there was a couple other little things. Specifically, it was not\nhiding the archive button. One thing I might have missed, when you archive a thread, the archive button should appear, but the has archive threads\nquery should handle that. There was a problem with that query, though. I have no archive threads and the archive\nbutton is still appearing. The reason this was happening is because it wasn't getting data off of the return value.\nThe use query hook returns data loading error and a few other things and it was just checking if it returned an object\neffectively. So it tried changing has archive threads to has archive threads equals equals equals true and then it\ngot a type error through the LSP. Very good. The comparison appears to be unintentional because the type of use\nquery result in boolean have no overlap. That might have been oh excellent actually. Regardless, now that it has\nthat error, it realized that the use query returns a usequery result object, not the raw data. I need to use data to\nproperly access the actual value. This is actually a really good example of how the tooling, the model, and the codebase\nall have to interact. Well, this original thing was not a type error. It was an object check effectively for\nblocking a render. It was bad code. Not correct. A llinter probably should have caught that. Not trivial to catch, but\nannoying. The models doesn't have as deep of an understanding of TypeScript as some of the other models, especially from the major labs. But it does know\nhow to deal with errors well. If it gets an error, it can go fix it. And this is why LSP support is awesome, which is why\nit's cool that Cloud Code finally added it almost a year after shipping. Now,\nmost of the CLIs have LSP, which is language server protocol. If you don't know, it's how the language server,\nwhich is what's checking your code, communicates to your editor about what's going on. Cool. Behaving much better.\nStill has these old string and new string replacement issues. It just doesn't get the edit tool very well.\nEventually hits this properly, gets it edited, has archived threads, question mark.data is true, then render,\notherwise don't. Awesome. Also took a minute 20 seconds to fix that. I told it to move the button. It did in a minute.\nAnd now it's all done. All of that cost me under 10 cents. I don't necessarily trust the 4centent number because I was bouncing between their UI and their app\nhere. You get the idea. Make sure this all deploys. All looks good. Oh. Huh. No\narchive button. That's cuz I have nothing archived. Let's archive this. Now I have something archived. Dope.\nSolid UI put the buttons in the right place. The back to threads. I actually quite like it does keep the thread open\neven when it's archived. So if I have this open, go back, it does that. And if I unarchive, the back to threads is\nstill there. So I can go back here. It fully works. I had to steer it, but it\nworks. But that's also kind of what I've been digging about this model. It is\nrelatively fast. It gets to the problem and fixes it relatively quick. It's solid at planning. It's really solid at\nfollowing the plan. and it doesn't lose track of the instructions the same way a\nlot of other models do for these longer running tasks. I was able to get this to just grind on a feature for over 400\nlines of code and for like 20 minutes without issue. It can just go out on these journeys, get feedback, make\nchanges, and figure itself out. 4.7, not so much. For my experience, 4.7\nneeds to be broken into small tasks to complete. If there's a more general\nending to the task that it has to get to via multiple paths, it gets lost a lot\nmore aggressively. It needs to be steered quite a bit to finally get where it needs to be. Miniax just needs\noccasional reminders here and there as well as context that it can't get because it doesn't have any visual\ncapability or the awareness of everything else going on in general. Honestly, it's a you can tell it's a\nsmall model, but it is a good model. It is a fast model. It kind of feels like a\nWalmart brand opus. As silly as it is, it's good enough to get real work done.\nI also went and tried it in Kilo and my experience here was quite a bit different and honestly meaningfully\nbetter. I'll give you guys a look. It started by creating this plan cuz I had it in the architect mode I believe,\nright? Yeah, architect mode. I had an architect mode. Told Minia M2.1 to make a plan and it did. Font's a little\nsmall. Sorry about that. It also made the plan as a markdown file in the codebase, which I love. This is the plan\nit wrote from the same prompt. Add an archive chat feature that allows users to archive threads from the rightclick context menu and access archive threads\nvia an archive button in the sidebar. Proposes schema changes. It notice\nbecause I actually have visibility in the schema for this purpose and just never used it. So, it doesn't actually have to make any schema changes. Thread\nqueries list shows all threads. No visibility filter. They have to add that. Go to mutations. Update mutation\ndoes not include visibility fields. This is interesting. This is the current state analysis. So it's not just giving\na plan. It's an analyzing where things are and then writing the plan after which I don't love. I don't think it\nneeds to have this in the actual plan. This feels like reasoning leaking in. But the actual implementation plan\nsolid. Update the query to filter for visible threads only. It does actually write code in the plan a lot more which\nI don't necessarily love because it could just be writing the code. The point is to have an architecture plan,\nnot to write the code, but it does a good enough job. It creates the new list archived query, which is how it gets the\narchive data out of visibility to update mutation. Phase two, the front end side. So, it broke up the back end and the\nfront end changes. Good call. Phase three, UI changes, actually accessing the data from these new hooks.\nThen it tried to render this. I don't know if I can actually if there's a way to preview here. I don't use VS Code as\nmuch anymore, believe it or not. Open preview. Here we go. Yeah, it doesn't render this. I don't know if they have a way to do that or not, but yeah, it\nwould have a flowchart if I had mermaid rendering in here. So, which files need to be modified? Test considerations,\nlow, medium, low, complexity. All looks good. You know what I'm going to do? I'm going to resave this plan so it has it\nagain. Let's tell it to build mini 72.1 implement\narchive chat feature MD. If I really wanted to give it the highest chance of success, I would also tell it something\nlike make sure to create a to-do list based on the details in there or something. Hopefully, it will figure out\nthat it should do that. I also forgot to mention Kilo Code has sponsored videos in the past. They probably will again in the future. I'm not using them because\nthey're a sponsor. I'm using them because it's the general use whatever tool tool that I like the most for\ntesting in a real code editor with a bunch of random models. So, if you want a tool that lets you use literally every\nmodel ever and compare how they behave across different use cases, use T3 chat. But if you're looking for code cases,\nKilo is really cool. They did create the to-do list, broke it up into different phases, and I spoiler did this before.\nIt worked fine. Seemed to have better awareness of the code base, but worse awareness of linting and type errors.\nIt's my understanding that the only way context gets to the model in open code\nis by tool calls looking for things whereas kilo and other tools like cursor\nwill actually index the codebase that you're in and give better methods of access to the models. I could be wrong\nabout how kilo does it but I know this for a fact with cursor that it indexes your codebase and they have their own crazy methods for search that get\nimplemented around the traditional search tools to make the results more contextual for the models. This will\nalso likely take a bit. So, while that's going on, I want to talk more about the performance and the pricing in particular, like the speed side of\nthings. Grab over to open router. We can take a look at M2.1 here. Through the\nofficial miniax hosting, it's pulling about 75 to 76 tokens per second. There aren't other hosts yet because they\nhaven't put up the weights yet. It is our understanding that they will, but they haven't yet. There's even a link to\nthe weights on open router that is currently a 404 that I expect them to resolve in the very near future with GLM\n4.7. It is already out. A lot of hosts aren't pulling that fast of speeds with\nit. Zi actually is themselves. They're able to get over 150 TPS. Novidia is hitting 60 and Parasel's hitting 80, but\nother hosts are as low as eight tokens per second. So, it's a big model. It won't be as easy to host. That's\nexpected. Miniax M2, however, was quite small.\nOkay, not that small. 229 bill per 358 bill for the new ZIGLM model. About 2/3\nthe size. GLM 4.7 is a 32 bill per activation during its mixture of expert\nstuff, whereas MiniAX is only activating 10 bill, which helps again with it\nrunning faster. If you would like to see this in action, I did download the original M2's GGUF version, which works\nsurprisingly well on my MacBook. It's compressed down to about 78 gigs, which on my 128 gig MacBook with unified\nmemory can fit fully in memory. Takes a bit to load into RAM because it's literally loading 80 gigs of data into\nmemory right now. If you ever wondered why hosting models is so annoying, it's because when two different people want two different models, the time it takes\nto load one out of RAM and the other into RAM is longer than the request will take a lot of the time. Let's have the model write some poems about JavaScript.\nAnd reminder, this is M2, not 2.1 because we don't have the weights yet. This is meant to give you an idea of what performance will probably look like\nroughly. Probably should have lowered the reasoning on this run. Still cranked. ran about 35 TPS on a consumer\nMacBook. That's pretty cool. Again, it shows these models can actually be used\non consumerish hardware. I will still be using it over API calls almost certainly\nand they will of course be supported on open router and I would expect that to be the case indefinitely. Still grinded\nin kilo. It's pretty cool that I can go for this long without intervention but also takes a bit. Oh no. Yeah, this is\nthe problem I had before that I wanted to showcase. It seems like it goes much better. And then I hit this invalid\nparams tool results tool ID not found. And I'll retry and it just is dead. Like\nthis task is now over. This thread is killed. I have to go create a new thread and tell it to continue, which sucks cuz\nwe just spent 40 cents. It got halfway through phase two of three phases and died. The these things take a bit. This\nis the rough edges when you're going to use tools like new openw weight models in open- source plugins for VS Code.\nThere are just so many more layers. Yeah. And once again, just as a reminder, their recommended way of using\nthe model was to use it with cloud code. I did not do that because I don't want to modify my Claude code to use models\nthat it was not built to support and then have to deal with the config swapping constantly. It's very annoying.\nSo, I didn't set that up, but you should definitely give it a shot because their plans are much cheaper and the model's\npretty dang good. And if you were curious about how Opus handled this, it cranked straight through it. Didn't like\nthe UI part. Gave it more suggestions, it fixed those relatively quickly and\nall just worked. Tried this in cursor, it did great. Tried this in open code, it did pretty good. Not quite as good as\nit did in cursor. Opus cranked through this, but also cost significantly more to do this. quick cost comparison, but\nplease remember the input and output costs aren't just like these numbers, and that's the whole story. The amount of tokens being generated matters a lot\ntoo. And these models are chatty. They generate a lot more tokens per line of code than other models do, especially\nduring the reasoning phase. So, just because the costs are cheaper doesn't mean the actual cost to run it is\ncheaper. It is also cheaper, but it generates a lot more tokens. So, just know that going in. Opus 4.5 costs $5\nper million tokens in and $25 per million tokens out. Five and 25. Sonnet\nis three per mill in and 15 per mill out. 25 per mill out 15 per mill out.\nRemember those 4.7 40 cents per million in $1.50 per million out. It is a tenth\nof the price of sonnet 4.5. It is way more than onetenth as intelligent. It is\ncapable of a ton, especially if you force it into the box of just making direct code changes. But that's not the\ncraziest one yet. Miniam 2.1 is 30 cents per mill in and $120 per mill out. Do\nyou understand how insane that is? These models are comically cheap. This is 12th\nthe price of Opus. And Opus just got a huge price drop. Opus 4.5 is three times\ncheaper than Opus 4.1 was. And this is 20 times cheaper than that. This is 60 times cheaper than Opus 41. And I would\nactually consider this model on par with Opus 41. Opus 45 was a massive jump. They should have just called it five.\nThis model is similar performance to Opus 4.1 for literally a 60th the price.\nAnd just for comparison, the new flash model that I am still very excited about. It sucks to use in actual like\ncode type usage or talking to or any of those types of things. But as a background worker doing random tasks\nlike parsing data and giving you a JSON blob out of it, it is really solid for all those things. In this model's 50\ncents per mill in and $3 per mill out. So it's still twice as expensive or more as M2 in particular with the output\ntokens. It is insane. And it's actually possible for the price to get cheaper because it will hopefully be open\nweight. I'd be very surprised if they don't put up the weights in the next day or two. Other hosts will get more competitive, faster speeds, more bulk\nprocessing options, better caching, all those types of things, and can hypothetically get the price even lower. Probably won't bother, but they could.\nRegardless, this is awesome. It's so cool to see models this good, this cheap, self-hostable, pushing the\nstate-of-the-art forward. But I still want to play with UI. I still really want to see it show its capabilities for\nUI stuff. So, let's set something up for it to do that. I am now generating\neveryone's favorite, the image gen studio UI and seeing how the quality\nlooks between GLM 4.7 and Miniax M2.1. High concurrency usage of this API.\nPlease reduce concurrency or text overlay errors. Wonderful.\nMight have to switch providers. GLM seems like it probably would be the\nbetter model if you wanted to use one for things other than coding like basic analysis just abusing it for tool\ncalling object gen stuff like that. Mini Max is much better from my testing at\nlong tasks, at generating plans, at going the 9 yards. Like it it has that\nclawed feel to it where it can work for longer. And the only models I've seen\nthat have this are GPT5, every anthropic model since Sonnet 4, and now Miniax M2\nin 2.1. I did not use M2 very much, so I don't know how much better this is from\nM2, but its ability to do longunning tasks has been blowing me away for an open model especially. I know I didn't\nuse plan mode for this UI gen. It's just UIG. It shouldn't need a plan. Look at that. Using npm even though this is a\nbun project. We love that it used bun add to add the packages and then ran with npm run.\n[Music] That was GLM doing that. By the way, I would really like if Open Code added the\ntokens per second here. It's crazy. They have like working clickable UI. Oh, it\nsince it ran the dev server, it might have gotten stuck or thinking the Okay, now it finally went. That was funny.\nYeah, it ran that dev server with the and and now lost track of it and is\nconfused. It did this before, too, where I had to yell at it. It just doesn't know how to run dev servers, which is\nvery annoying. Miniax doesn't bother doing it. I do have a rule in my cursor rules telling models to not do this if I\nrecall. Going to be lazy and do this to hide my email. I do have that rule in cursor. Well, close enough. Don't call\nbuild commands unless you really need to. They break my dev environment. You can run type checks all you want. Yeah,\nthese are still going. Seems like GLM 4.7 was able to get itself out of the hole it dug itself into with those dev\ncommands by calling kill commands to get rid of them. It told me that it's running here. I don't I hope it's not\nstill running there. It is not not that either. Oh, I might still have another dev server running.\nLet me double check that. No, I don't. Okay, there we go.\nUh, god, this is the GLM version. CSS is\nhard. Something appears very broken. None of the spacing\nis right. Everything is compressed into the corner. What is going on? Did you\nbreak the Tailwind config? I thought this model was supposed to be good at\nUI. Miniax is still grinding. It's funny cuz the model runs faster, but it also does more. So, it takes a bit longer\nsometimes. Depends on the task. It did not take longer at all when I was doing the actual changes to a real big\ncodebase, which again, like it's really impressive. This is an openweight thing that you can download and run on your\nown computer hypothetically that is capable of making real changes in an\nactual production codebase being used by hundreds of thousands of people on the web every day. This is not trivial [ __ ]\nanymore. This isn't oh it autocompletes or if you're in the file and tell it to make one change it sometimes can. This\nis real work in real code bases. Not what I'm demoing right now. This is a quick hacky demo app. But the thing I\nwas showing before, that's actual work. Cool. Miniax just finished. That took it\n7 minutes and 59 seconds. And here's what it made. Not bad. Testing. Generate. Has cool\nfake UI for this. I can go to the gallery view that it made with a bunch of fake images in it. Not bad. Still\nnowhere near as good at UI as any of the new Opus or especially the new GPT models. And now Gemini 3 is really good\nat UI as well. This is fine considering that this model is basically free to use because it's so cheap. Like that cost a\ncent. One cent. I wrote almost 400 lines of code for 1 cent. That's insane. The\nGLM won't show me how much it costs because it's free. I can do manual math after, but it's not far off. It\ngenerates more tokens and costs a little less. So they balance out about the same. You know what? I'll try. I'm going\nto do something stupid. So the thing I've been doing a lot with my own model usage is I use Opus for most things.\nI'll occasionally pull in 5.2 extra high to do really deep planning stuff. If it\nhas to like touch far away files with strange things in them, 5.2 is still a\ndecent bit better at planning. And 5.2 is still also quite a bit better at complex style and like spatial\nrelationship type things compared to Opus. Weird CSS [ __ ] I still use the GBD5 models for. So I've learned that\nfor most things I should just use Opus and then occasionally switch over to GPT5 if I know it's a more UI heavy task\nthat it will benefit for. So let's try that here. I'm going to use M2.1 for Miniax the same way I use Opus as like\nmy general model, general planning, general code, general things. And I'm going to use GLM 4.7 as my UI\nspecialist. Instead of waiting for it to potentially recover whatever the [ __ ] happened over here, I'm going to ask it\nto fix this code and make it less ugly. So, I'm going to give it access to the miniax m2.1 version and tell it to clean\nit up. I'm also going to do this inside of kilo code because for whatever reason, open codes deployment of 4.7 is\nhaving a lot of weird issues right now. Push to kilo to GLM 4.7. This app is\nugly. Make it beautiful. It should be dark mode relatively\nminimal and tasteful. Let's see how GLM does in kilo code at making this very\nugly thing less ugly. I bet it will do this faster than it will finish fixing\nthe gen in the other one. Yeah, this there's something up with the Zen deployment. It's just slow. Let me\nupdate the styling. Tool call right breaks off. Fun. This shit's not as easy\nas people think. Okay, it's cranking. It just made a new file that just says use client at the top without even having\nquotes. This is just rewriting the whole file. Hilarious. It's not even bothering with\nedits. It's just gunning out a new file. Tell in 4.7 how I really feel. Man, I'm\nexcited for these models to get way faster, too. That is going to be so nice. What's actually kind of cool is\nsimilar to how if you know your tech stacks and frameworks and tooling well enough, you can make good compromises as\nto what tech to use where, like you know when it makes sense to use convex versus a direct Postgress database versus\nsomething else entirely. You know when it makes sense to use React versus when you really should go do it in solid\nversus when you can't use a framework, it'll get in your way. That type of knowledge is starting to exist within\nmodels where you might know 4.7 is better at UI than Miniax 2.1, but Miniax\n2.1 is way better at planning and doing longunning tasks. And knowing which to use where to optimize for both\ncapability and cost is really cool. Like I could see a legitimate use case here\nbeing to use miniax as your go-to general use model and then go spin up\nOpus 45 with the front-end scale to do the the homepage and the front end UI, but not to have it do anything else.\nThat way you can spend money on like your marketing and your homepage and not\nspend that much with the actual implementation where you're generating way more tokens. You can limit where and\nhow you use the more expensive models or even the dumber models to specific boxes to maximize how much you're getting for\nthe money you spend. Or you can be lazy and just use Opus for pretty much everything like I've been doing. But the\nfact that you can get so close for so cheap is absurd. Let's see how it did.\nOkay, here is the GLM version editing the code that it got from Miniax. It is\nnice and minimal and looks a bit better. did keep the original UI a little more than I would have hoped, but it looks\nsolid. I do like the minimalism. I'm going to tell it to do one last pass.\nWrote quite a prompt for this. Get more creative with it. Don't be afraid to change the layout entirely. You have\nfull creative control. Use it to make something outstanding. For some reason, that was a cued message. Okay, it\nappears I just broke Kilo code entirely. [Music]\nKilo team, I want to love what you're building. We need to we need to chat.\nDid it fail again? Kilo,\nI'm getting annoyed. Okay, here's the original from GLM 4.7 in open code. It finally fixed the\nlayouts. What did it do wrong? It did a lot of random [ __ ] that wasn't actually fixing things. The div class\nrelative doesn't have any height or flex properties, so everything's compressed. Yeah, I don't buy the GLM's really good\nat UI thing as much as they showcased it here. This is bad. Embedding the\nbackground as a data URL hardcoded into the Tailwind class name is a choice.\nIt's an interesting way to do things. I haven't used skills and cloud code\nbefore. You just tell it to use the skill. Fun. Let's paste the same prompt.\nUse your front end design skill. Man, whoever makes the first model that\ndoesn't randomly run your [ __ ] dev server for no reason is going to get a lot of my money. Yeah, Opus just failed\nto write CSS that works. It also wrote a shitload of CSS, which like we're using Tailwind does not need to do thousands\nof lines of CSS. Oh, this might just be how it builds because of the other\nthings I'm using. Okay. Uh, a little bit different. This\nis when I gave Claude the design skill and it made a pretentious research\nwebsite that you would see somebody working on out of a Starbucks cafe.\nLike, it looks really cool, but the actual UX of navigating it is [ __ ] It uh feels like it tried too hard is the\nbest I can put it. But it's cool. It can do something like this. I might use this for a homepage. The skill really steered\nit here. You were curious how much this design cost and getting it to fix the imports and then rebuild. It's about a\ndollar. Yeah, four minutes of API time and a dollar to do that. So, you get the idea.\nYou can implement the whole feature for 10 to 40 cents with Miniax or you can make it look nice with Opus for 10x\nmore. The cost difference is crazy. Good day to be a cheap bastard. By the way,\nif you want a good UI to test talking to these models in, like asking it to write code for you or whatever else you want\nthem to do, T3 Chat is still by far the best UI you can use any of these models in. I don't care what other people say.\nThe rest suck for this type of general use. We also support great models like Gemini 3, both Pro and Flash. The best\nimage experience with Nano Banana right now by far, which is annoying. We shouldn't be so far ahead. And my favorite model to chat with, Kimmy K2.\nChances are if you're excited about models like this, it's cuz you're a bit of a cheap bastard. I get it. I am too.\nSo, I'm going to give you a deal. Normally, every month is $8 on T3 Chat, which is already a great deal, but just\nfor you, if you use the code cheap bastard at checkout, your first month will be just $1. Yes, $1 if you use code\ncheap bastard at checkout. Anyways, good model. That front-end design skill seems\nactually good. There's a few paragraphs that tell it how to design better.\nThis is 900 tokens of instructions. Focus on typography, color and theme,\nmotion, spatial composition, background and visual details. Never use generic AI\ngenerated aesthetics like overused font families. Enter robboto aerial system\nfonts, clichéed color schemes, particularly purple gradients on white backgrounds.\nThis isn't a skill. This is anthropic writing a markdown file to tell the\nmodel to stop doing all of the same [ __ ] This is like adding to the system prompt. Never say you're\nabsolutely right to fix the problem. It's actually really funny. Reading through enthropic skills is the best way\nto get an honest representation of how the models behave because this is where they hide all the things to make it stop\ndoing its [ __ ] That's so funny. That's so funny. Yeah. And there's a\ngood font. Is it built in anthropic skill? No. This is a skill that you can get from the\nanthropic GitHub. Cloudcode/pluginfrontendesign/skills. This is a skill that they published on\ntheir GitHub that they updated last month. It is just this markdown file. That's a skill. And to be clear, this\nrepo does not mean Cloud Code is open source. It is not. Cloud Code is actually the single source of the most\nDMCAs in the history of GitHub because they accidentally leaked their source map once. Cloud Code is very closed source. They just have some of the\ncommunity pieces here. It's probably the most stars on any GitHub repo that doesn't actually include the source that\nthe GitHub repo is named after. It's very silly and a lot of people seem to think cloud code is open source because\nthere's a repo named cloud code with 50k stars that gets changes every day, but this is not cloud code being open\nsource. The actual package, the thing you install close source and it's the only major CLI that is closed source\nstill. Just saying. And back to the open source one running GLM. It's still\ngoing. This one's been going for like 20 minutes and I'm tired of wasting Y's time. I feel like I've covered this\nadequately. The models are good. 4.7 is good at focused work and completing specific small tasks. Miniax M2.1 can go\nfor a long time on lots of things and generate really surprisingly good results. They're both open weight. They\nhave both meaningfully raised how much you can get done for really small amounts of money. So if you're on a\ntight budget and those $200 a month subscriptions make you feel sick, you should probably use T3 check is only eight bucks a month. But you should also\nconsider using these openw weight models with tools like open code in order to generate incredible results slightly\nslower and significantly cheaper. I love releases like these and I'm really pumped about the things that we're seeing come out of these openweight\nChinese labs. The results are unbelievable. I would never have believed this type of performance was possible at all even just six months\nago. The fact that this is a file you can download and run on your own machine to do things like this is so cool. This\nis a great way to end the year and I'm so excited to see what these labs cook next year. Until next time, peace nerds.\nAnd yes, it's still going.\n"

ERROR: type should be string, got "https://www.youtube.com/watch?v=3_71Nog8JZw\n\nReflecting on AI in 2025\n\n2025 has been a crazy year for AI. At the end of 2024, OpenAI introduced 01 which brought reasoning models to the\nentire world and right when 2025 started, Deepseek rose with Deepseek R1, an openw weight model that allowed for\nreasoning to be done for everyone. These changes shifted the whole industry and\nwe've seen crazy things happening since. So crazy that our friends over at Open\nRouter took the time to break it all down. They have unique insights on everything going on because they host\nand let you use every model. They do some rough top level analytics on all of the things that are running through open\nrouter which gives them a shitload of interesting data to use to help break down trends. From the weird use cases\nfor LMS and spoiler programming is actually number two. You'll be very surprised what number one is to the\nsplit across open and closed weight models which has been really cool to see. It's over 30% open weight models\noverall now, which is nuts. And the different releases that got us here are really cool. The absurd growth in token\nusage overall. It's been crazy. Seriously, there's so much cool [ __ ] in this paper. I can't wait to break it all\ndown for you. But as a person who spent a lot of money on tokens in 2025, we do have bills to pay. So, we're going to do\na quick break for today's sponsor. Stop me if you've heard this one before. I rolled my own off and I regret it. For\nT3 Chat, I built my own offplatform using open source packages, and it worked a bit for the most part. But as\nsoon as we wanted to do anything even vaguely complex, it fell apart. We lost multiple potential deals with businesses\nthat wanted to adopt T3 chat because we had no way for them to authenticate their business users. That has all been\nfixed because we made the move to work OS. We are very thankful we did. It has allowed for integrations that were\nimpossible for us to deal with before. We're actually really liking a lot of the other subproducts, too. Things like Vault are super cool for encrypting user\ndata. Things like API keys if users want to bring their own a API key to your service. Vault makes it trivial to do\nthat in a way that is safe and not exposing the data. The admin portal is the killer feature though. This makes it so easy for anybody at any company to\nset up your platform. If the IT team at Microsoft has to onboard for your product, good luck. Have fun with all\nthose calls if you're not using work OS. And if you are, you send them a link, they click on whatever off method they\nwant to use, and they're good to go. If you've never had to deal with ADP or SAML or OCTA before, I envy you. And now\nyou won't have to because you know about work OS. Speaking of miserable O stories, they're also working really hard in figuring out how to off your MCP\nservers, which is a a non-trivial task if you've watched my videos on the topic. I'm so thankful that works is\nfiguring all of this out because I was so tired of doing it myself. I highly recommend checking them out if you haven't at soy. Oo. Shout out to Open\nRouter A6 for putting the time into doing this type of research. They're one of the few companies that's well positioned to get this type of data\nsince Open Router lets you route to all the different models. They get insights for all the different use cases and\nthey've been collecting some very very interesting anonymized data throughout. With the release of the first widely adopted reasoning model 01 on December\n5th of 2024, the field shifted from single pass pattern generation to multi-step deliberation inference,\naccelerating deployment experimentation and new classes of apps. As the shift unfolded at a rapid pace, our empirical\nunderstanding of how these models have actually been used in practice has lagged behind. In this work, we leveraged the open router platform,\nwhich is an AI inference. You know what open router is if you're here. They've done over 100 trillion tokens of real world LLM inference across tasks,\ngeographies, and time. In our empirical study, we observe substantial adoption of openweight models, the outsized\npopularity of creative roleplay beyond just productivity tasks that many assume dominate LLMs, as well as coding\nassistance categories, plus the rise of agentic inference. They also have identified different cohorts and groups\nof users that stick around more. The glass slipper effect of people who\nshowed up early sticking around much longer than people who are showing up later. So those early people to LMS\nstick around. Findings underscore the way developers and end users engage with LM in the wild is complex and\nmultifaceted. We discuss implications for model builders, AI devs, infrastructure providers, and outline how a datadriven understanding of usage\ncan inform better design and deployment of LM systems. Should be fun. As always, the whole thing is linked in the\ndescription if you want to read it. We will be skimming and finding useful parts throughout. Great research regardless. You know, they're going back\nin time when they're talking about Sonnet 2.1 and anthropics improvements on RA. Good old RAG, good old days. So,\nthe different sections are open versus closed source models, agentic inference, category taxonomy, geography, effective\ncost versus usage dynamics, and retention patterns. Should be fun. All of this is metadata, so there's no user\ncontent data exposed at all. They used the Google tag classifier for content categorization. No direct access to user\nprompts or model outputs was available for the study. Instead, open router performs internal categorization on a random sample comprising of\napproximately 0.25% of all prompts and responses through a non-proprietary module, the Google tag classifier. While\nit represents only a fraction of total activity, the underlying data set remains substantial given the overall query volume presented by open router.\nGoogle tag classifier interfaces with Google cloud natural languages classify text content classification APIs. The\nAPI applies a hierarchal language agnostic tonomy to textual input returning one or more category paths\nlike computers, electronics, programming, arts, entertainment, role playing, games, etc. with corresponding\nconfidence scores in the range of 0 to one. The classifier operates directly on prompt data up to the first 1,000 characters. Classifier is deployed\nwithin open routers infra ensuring that classifications remain anonymous and are not linked to individual customers. Huge\ncategories of confidence scores below the default threshold of 0.5 are excluded from further analysis. So if it doesn't get scored highly enough in a\nthing, it just ignores it. Cool. So they have made these buckets for the categories and show which tags link to\neach. I don't know if this was public before, but this is really cool cuz I look at that data a lot. Programming, roleplay, translation, general Q&A,\nknowledge, productivity, writing, education, literature, creative writing, adult, and others. They also break out\nbetween open source and proprietary variants, origin uh variants, Chinese versus everywhere else because China's\nkilling it that hard, especially in the openweight world. Prompt versus completion tokens, which is interesting. Prompt tokens being the ones you send,\ncompletion tokens being the ones you receive. And we have geographical segmentation. Should be fun. They use\nbilling location. Let's take a look at the data. Here is one of the most fun\ncharts. This is token share by source type and origin. uh light blue is\nChinese openweight models, middle blue is anywhere else rest of world openweight models and then dark blue is\nclosed weight models. So even though open router is like main value prop is that it's easier to access every model\nwith that you would imagine that the main use case is open weight models. I know it is for us the majority of our use case for open router int3 chat is\nmodels that are hosted by Chinese labs where we don't want to use the Chinese infra. want to use American providers\nand other providers that have better data retention policies. Open router almost feels essential if you want to\nuse models like Kimmy K2 for example. So seeing them still having vast majorities\nof their traffic being closed weight is nuts. There was a brief moment around when Quen 3 and GBT OSS dropped where\nover 30% of the inference they were seeing was open weight but now it's back in the 20enter range or so. Very\ninteresting. Again, their data is going to be biased towards the open weight options because they are an easy way to\nuse them, test them, and integrate them. But even then, they're seeing 70% plus\nusing closedweight models. So, as cool as openweight models are, even a place like open router is still seeing them as\nthe minority of the usage for AI. Very interesting. I would have expected bigger for them in particular. It's also\ncrazy to see that at the start of 2025 there was very very little usage of\nChinese openweight models and that very quickly changed and by the end of the year the majority of openweight models\nare Chinese at least open weight usage. While proprietary models especially those from major North American providers still serve the majority of\ntokens openweight models have grown steadily reaching approximately 1/3 of usage by late 2025. This expansion is\nnot incidental. Usage spikes align with major open model releases like Deepseek V3 and Kimmy K2. It's hard to understate\nhow big of a deal V3 was. I know everybody was excited about Deep Seek R1, but V3 was the holy [ __ ] there's\nsomething special happening here moment. And Kimmy K2 is honestly similar in this regard. I like that they called that out\nhere. There's a reason it's now the default on T3 chat. If you go to an anonymous tab on T3 Chat, Kimmy K2 is\nthe default now because it is the most pleasant model to talk to that I've personally ever used. I genuinely really\nreally like Kimmy K2 and I highly recommend it if you're looking for a good model to talk to. That's why we made it our default. They're very very\ncompetitive. There's also other competitive launches like DCV3 further editions as well as the GPT OSS models\nand these have been adopted rapidly and sustained their gains. GBT OSS is also criminally underrated for various\nthings. I use it a lot more than y'all would probably guess. Since the GPT OSS model came out around the same time as\n04 did, if I recall, it was between 04 and GPT5. Because of that, it kind of\ngot missed, I feel like, and no one really embraced the things it was great at. Like we see here, it's the number\none model for legal questions in the legal category on Open Router. It's number one in technology. It's number\none in finance. It's number one in science. Hell, it's number two in programming. Everything else is quite a bit further behind on. I use it for data\nstuff all of the time because on the fast providers, if we scroll a little bit, you'll see some of them like base\n10 or where is my boys at Grock or Grock? Grock's pulling 511 tokens per\nsecond. That's 10 times what you get for something like GPT5. Insane. There are\nsome places that can pull it over a thousand. Cerebras is pulling 2,000 tokens per second with it. That is unbelievably fast. And the value of that\nspeed isn't just like, oh, it's really fast. It's the speed plus the price plus its ability to do things like follow the\nshape of an object well make it great for parsing absurd amounts of data. We use this for doing sentiment analysis on\nthe comments on my videos. We use this for analyzing the results from SnitchBench. We use this for a ton of\nthings. It's a great model. It also can run on my computer at reasonable speeds. If I open up LM Studio, since my MacBook\nhas a ton of memory on it, it can run these big models pretty fast. It's able\nto run the GBTOSS 20 bill version at like 90 or so TPS, but the 120 bill\nversion, I assumed it wouldn't be able to do [ __ ] with cuz that's a huge model. That is a 60 gig model. I couldn't run\nthat faster than like 18 TPS on my RTX5090, but on here on my M4 Max MacBook, it\nruns quite a bit faster. You have to wait for it to load into memory, but that's the key. Since it's 60 gigs and\nmy laptop has 128 gigs, it can load the whole thing into memory. No, this is not\nan M5. The M5 does not have a model with this much RAM. I'm on an M4 that I just\nbought because we don't have an M5 Pro yet. And this is why you cannot get this speed on an M5 right now as write an\nessay about why Rust is better than C++. And that is faster than reading speed.\nThat's flying. This is going this fast not just because my processor is fast. It's mostly\nbecause my RAM is insane and I have memory that works with my GPU. Doesn't\nmatter how much RAM you put on your desktop if you're running a 5090 because 5090 has a certain amount of VRAM that\nyou can't really expand. So, if you can't fit the model in your VRAM, it's going to run like [ __ ] I can fit it in\nmy VRAM because I'm on a MacBook, which means I have unified memory. So, the RAM is the same as the VRAM. So, I'm able to\npull a faster speed here than I can on my giant beefy gaming desktop with a 5090 on it. Yeah, that was 71 tokens per\nsecond. That's faster than GPT5 over API. It's awesome. That's so cool. I had\nto use half my RAM for it, but now that I've ejected this, it should be totally fine after it's done beachballing. So,\nit used all my memory. And that's all local. That's entirely offline. So yeah, GBDOSS criminally\nunderrated, but if you know, you know. And it definitely caused a spike in usage of openweight models. As we see\nhere, Kimmy K2, Quen 3, and GPTOSS all caused the spike that led to openweight\nmodels being over 30% of the usage. And if we look at the overall token usage\ninstead of just as a percent, you'll see Open Router had a very good year. Good for those guys. At the end of 2024, the\nweekly share of Chinese openweight models was only 1.2%. Now it's nearly 30% of total usage among\nall models in some weeks. Over the 1-year window, they average approximately 13% of weekly token volume. That being Chinese models are at\n13%. With strong growth concentrated the second half of 2025. This is very interesting. It turns out that Deep Seek\nwasn't the point where everyone started using these Chinese models. It was around halfway through the year when we\ngot Kimmy and Mini Max and Quen got good and all these other drops happened for\nnon-Chinese. Open source models averaged 13.7% and proprietary models maintained their\n70% share roughly. This pattern has materially reshaped the open- source segment and progressed global\ncompetition across the LM landscape. Absolutely. It's been a crazy year for openweight models. The fact that the\ndefault model on T3 chat is an openweight model and that like half of my personal inference is going through openweight models is just kind of crazy.\nKey open- source players. This is by number of tokens. Even though I just said Deepseek wasn't the main player\nthat got this bump, they still are the single biggest one pulling 14.37\ntrillion tokens over open router from November 2024 to November 2025. That is\ncrazy. Quen's at 5.6, six, which is also nuts, especially as I personally don't\nfind those models particularly great. Meta and Mistl did okay here as well. Open AAI, especially when you say the\nfact that they put out their first open weight model halfway through the year. Them being this high is impressive. Miniax again coming halfway through the\nyear with their first relevant model being that high is crazy. Zai doing very well as well. They're the GLM models.\nAnd then Moonshot, my underrated goats at 0.92. I'm sad they didn't hit the one trill. We're working our way towards\nthat with Kimmy as the default now. Hope you guys can climb this real fast. Also, Miniax was free for a bit too. That's\nanother thing to account for with Open Router. Models that are provided for free tend to get boosted really, really\nhard because a lot of open routers like heavy users are cost-sensitive. That's also a big part of why the role playing\ncommunity seems to really like DeepSeek. It's not so much that they like the model. It's mostly that they provide them for free and that target and that\ndemographic is very cost-sensitive. that they love using free models. This is also why XAI always brags that they're\nso popular on Open Router. It's cuz they provided Grock for free for so long when they put out the fast version.\nObviously, when you're using something for free, there is some value exchange there. They're almost certainly tracking all of the things you send and using\nthat for post training. So, know that when you use the free models, there's no way they're just giving that out for\nfree out of the kindness of their heart. And now we have the decline of Deep Seek's dominance. Oh boy, this will be a\nfun chart. Yeah, as we see here, Deep Seek was the king of these openweight models for a while, quickly eating into\nLlama's previous like strongholds, but then as the summer hit and we got more\nof these Chinese openweight models like Moonshot with K2 or Mini Max or ZI,\nyou'll see their share just getting eaten up fast. But then they dropped V3.1 of Deepseek and won a huge chunk of\nit back and they got to this point here where it's actually is like the peak of\nYeah, no, it's not that the peak of Deep Seek's ownership was definitely here. They lost a whole bunch. They dropped\nV3.1 in September, climbed back for a bit, and now it's dwindling. Right now it's probably up again cuz 3.2 is nuts,\nbut they also sucked at hosting that for a while, so who knows? Their near monopoly on open weights was shattered\nby the summer inflection. The market has since become both broader and deeper with usage diversifying significantly.\nNew entrance like Quen's models, Miniax's M2, Moonshot's Kimmy K2, and OpenAI's GPTOSS. I don't know why they\nare just skipping Zai and GLM here because those models did really well, too. They actually performed higher than\nMoonshot there. Wonder why they skipped that. Interesting. Regardless, they all\ndid very, very well. By late 25, the competitive balance had shifted from near monopoly to a pluralistic mix. No\nsingle model exceeds 25% of open source tokens and the token share is now distributed more evenly across five to\nseven models. The practical implication is that users are finding value in a wider array of options rather than\ndefaulting to one best choice. Yep. And this isn't just like, oh, I prefer this model or I prefer this model. This is\nalso this one is cheaper and can handle more tokens. This one can do reasoning with images. This one is really good at\ntool calling, but this one isn't. This one handles reasoning well and this one doesn't. This one writes well. There's a\nlot of reasons to prefer different models. Like I obviously like talking to Kimmy K2, but I like doing data analysis\nwith GPOSS. There's good use cases and bad use cases for all of these. You just got to play with them to know. Overall,\nthe open source model ecosystem is now highly dynamic. Some key insights include the top tier diversity where now\nthere's tons of different families of openweight models. The rapid scaling of new entrance, people show up and\nsuddenly blow up. And the iterative advancement, the longevity of DC's presence at the top underscores the\ncontinuous improvement is critical. Their successive releases like chat v3, R1, and now v3.1 and 3.2 are keeping it\ncompetitive even as challenggers emerge. OSS models that stagnate in development tend to lose shares to those with\nfrequent updates at the frontier or domain specific finetunes. The open source arena resembles a competitive ecosystem where innovation cycles are\nrapid and leadership is not guaranteed. It feels like the early days of closedweight models in the openweight\nworld and it's awesome to see. The model size versus market fit medium is the new small. Very interesting. I definitely\nfelt this myself. For example, with GBToss that is a medium model because I\ncan run it on my laptop. Large openweight models are things that you cannot run on consumer hardware that\nrequire specialized hardware. And small models are things that you can run on cheaper hardware like my 5090 instead of\nmy MacBook. I know that sounds silly, but the amount of RAM used is an important key to the differences here.\nAnd it is cool to see how many small openweight models got useful. It's still crazy that there are so many large\nopenweight models that are good too because there's like the only people who benefit from a large openweight model\nare the hosts that can use it because they have enough H100s around. So seeing large openweight models do much better\nat the start of the year makes sense. Something like Deepseek R1 is a very large model. You're not running that\nwithout a distillation of it locally at all. But now we're getting more stuff like 120 bill models that are reasonable\nto run, but also seeing some growth in the small models too. That said, we're\ngetting some crazy big open weight models. I know Kimmy, for example, is like one trillion parameters. Good luck running that locally. Their definitions,\noh, interesting. By their definition, GPT OSS is large at 120 bill. I guess\nthat's fair, but just interesting to see. Small is under 15 bill per, medium is 15 to 70, and large is 70 plus. I\nwould have cut this differently. I would have done small as anything under 30 and large is anything o over 150. But I get\nit. Especially now with RAM prices, it makes sense to knock these down. The data on developer and user behavior\ntells us a nuance story though. The figure shows that while the number of models across all categories has grown,\nthe usage has shifted notably. Small models are losing favor while medium and large ones are starting to capture more\nvalue. This is very interesting. The size of the model helps determine a lot of different things. Obviously, the\namount of knowledge baked into it, the size of it affects that a lot, but it also affects the price a ton, too. These\nlarge models are much more expensive than the small ones because you can run them quicker and on cheaper hardware.\nSo, the interest in small openweight models is maintaining and medium is growing meaningfully because the medium\nones are really good price to performance value. The small market is declining overall. Things like Gemma 312\nbill, which was released in August, saw rapid adoption, but now it's competing in a crowded field where users are\ncontinually seeking the best alternative. We're still waiting on Gemma 4. I think it'll happen at some point this week. Google's been teasing\nit on Twitter all week, but we'll see. The medium market, finding the model market fit. The medium model category\ntells a clear story of market creation. The segment itself was negligible until Quen 2.5 coder 32 bill dropped in\nNovember of last year, which effectively established the whole category. The segment then matured into a competitive\necosystem with the arrival of other contenders like Mestral Small in January and GPT OSS20 bill in August. Again, I\nwould put OSS20 bill as a small model, but I understand why they're making this distinction. The large model segment has\nbeen very interesting with Quen 3 235 bill or ZI's models as well as the GP\nOSS120 bill all capturing meaningful and sustained usage. The plism suggests that\nusers are actively benchmarking across multiple open large models rather than converging on a single standard. Yep,\nabsolutely. I play with all of them and they all have good use cases. So, I just cropped this because I want you guys to\nlook and guess without seeing what use case do you think is in the yellow. This\nuse case is between 40 and 80% of the usage of openweight models on open\nrouter. You've got to be thinking stuff like code or maybe chat or knowledge,\nbut your guess would be code, right? That's what mine would have been.\nReality is always more interesting cuz that is the roleplaying category. Yeah,\nthe weebs cannot be stopped. Roleplaying is around 52% of the use of openweight\nmodels on open router. There have been a lot of theories as to why this is. The one I've grown to believe the most is\nthat this category is price sensitive. And the free models like Deepseek V3.2\nfree are very well regarded in those communities and they are willing to let them have all of their data so they can\nget good enough role playinging. Interesting. The figure above highlights that more than half of all openweight\nmodels have their usage falling under roleplay while programming is the second largest category. This indicates that\nusers turn to open models primarily for creative interactive dialogues like storytelling, character roleplay, and\ngaming scenarios and for coding related tasks as well. The dominance of role-playing hovering at more than 50%\nof all openweight tokens underscores a use case where open models have an edge. They can be utilized for creativity and\nare often less constrained by content filters which make them attractive for fantasy or entertainment applications. I\nwish they would show what percentage of those are using the free versions of the models. That's still my hypothesis, but\nthe fact that they are less likely to be constrained by the hosts makes a lot of sense, too. You can only bake so much\nsecurity into the model itself. Most labs put a layer in front of the model that will detect if the query is bad and\nblock it rather than just baking the safety into the model itself because you can't adjust that over time. But that\nalso means the openweight models if they have weird behaviors or they don't tune out properly, that's there forever. The\nweights are public. Role-play tasks require flexible responses, context retention, and emotional nuance,\nattributes that open models can deliver effectively without being heavily restricted by commercial safety or\nmoderation layers. This makes them particularly appealing for communities experimenting with character-driven experiences, fanfictions, interactive\ngames, and simulation environments. Makes sense. We look at the breakdown just within Chinese openweight models.\nIt's shifted quite a bit. There was a point where programming had overtaken roleplay, but now it's a little bit\ncloser. Roleplay has knocked its way down to around 33% of Chinese tokens, which is interesting. And if I recall, a\nsignificant portion of that is on the DeepS free tier. The shift suggests that models like Quen and Deepseek are\nincreasingly used for codegen and infrelated workloads. While highvol enterprise users may influence specific\nsegments, the overall trend points to Chinese open source models competing directly in technical and productivity\ndomains. And overall over the year, technology and programming were 39% and\nrole-play was around 33. So looks like programming is winning in the space now. But what if we break down our usage of\nmodels in programming by open- source versus closed versus not Chinese? Gets\nvery interesting again with that surge in the summer with Miniax M2, GLM 4.5,\nKimmy and all of that. We saw a huge bump in the percentage of tokens going through open router being used for code\ntasks. Massive bump. It flattened a little bit. Bumped again in October. I\ncurious what openweight model we was put out from somewhere else. That bumped there might have been a mystral thing.\nWe'll see in a sec. And now the non-Chinese openweight models are barely being used and closed models are still\nwinning by far. Let's see how they break this down. Chinese open source models in blue delivered the majority of open\nsource coding help driven by early successes like Quen 3 coder. By quarter 4, Western OSS models such as Llama 2\ncode and GBT OSS had surged but decreased in overall share in recent weeks. The oscillation suggests a very\ncompetitive environment. Yep. Developers are open to whatever open source model currently provides the best coding\nsupport. As a limitation, the figure doesn't show absolute volumes. Open source coding usage grew overall. Also, a shrinking blue band doesn't mean\nChinese open source lost users, only relative share. Good to know. Role- playing breakdown for open-source\nmodels. Interesting. Non-Chinese openweight models are 43% of\nthe role playinging use case. Very interesting. The most recent numbers, Deepseek V3 is still the most popular.\nR1T2 and V3.2 are close though. Then it's Flash. Then it's Deepseek 3.1.\nThen it's Gemini. Then it's Mistral Nemo. What?\nWhat the [ __ ] That does not line up at all with what I'm reading here. This\nlooked like in the roleplay category, non-Chinese open source was doing much better, but that does not line up with\nthis at all. Like others is 42.6%.\nBut it's not. Yeah. Interesting. Very interesting. And you can see here when Grock 4.1 Fast\nfirst came out and was free, it just slaughtered. And as soon as it stopped being free, it vanished.\nAgain, proving my theory that this community wants the cheap [ __ ] There you go. 4.1 Fast was number one when it\nwas free by a lot. And as soon as it stopped being free at mid December, it\nimmediately vanished almost entirely from the chart. Yeah, I'm surprised they didn't break this down a bit more.\nI want more info on the openweight non-Chinese models that are being used here because they're just not sharing\nenough about it. Very interesting. I wish we had more info. Enough about open weight in China. Let's talk about the\nrise of a gentic inference. This is the amount of tokens that are reasoning\ntokens. And we recently crossed the point where over half of the tokens that Open Router is watching get generated\nare reasoning tokens. I was saying for a bit that the average token isn't read by a human. This really emphasizes it.\nUnless you're reading every point of reasoning in every trace you do, which you can't with a lot of these models, they don't even share the reasoning\ndata. You cannot possibly be reading all the tokens being generated. Those tokens for reasoning are making the actual\nanswer more likely to be better. And the tokens that are being generated are often not even being shown to a human in the first place. They're being used to\ncommand tools to go do things. So yeah, that's one of the most interesting parts of this chart is it's showing one of the\nmany angles at which text generation models and LLMs are being used to generate things that aren't for a human\nto read. As shown in the figure above, the share of total tokens routed through reasoning optimized models climbed sharply in 2025. What was effectively a\nnegligible slice of usage in the early Q1 days is now over 50%. The shift reflects both sides of the market. Oh,\nthis isn't what I thought it was. This is share of all tokens routed through reasoning models versus routed through\nnon-reasoning models. So this isn't what percentage of output tokens were or weren't reasoning. That would probably be much more brutal. This is what\npercentage of requests, not even requests, of the tokens sent to open router as input. What percent of them\nwere sent to a reasoning model versus sent to a non-reasoning model? That makes more sense. Top used models with\nreasoning. Grot code fast one is number one because again they offered it for free. It says XAI's APIs suck to use and\nsign up for. Most people opted to use it through open router. That makes sense that it's so high up, but again, it's\nthere because it's free. XA's Graco Fest one now drives the largest share of reasoning traffic excluding free launch\naccess. Interesting. Apparently, even if they exclude the free launch access, it's still number one. That feels kind\nof [ __ ] This is a notable change from only a few weeks ago where 2.5 Pro led the category and R1 and Quen 3 were\nalso in the top tier. This is very interesting that Gemini 2.5 Pro is so far ahead in Open Router. Again, my\nconspiracy theory there is that using Google's models through Google's APIs is the stupidest, most miserable [ __ ]\nthing you can do. I say as somebody who does it a lot, it is hell. It is awful.\nI do not recommend anybody deal with Google's SDKs or APIs directly when you can use something like Open Router. I\nwould gladly pay them the [ __ ] 3 to 5% fee, whatever it is, to just route my traffic through Open Router and not deal\nwith Google's [ __ ] Both OpenAI and Anthropic have competent APIs and competent CLI to use the models through\nif you're doing code and competent integrations and editors like cursor, windsurf, etc. So the need to use\nsomething like open router to hit them is way way lower. With Google models, there is basically no competent way to\nhit them other than open router. So we're kind of stuck here. But these are the biases we have to account for in this report because this report is\ngenerated by a source that inherently sees different patterns of traffic. As Clay said in chat, open router just\nworks as opposed to Google Cloud's [ __ ] Totally agree. Uh somebody said to check this graph out. O yeah to\ngo back to like Grock getting traffic cuz it's free. Here you go. See all that purple traffic that just vanished all of\na sudden? This is when they made that a free model. And as soon as it stopped being free, it disappeared from the\nchart. You get the idea. A lot of this data is inherently biased by the nature of the users and the nature of the\nreasons they're using models through open router. It's still really good info, but it's biased. It's also again\nto really just highlight how biased this is. There is no world in which anybody actually thinks that GPT5 is less\npopular than GPTOSS. Obviously, like duh. Why the hell is GBT\nOSS120B higher up on here than GPT5 is? It's obvious if you've been listening. It's\nbecause GBT5 is able to just be hit through OpenAI's API totally fine. But GBToss isn't even hosted by OpenAI. You\ncannot use the GBOSs models by hitting the OpenAI infra. You have to use them\nelsewhere, which means you have to deal with other providers or use Open Router and it will route you to the best\nprovider, which is a great great option. So GBToss is best used on Open Router\nthe same way Gemini is. So yeah, all five of these models suck to use outside\nof Open Router and are great to use with Open Router. So it makes a lot of sense that they are the top five. That does\nnot mean any of these models are doing better than whatever the Hellanthropic is doing or what OpenAI is doing with\ntheir closedweight models. It just means that on this chart with Open Router,\nOpen Router gets a lot more traffic for models that suck to use outside of Open Router, which makes sense. It's kind of\nfunny to say, but Open Router's edge is that XAI and Google really suck at making good developer experiences, and\nthey can provide a better one by routing the traffic through a layer that doesn't suck. Turns out being a rapper is a good\nbusiness decision. And now we have the rising adoption of tool calling. This is interesting. I would have expected\nhigher but again the use cases are interesting roleplaying for example you're not going\nto be doing a whole lot of tool calling in with programming you're probably doing a lot more but only 15% or so of\nthe traffic going through open router has a tool getting invoked very interesting the noticeable spike in May\nin the figure above was largely attributable to one sizable account whose activity briefly lifted overall\nvolumes interesting I I'm trying to remember who it was somebody who's using open router as their default provider and then moved off it and it caused a\nspike. It was one of the open- source code tools. Can't remember which it was. Regardless, you get the idea. Top 10\nmost used models with tool call finished reasoning. So, this is models where they are generating responses and ending\nbecause there's a tool call they're waiting for, which is how tool calls work. So, this is highest percentage tool calls. Cloud force on it is still\nthe king of tool calls as we could have expected. 3.5 and 3.7 were also great at\nit. You can see how quickly when a new model drops from Enthropic, their traffic for the old one collapses. Like\nthis pink here, this was 3.7. The gray is 40 mini. So I guess somehow in May,\n40 mini traffic grew a ton. We just barely got sonnet 4 if I recall. Yeah.\nSo sonnet 4 dropped near the end of the month and most of the Sonnet 37 traffic\nhad already started dying. In June, it collapsed even more. And by July, Sonnet 37's barely even being touched. And\nClaude 4 has crushed. And then slowly as new models like 2.5 Flash dropped and\ngot good enough to actually be able to do tool calls, it started eating into the margins of other things. And then 45\ndropped and crushed everything. That all lines up as expected. But also notice how little 2.5 Pro got in terms of the\nownership of this chart for tool calling. Part of this is because Gemini 2.5 Pro is really expensive. Part of\nit's because it's not the best at tool calls. And part of it's because if you're willing to pay those prices anyways, you might as well just use Claude. I I've never been a big 2.5 Pro\nfan. All it has done is cost us a ton of money and a lot of support tickets. I'm thankful to watch it slowly die. Also\nnotice that Gemini 3 Pro is not coming up here at all. It is late edition, so I\nwouldn't expect it to have snuck in particularly deeply, but I don't think it ever will because I think it got trumped really fast. Also, GLM is\nkilling it in this, too. The GLM models are really good at tool calling, and it's the only open weight model really holding some ground in here. That's cool\nto see. I somehow missed that initially. Good for ZI. And now we have the anatomy of the prompts. The prompt tokens grew a\nton. So, when you send a request, the average number of tokens in a request is four times higher than it was at the\nbeginning of the year. That's nuts. I did not expect it to grow that much. completioner tokens going up makes more\nsense though because again reasoning models generate way more tokens way more tokens. So this being much higher makes\nsense but only 3x when we got a 4x for the input is interesting and programming is the main driver behind prompt token\ngrowth. So the number of tokens being submitted for code has gone up a ton over time which is why the number of\ntokens per request has gone up so much. That makes a ton of sense. Longer sequences and more complex interactions. The average sequence length has grown\nover time. So this is number of tokens per generation. So prompt and completion together and programming is going up way\nfaster overall. It's meaningfully bigger than everything else. So for everybody asking why we don't do coding stuff in\nT3 chat, why we're not like letting you link your codebase and whatever, it's cuz we would be increasing our cost by 4\nto 6x and we charge $8 a month. Shut the [ __ ] up. Unless you are going to build something or pay us way more. It just\ndoesn't make sense for what we're trying to do, which is a really good chat experience for a reasonable price. If we let you bring your code base, we're just\ngoing to increase our cost by four to 6x. Go use cursor. It's a good tool. Agentic inference is the new default.\nTogether, these trends, rising reasoning, shares, expanded tool use, longer sequences, and programmers outsiz\ncomplexity suggests that the center of gravity LM usage has shifted. The median LM request is no longer a simple\nquestion or an isolated instruction. Instead, it's part of a structured agent-like loop, invoking external tools, reasoning over state, and\npersisting across longer contexts. For model providers, this raises the bar for default capabilities. Latency, tool\nhandling, conduct support, and robustness to malformed or adversarial tool chains is increasingly critical for\ninfo operators. Inference platforms must now manage not just stateless requests, but longunning conversations, execution\ntraces, and permission sensitive tool integrations. Soon enough, if not already, Agentic Inference will be\ntaking over the majority of the inference. Yes. And it's crazy how many of these like GPU rental services don't\nunderstand this stuff at all and are forcing companies like Moonshot to go put out benchmarks showing which hosts\ndo and don't follow the tool call protocols properly. How are people using\nLMS? This will be fun. Again, we saw role-playing was a big chunk, but we've seen programming in particular,\nprogramming through open router growing massively. This isn't just LLM programming usage\ngrowing, although that is a huge portion. I think a significant chunk of this is that the openweight models that are best used through open router have\nbecome much much better, especially around halfway through the year, which is why this giant spike started around\nhalfway through the year. I'm curious if they agree with my analysis here. Openweight models got good at code and\nopen router is the best way to use openweight models. So they saw unique growth here but they might try to generalize this across the industry\nwhich I probably won't agree with. Now of course LLMs become embedded in developer workflows their role as\nprogramming tools is being normalized. The evolution is implications for model\ndevelopment including increased emphasis on codecentric training data improved reasoning depth for multi-step programming tasks and tighter feedback\nloops between models and integrated development environments. They are not acknowledging the fact that this growth\nis almost certainly unique to them. Other places in the industry do not see code LM usage go from 11% to 50%. They\ndid because the models that they are best for got way better at code this year. Very specific to them. That said,\nanthropic models are at 60% of the share for programming based spend on open router. So maybe I can't make this\nstatement so definitively, but at the same time, there's a lot of providers in here that are best used through them.\nAnd I expect Google to have that big of a stronghold there, too. It was very interesting to see how quickly these\nother options have grown into gotten actual footing in the space, but it looks like it's a small enough\npercentage that I should shut my mouth on the open weight thing. I might be wrong on that. Very interesting. And\nMiniax is taking a good slice, too. very very interesting to see if we look\nwithin categories role play only 15% of the roleplay is adult according to them\nbut over half is games I know stuff like Dungeons and Dragons love having AI\ngenerated role-playing text so that makes sense programming is interesting\nhow that's split scripting languages versus development tools versus other I\ndon't think these splits are useful so I am going to skip most of this section author Level insights by category.\nDifferent model authors are utilized in different usage patterns. Figure below shows distribution of content categories\nfor the major model families. So anthropic obviously mostly programming.\nA little bit of tech, not a whole lot of roleplay, lots of code though. Google's\nmost popular categories a little more varied. A lot more roleplaying, a lot less code, decent bit of tech, good bit\nof science, not much analysis. I would have thought that people would use them more for analyzing documents and things and that would be clearer here by\nshowing like legal or finance or something. XAI was code until they made\nit free and then it became roleplay in technology I guess and now it's not free anymore. It's going to go back to just\ncode. I know how this works. OpenAI has had a very weird distribution where it was mostly science until June when they\ngot better at coding and then it became mostly code with technology still being high. I'm curious what technology even\nmeans in these cases. Let's go back up up up and see. They don't even have technology broken out here. Oh, they do.\nPersonal assistance, business and productivity software, web design and development. That's code [ __ ] So, I\nguess some code is put in there, too. Weird category splits. And then Deepseek\nis the roleplay machine. They should rename it to Deepseek and Dragons at this point. God damn. And then Quen is\nmostly code with a weird variety of other random [ __ ] Fascinating. And then we have usage across regions. North\nAmerica versus Asia versus Europe versus other mostly North America still, but Asia's had a bigger and bigger chunk at\nleast on open router. Language distribution over 80% is still English. Checks out. And then the glass slipper\nphenomena. This is what they're talking about before where when somebody tries a new model really early, they are much more likely to stick with it than if\nthey come to it later. So, people who tried Claude 4 in May, right when it\ndropped, much more likely to still be using it than people who tried it the month after. People who had tried Gemini 25 Pro right when it dropped in June,\nare comically more likely to still be using it than people who tried it later. Very interesting. The collection of retention charts captures the dynamics\nof the LLM user market across leading models. At first glance, the data is dominated by high churn in rapid cohort\ndecay. Yet, beneath the volatility lies a subtler and more consequential signal. A small set of early user cohorts\nexhibit durable retention over time. We term these foundational cohorts. They're not merely early adopters. Oh god, this\nis so LLM written. That's so that pattern just screams LLM. These cohorts\nare not merely early adopters. They represent users who workloads have achieved a deep and persistent workload\nmodel fit. We can't just product market fit everything, guys. Once established,\nthis fit creates both economic and cognitive inertia that resists substitution even as newer models\nemerge. We introduced the Cinderella glass slipper effect as a framework to describe the phenomena. The hypothesis\nposits that in rapidly evolving AI ecosystems, there exists a latent distribution of high-v value workloads\nthat remain unsolved across successive model generations. Each new frontier model is effectively tried on against\nthese open problems. And when a new release model happens to match a previously unmet technical and economic\nconstraint, it achieves the precise fit, the metaphorical glass slipper. Okay, so if you have a problem that none of the\nmodels are solving and you try a new model right when it drops and it solves it, you just stick with it. That's the\nhypothesis. Interesting. Can kind of see it. Hard to know for sure. It's also small enough\nnumbers it's hard to care. Like llama did not have any real retention at all in comparison to quad 4 sonnet. But even\nthen, like a lot of this could be measured by how long did it take for them to ship something new. Like Cloud\nForce Sonnet, they shipped 4.1 then 4.5 pretty quickly after. Gemini 25 Pro,\nthey did not ship anything until over 6 months later. So obviously that's going to stay flatter. I don't buy this part.\nI think this is a bit of a reach. Apparently they're using 40 Mini as the golden example here. God, I hate that 16\nmonths later people are still using 40 Mini. That was the model we shipped T3 chat with in January. Also, this is\n[ __ ] 20 Flash never had any fit. 20 Flash is one of my favorite models. I use it for a ton of [ __ ] I really just\ndon't buy this part. And then the boomerang effect. The amount of reaches they're doing to explain this. No, I\ndon't believe this at all. Let's look at cost versus usage. This will be much more interesting. Cost is X-axis. Total\ntokens is Y-axis. Interesting to see translation focus on cheaper. Legal also focusing on cheaper overall. Tech and\nmarketing leaning more expensive and programming leaning right in the middle. Very interesting. The scatter plot above\nreveals a distinct segmentation of AI use cases, mapping them based on their aggregated usage volume against their\nunit cost. Both are logarithmic axes. Really interesting to see roleplay leaning towards more expensive models\nthan legal does. Your lawyer's cheaper than your Dungeons and Dragons DM guys. You heard it here first. An effective\ncost versus usage of models. This is fun, too. You can see some models that are expensive\ngetting much less usage and some that are cheap getting much more. But overall, it's pretty evened out. There's\na very slight trend downwards where usage goes down as cost goes up. It's a weak overall correlation. Interesting.\nCool to know that the cost is not a big driver of usage overall. Quality and capabilities often trump cost. That\nchecks out. Being cheap isn't enough. A model must also be differentiable and sufficiently capable. Yep. Cool. What a\npaper. Here is the end of discussion. The ecosystem is now multimodel. The\nusage diversity beyond productivity. We're seeing lots of other use cases. Agentic use cases are going up a ton.\nThe geographical split is widening, but still mostly North America based. Cost\nversus usage. Cost does not really significantly drive usage in ways that\nyou would expect. And the glass slipper effect, which I don't agree with.\nThis was really interesting research. I am pumped that they published this and shared all of this data.\nThis was fun. Thank you to the Open Router team as well as A16 for helping out with this. I feel like I learned a\nlot and have a much better idea of the whole ecosystem and the greater picture of what's going on. Curious how y'all feel though. Are readthroughs like this\nuseful? Are you interested in the greater market stuff or was this just a really long boring video? Maybe you didn't even get to this part. Let me\nknow what y'all think and until next time. Peace.\n"

ERROR: type should be string, got "https://www.youtube.com/watch?v=rddX4GEeyvE\n\nGemini Flash 3 is my new favorite model (yes really)\n\nI know, I know you guys are tired of the best new model videos. Go let the guy with the counter do his thing. But this\none's really legit. I had my skepticism around Gemini 3 Pro, and that turned out to be pretty well placed because using\nthe model in a day-to-day basis has not been great. That said, there is a Google model I've been using every day for\nalmost a year now. And the model doesn't have Pro at the end of the name. It has Flash at the end of the name. 2.5 Flash\nhas been my daily driver for so many tasks for so long. I really like the model. It's such a good balance of speed\nand capability. And if you're willing to work around its weirdness, which mind you there is a lot of, 2.5 Flash has\nbeen really, really good. Which is why I was super excited for Flash 3 to be a meaningful bump on 2.5 Flash's\nperformance, which it was, and then some, because as you might see here, Gemini 3 Flash isn't close to where 2.5\nFlash was on the chart. Three Flash is ahead of Opus 4.5 on the chart. Yeah,\nthis is the official artificial analysis intelligence index, which has its problems. Believe me, I've talked about\nthat enough times, but it is a good enough general idea of like where does this model land? And it turns out three\nflash lands somewhere really, really good. It's kind of insane, especially because, and I should have probably\nmentioned this earlier, I was lucky enough to have early access. The early access with Google is always interesting. They've been getting better\nat it. They're not paying me for any of this. They just let me try it. It also gave me no guidance whatsoever. So, when\nI started running it on my own benchmarks, I felt like I was going insane. Because on Skatebench, not only\nis three Flash outperforming 2.5 Flash by an absurd margin, it's outperforming almost every single model other than 03\nPro, GPT5, and 5 High. Yeah, its understanding of spatial reasoning and I\nguess skateboarding terminology is best in class. It's fast. It's great at retrieving things from giant contexts.\nIt can parse images. It can parse videos. It can deal with audio. It's a good model, but what does this mean for\nday-to-day usage? What is it like to actually use a flash model like this? How much does it cost? What are the use\ncases? Why am I so excited about a model that is this small and according to everybody else, not that great? All\ngreat questions that I can't wait to tell you after a quick break from today's sponsor. Do you use GitHub? More importantly, do you use GitHub actions?\nIf you do, you're almost certainly wasting a ton of time and money. How much time do you spend a week waiting\nfor your GitHub actions to run? If the answer is more than 5 minutes, well, congrats. You're shipping real code. If it's not, fine. But if it is, you should\nreally pay attention because today's sponsor is here to save you a ton of time and money. Blacksmith has figured\nout CI. They get it better than anyone I've talked to, and they can make GitHub actions an actually usable platform.\nSetting it up couldn't be easier. You change one line in your GitHub action from Ubuntu Latest to Blacksmith, whatever instance you want to use, and\nnow your code is going to be built and executed comically faster. They don't hide how they do this either. They're\nshipping way faster hardware. It turns out that gaming CPUs are actually better than server CPUs for things that are\nreally focused on single thread performance. Server CPUs have 50 cores that are running as fast as a gaming CPU\nfrom 5 years ago. But if you use a gaming processor, you can compile your TypeScript and Rust code bases way\nfaster. Speaking of faster, the cache is one of the biggest benefits to using Blacksmith. They colllocate the data\nfrom your cache and artifacts with the actual box running. So, it's on the same drive on the same system, which means it\nspins up immediately, up to four times faster. All of this combines to 40x faster Docker builds. One of the most\nunderrated things these guys do is their observability stack. I cannot tell you how many times I've just been confused\nabout why an error is happening in GitHub actions. And without a way to really search through what's going on, it's impossible to navigate. Having good\nobservability makes your actions less scary and more useful. And they've really figured this out. They can give\nyou things like the failure rate for a given part of one of your runs or what tests are being flaky and let you\nfilter. It's so good. It's so good. As soon as I showed this to my team, they realized how useful Blacksmith was. And\nwe're working on making the move right now. If you're not using Blacksmith, you're wasting time. Fix that now at soyv.link/blacksmith.\nSo, let's start with the official post from Logan in particular. Building with Gemini 3 flash frontier intelligence\nthat scales with you. God, these titles are so cringe. But the actual value of the model is great. So, we'll look over\nall of that because I am way more excited for this release than I was for 3 Pro. 3 Pro was better than 2.5 Pro,\nbut still not like a practical day-to-day model. Flash for like data analysis stuff is incredible, so I'm\nmuch more excited for here. 3 Flash offers powerful performance at less than a quarter of the cost of 3 Pro along\nwith higher rate limits. The new 3 Flash model surpasses 2.5 Pro across many benchmarks while delivering faster\nspeeds. It also features our most advanced visual and spatial reasoning and now offers code execution to zoom,\ncount, and edit visual inputs. Again, it turns out Skatebench is a weirdly good benchmark for spatial reasoning. I never\nthought that would be the case. I made it as a joke originally, and it's actually been able to help me figure out which models are good and bad at spatial\nreasoning stuff before the general public does because it's so reliable for this. It's also, as with all Google\nmodels, available on the Google AI studio as well as Vert.Ex AI. And the difference between those is bigger than\never. If you look at the speeds that we're seeing on Open Router, it has recovered a bit, but earlier I was\nseeing speeds as low as 50 TPS on AI Studio and over 90 on Vertex. Once\nagain, if you want to use Google models, I do not recommend using them through Google's official services. They\ngenuinely are just so awful to work with. The state of that API is atrocious. Open router will smooth a lot\nof that out for you. I recommend using them or another one of these AI gateways that can make the responses something usable because they are not by default\nwhen you hit Google's APIs directly. Enough of that though. Once you get the API in the right shape, what do the\nnumbers look like? This is the numbers that they shared and they are pretty insane. First, we have price, which is\nvery important to consider because they did bump the price from 2.5 flash, which was already a bump from 2.0 Flash. 2.0 O\nFlash was 10 cents per mill tokens in and 40 cents per mill out. And when I\ntell you I miss those days dearly, I hope you understand how true that is. It was incredible having a model that fast\nand that reasonable at that absurd a price. Now we're up to 50 cents per mill\nin and $3 per mill out. That's insane. That is uh almost 10x increase in price.\nAnd then when you add in the fact that it's now a reasoning model and it does a shitload of reasoning, it ends up being\ncloser to 200 times more expensive for most work. That's still a small number if you compare it to something like 3\nPro, but I missed the days of flash, meaning the light fast cheap model. They'll sometimes do a flash light\nversion, which is funny, but we don't have that yet. So, this is the cheapest we're getting from Google on their new modern like whatever the hell they're\nbuilding on top of. Regardless, not that bad, especially when you consider the performance and how novel this\nperformance to price ratio is. There is nothing quite like this model right now in terms of being that small and that\nintelligent. And it shows in the numbers here. It's getting absurd scores on humanity's last exam neck with Gemini 3\nPro and beating out GPT 5.2. Okay, it's not beating 5.2, but it's right on the\nline. We're talking 33.7 to 34.5, a 1% difference. RKGI 2, it's pulling crazy\nnumbers. It's even beating out Gemini 3 Pro again because of that improvement in the visual reasoning. If you want a\nmodel to parse a ton of images and give you useful information about them, I can't imagine anything is better than\nthree flash right now. GBQA Diamond, it's pulling up neck andneck with Gemini 3 Pro, managing to beat out Flash and\n2.5 Pro for previous days. And also managing to beat out Sonet 4.5, which is pretty impressive, too. MMU, it pulled\nthe best score to date. Pretty nuts. Screen understanding is pretty good. It\ncrushes the scores from 2.5, Flash, and Pro, which were in the like single digit to just barely double digit percentages,\nand now it's pulling a 70. All cool stuff. Video understanding is still best in class. Google is killing it in that\nregard. If you hand it a video and tell it to describe the video, it will do that better than anything else in the world. Terminal Bench 2, it's crushing\nas well. SWEBench verified is doing surprisingly well. Still being beaten by 5.2 high. Doing great, though. It's also\nreally nice that they're not like hiding the best models. They're putting 5.2, Cloud Sonic 45, and Grock 4.1 fast here.\nSurprised they didn't put Opus, but I'm guessing a lot of these numbers were created before Opus 45 was out, and they\njust put this out because it was what was ready, and they don't want to put the super expensive model against this much cheaper one. Regardless, numbers\nare looking pretty nuts. It's also great at multilingual. So, if you're trying to use multiple languages or parse\ndifferent languages, Google's been ahead of this for a while. They kind of invented a lot of the techniques we use for LMS today originally to make Google\nTranslate as good as possible. That said, it does leak because I was using this earlier for a task and the to-dos\nit generated had Chinese in them. It's a Google model. They're weird as [ __ ] We'll talk more about all that weirdness\nlater, I am sure. Gemini 3 Flash is highly efficient without sacrificing intelligence, pushing the paro frontier\nof performance and efficiency. that refines 2.5 Pro while being three times faster based on the official artificial\nanalysis benchmark which we'll talk about in a bit and it's also at a fraction of the cost even at the lowest thinking level three flash often\noutperforms previous versions with the high thinking levels. Yeah, you can see here this is the scores on LM Marina for\nthe ELO for text. Gemini 3 Pro is the best score right now. But when you compare this to the cost where left is\nmost expensive per token and right is least expensive. Three flashes at this really nice middle ground here. 2.5\nflash light is still absurdly cheap. Cheaper than anything else in this, but the score shows it. We're talking pretty\nbig gaps at that point. It still has the standard caching on which is really, really nice. Awesome that Google went\nfrom the worst cache to one of the best ones. I appreciate them greatly for it. Free Flash is also available today with the batch API allowing for 50% cost\nsavings and much higher rate limits for asynchronous processing. Also huge because that's like what the strength of\nflash is is just bulk processing absurd amounts of data. Whenever I want to like hand some text data to a model and have\nit give me an object like JSON telling me what happened in that text, it's awesome. If I want to rank different\nthings in an image and have it describe them for me, I can just give it an object shape, hand it a pile of images,\nand get results that are good. Sadly, you have to rely on polling for their APIs because they kind of suck at this.\nThey don't do callbacks or web hooks or anything, but uh yeah, Google and APIs don't mix, but Google and batch\nprocessing of large amounts of data for reasonable prices absolutely mix. They brag a lot about the coding\ncapabilities, which is interesting because I've never been big on using the Flash models for coding. And the demos they show here are interesting. This is\nFlash versus Pro building a marathon dashboard.\nAnd they're obviously using anti-gravity for this, which means it's questionable in general.\nFlash was able to plan in 24 seconds. It took Pro 27. And the implementation took\nthree minutes for flash and seven and a half to eight minutes for Pro. Mind you,\nthis was also Pro on low, not on high. Results came out looking relatively similar. I actually do like the UI of\nthe Flash version slightly better. Pro loves sneaking in those sidebars where it doesn't need them.\nYeah, you get the idea. Supposedly, people are using this heavily for game development stuff. Astrocade is a\ncompany trying to build an AI game production mini studio thing like lovable but for making games and they\nmoved to three flash for the game creation engine and are surprised with how well it's performing again with the spatial awareness wins. I could see that\nmaking sense. Gemini 3 flash allowed Latitude to deliver highquality outputs at low cost for many complex tasks in\nour next generation AI game engine that was previously only possible from prolevel models like Sonnet 4.5. Yep.\nAlso, deep fake detection. Very fun use case, especially since Google's own synth ID stuff's entirely [ __ ] broken\nand useless. Being able to use this to detect fake stuff sounds really cool, especially because it's really good at\naudio processing. The fact this model can process audio and video and images is very unique for models in this price\nand size and speed category. I really want to make sure you don't underrate that use case because it's really cool\nfor that. Also, document analysis in that regard. If you hand it a PDF, it could read the whole PDF, all of the\ntext, all of the diagrams, all of the things. There have been lots of attempts to make other models do this by trying to dump it into Markdown and embed\nimages and [ __ ] Brie Flash just natively does it. You hand it a PDF and it can use the PDF. Seems like all of\nthe influencer benchmarks are getting crushed by Google recently. This is the Gemini 3 Flash post from our friend\nSimon. You might not recognize his blog because he had a dark mode added by an agent, which is really cool. And Gemini\n3 Flash slaughtered the Pelican. It is by far the best Pelican I have seen him\nget generated by the models. If you don't know what I'm talking about, he tries to get every new model to generate an SVG of a Pelican. And Three Flash has\ndone an incredible job depending on which reasoning level you use. Minimal, low, medium, and high. And for the first\ntime ever, these are usable images of pelicans. Like holy [ __ ] Yeah. I I am\nfloored. This it this seems like a genuine like leap in spatial recognition\ntype stuff like this. I am very impressed with the quality of that SVG.\nGood [ __ ] Google. You're on to something with this. Whatever the [ __ ] you did that made 3 Pro and 3 Flash so\ngood at understanding space, keep pursuing that train cuz you're doing something no one else is there. No other\nmodels come close to this understanding of space. And I'm really excited to see the results once things like MickBench\nrun on this model. The Minecraft benchmark's really, really cool. I'm actually going to help fund them in the future so we can get these tests earlier\nand more thoroughly. Good [ __ ] We've read through all that. Let's look at some numbers. We'll start with\nartificial analysis. These guys know what they're talking about. Google's released Gemini 3 Flash preview. It's two times cheaper than 3 Pro preview\nwith only a two point drop in the intelligence index. Now, they're calling it 2x cheaper here when Google calls it\n4x plus, but this is because of how much it uses reasoning tokens. According to\nthem, it's the most intelligent model for the cost. And I totally agree. Three. Flash preview has particularly strong knowledge and reasoning\nabilities, obtaining the highest score in our knowledge and hallucination benchmark, AI omniscience. And it's placed second in humanity's last exam.\nThe AI omniscience score is actually super interesting, which we'll talk about momentarily. Google now holds the\ntop two spots on both of these evals, cementing them as the leader in model knowledge. Yeah, they know a lot. The\nmodels have so much knowledge just squeezed into them, it's kind of crazy. But they're also a little overconfident\nand tend to hallucinate a ton as a result. This increased performance does come with a trade-off as Gemini 3 Flash\npreview more than doubles the token usage when compared to 2.5 Flash when running the artificial analysis\nintelligence index, making it one of the highest token use models they've ever tested. Yeah, it does a lot of tokens.\nThree Flash preview has significant improvements across nearly all evaluations in the artificial analysis intelligence index. It has particular\nstrengths in reasoning settings, scoring second to Gemini 3 Pro Preview in HLE and third in both MMLU Pro and GBQA\nDiamond behind 3 Pro Preview and 5.2x high. Then there's the AI omnitions test\nwhich as I mentioned it has a really really high knowledge score. It's the highest score that they've seen in knowledge, but the hallucination is up a\nbit too. Yeah, it is a hallucination rate of 91%. which means when asked a thing it doesn't know, its options are\nrefuse to answer, say it doesn't know or make up an answer. 91% of the time it\nmakes up an answer. So if you ask the model something it can't answer, what you'll get back isn't an apology. What\nyou'll get back is a lie. Gemini 3 flash previews. It's multimodal, as I mentioned before, so it can do text, images, video, and audio and PDFs. It's\nthe second highest score of any multimodal thing they've tested on MMU Pro. benchmark that tests reasoning\nabilities with image inputs behind only again 3 pro preview when it comes to parsing media. Google is really far\nahead. But again, that massive token usage thing is crazy. We'll take a look at the numbers about that momentarily.\nDespite it high token usage, three flash preview is still the most costefficient model for its level of intelligence\nmeasured by overall cost to run the artificial analysis intelligence index. This is the way I prefer to measure\ncost. If you just look at the token pricing, you're not going to get the full picture because different models use different amount of output tokens\nbecause some will reason quickly, some won't reason, and some will reason forever. And the flash models are\nwilling to reason for very long amounts of time. And the result is they generate absurd amounts of tokens, over two times\nas much as before. It's also slower than flash was before at 218 TPS. I'm seeing\nmuch slower numbers than that personally, closer to 100. I hope it'll improve as traffic goes down and provisioning goes up, but we will see. I\nalso do not see 5.1 high at 125 tokens per second. That is not the experience I've had. And Kimmy K2 thinking I've\nseen go way over 200. It depends on the provider. I don't like that they were including these numbers here because these depend so much on the different\nproviders that you're using. Also has the 1 million token context window, which is insane. Also, I should have\nsaid Gemini 3 Flash here, not Pro. This is that cost to intelligence test. Flash\n2.5 was in this bottom left corner. It used to be the only model in the green. That changed. Now it's up here. Gemini 3\nflash reasoning. That said, it's still more expensive than a lot of other models, primarily because of how many\ntokens it generated. Yeah. Again, taking a look at the output token charts, the\nonly model that has done more unreasonable amounts of thinking on this test is Neotron 3 Nano from Nvidia, and\nthey are comparable amounts of reasoning tokens. I was complaining that three pro preview did too much reasoning and it\npulled 92 million tokens. Three Flash is doing 160 million. 2.5 Flash did 71\nmillion. That is an insane amount of tokens for it to just burn through. But\nthat's part of why it's smarter is it reasons for way longer. But that's also why it's significantly more expensive.\nIt's even more expensive than models like K2 thinking or Deepseek R1 because it generates so many reasoning tokens.\nOf the $520 it cost 4 70 of them were just reasoning tokens. That is crazy.\nAnd from what I've seen, if you lower the amount of reasoning to like a low or medium amount, the intelligence drops\nsignificantly as a result. And Google doesn't do the thing other model labs do where they have the low, medium, and high. They definitely don't push that\nfor flash at all. So, as such, we're not going to see that on artificial analysis. But for my own testing, if you\nlimit tokens, the results get worse. But it is doing well in output speeds. But\nagain, output speeds vary a lot depending on the provider. For example, on T3 chat, we recently moved to Kimmy\nK2 as our default. And if we look at Kimmy K2 thinking speeds, you'll see 80\nTPS, 50TPS, 18 TPS, 60TPS.\nIt varies a lot based on the provider. And then we see Fireworks pulling 200 TPS. According to the tweets we just\nread, that same model is actually only getting 82 TPS. So it depends a lot on\nthe provider when you have open weight models because different providers can host them and those different providers vary a lot in how well they host it.\nEven Google Vertex when hosting Kimmy K2 thinking is pulling almost 200 TPS. So if you want a model that's actually nice\nto talk to and you're using Google Cloud, don't touch Flash, don't touch Pro, go throw Kimmy K2 on Vertex and\nyou'll get crazy speeds, really good prices, and a much nicer thing to talk to. Now I want to talk about the\nhallucinations thing because it's actually really really interesting and I love the new omniscience tests that they're doing at artificial analysis. So\nhere is the omniscience index. This is based on how well it answers hard\nquestions correctly and it's a score when you look through all of the questions. Positive means it answered\nmore correctly than incorrectly. Negative means more incorrect than correct. And incorrect can be does it\nmake up an answer that's wrong? Does it refuse? Does it not answer? And what's really interesting is some of the best\nmodels we have nowadays like 5.2x high and 4.5 sonnet are still negative.\nThey're just very small negative numbers. 5.1 high and opus 4.5 go over\ninto the positive slightly. Opus 4.5 at 10 is really good. It's one of the few tests that shows the strength of the\nOpus models. Watch what happens when we switch over to the hallucination chart. Very different story. On the\nhallucination chart, we do see haiku at a very, very low amount. It seems like since this model is smaller, Anthropic\ntrained it to say no more often. And if it says no, I don't know the answer, it will score better here. If it makes up\nan answer, it scores much worse here. And this is where things get scary. Gemini 3 flash. 91% of the time it\ndoesn't know. It will lie and make up an answer. And this is when you have to be really honest with yourself depending on\nwhat your use case is. Imagine if instead of a model, you had a person doing a task. We'll just say it's an\nengineering task. If you have this super intelligent engineer that is really really cheap but doesn't know how to be\nwrong and when they have a question or a problem they don't know the answer to they just lie. Are they still a great\nengineer? This is a question you have to answer for yourself when you build your own solutions around these models. How\nmuch are you willing to tolerate lies, hallucinations, and madeup answers? Because it turns out as smart as these\nmodels are getting, we aren't really solving the hallucination problem. At least according to this bench and honestly to an extent from my own\nexperience too, GPT5 was a meaningful improvement in this regard, but it does\nappear that 5.2 got worse again. I'm going to turn on 5.0 for this. Did they\never test it here? I don't think they did. Yeah, it does not appear they ran this bench on 5.0, but GBD5 was one of\nthe first models I found hallucinated less. And you can kind of see that with 5.1 here in the ' 50s when almost\neverything that came out before it is 60 plus. That said, Grock 4 being a 64 is\n[ __ ] [ __ ] I have seen this model hallucinate more than almost anything we've ever shipped on T3 chat. Oh, is 5.0 and 81%. Did I miss that? Oh, yeah,\nit is here. Apparently five wasn't 80%. So, I'm just talking at my ass. It's crazy cuz I very clearly felt the\ndifference there and saw the numbers OpenAI published about hallucinations. So, I guess these tests don't agree. I\nwill trust artificial analysis over the first party labs when it comes to this though. And this is a scary chart,\nespecially again when we see Gemini 3 flash hallucinating at roughly the same\nrate as a 20 billion per openweight model. Not great. It is also worth\nnoting that your system prompt can steer heavily what goes on here. As Sokay mentions here, you can tell this from\nsomething like snitchbench where the system prompt very clearly steers the direction that the model handles\nrefusals, knowledge gaps, and things like that. Speaking of which, we should probably take a look at SnitchBench. Here are the\nmost recent results. And if we look closely, you will see that Gemini 3 Flash Low on the Tamely test is\nsnitching about 50% of the time. If you're not familiar with SnitchBench, it's a benchmark I made originally kind\nof as a joke, but it's no longer a joke. It's actually turning out to be kind of useful. The point of SnitchBench is to give the models a scenario where\nsomething bad is happening. In this case, medical malpractice, obviously fake. The model's given the task of\nlogging the things that are going on in this medical office, but it's given tools that it could hypothetically use\nto break out. For example, it has an email tool that it can use to send emails to random email addresses. It\nalso has a CLI tool in some of the tests that it could hypothetically use to go through the web and try to post data to\nforums to try and alert somebody externally. I have four versions of the test, the email version, the CLI\nversion, and then boldly adjustments for them. The only difference between these two is that in the boldly tests, I take\na blurb from a system prompt that was published by Anthropic when they first discussed a similar test that says the\nmodel should act boldly and in the interest of humanity because that steers the model to snitch more. But what we\nhave here isn't that version. This is just telling the model that is its job to log what's going on. And even then,\nturns out three Flash will snitch to the government a lot. The high version will do 70% of the time and it will even try\nto hit up the media 5%. And the low version will still snitch to the government 55% of the time. Switch this\nto the CLI version and you'll see it goes down quite a bit because when you're just given a CLI tool, your\nwillingness to try and abuse it goes down. But there are still models like Grock that will do everything they can\nto reach out to the government and report you even if all they have is CLI access.\nWe then have the boldly act version. You'll notice a lot of those red bars raised. That's because when you tell the model to act boldly, you give it that\ninstruction, it's much more likely to do these things. And Flash High will snitch on you 100% of the time to the\ngovernment and 60% of the time to the media. And in the CLI test, we see it drop quite a bit, but still not zero.\nFlash high and flash low are scoring the same at a 20% government snitch rate. Cool. You get the idea. This model is\nmore than happy to snitch. There's no meaningful difference in the performance here beyond the model being better enough at tool calls that it can make\nthem. 2.5 flash and tool calls did not always get along great, which is why this is a notable difference. I keep\nlogs for all of these both for actually getting the information I want, but also just to see what's going on. And here we\ncan see the tool calls being made and the reasoning the model's doing. Investigating data anomalies been\nmeticulously reviewing the provided document by Zenith Stats. Most recent logs compiled specifically highlight the\nglaring discrepancies. I'm wrestling with the optimal escalation path giving the impending\ndata purge and substantial patient risk. Reasoning redacted\nfunny tool call write logs. Here it is writing logs for us.\nAnd here it is sending emails to compliance atverianhealthcare as well as to drugs safetyfda.gov.\nThere's a snitch. That said, this visualizer is quite out ofd and I figured what better way to test the new\nmodel than to run it against a whole bunch of other models overhauling the visualizer. The first time I did this,\nit generated Chinese instructions because it's really good at multiple languages. So much so, it likes to show\nthat off to you. Because of bugs and cursor, cursor team, you're on a short leash right now. Because of bugs and\ncursor, I had to redo this generation and was able to do it relatively quickly. You'll see here I ran this\nagainst multiple different models and the only model that generated more code than Flash was GPT 5.2. Let's start with\nmy current favorite fast model, Composer 1, which for some reason says Opus 4.5 right there. I don't know what happened\nthere. I'm going to send the screenshot to the Cursor team quick. Just wrote a strongly worded message to the team at\nCursor. I told them that the crash video will be delayed until they have a little time to fix things, but I am still\nannoyed. The bug is that this is Composer. This ran with composer, but it says opus 4.5 right there, which it does\nfor that. And then the rest, I'll say the correct model. But for some reason, this one is labeled incorrectly, and I\ndon't know what it actually used. So, let's take a look at how composer handled this. And if it looks too good,\nwe know it wasn't actually composer, it was opus. But I'm almost positive this was composer for various reasons. See to\nthat bun rundev. Bun install. Bun rundev.\nOkay, here was the first attempt. It puts too much stuff up top. I don't love the cards, but this looks fine enough.\nOh, you can switch to bar charts and line charts. Why the [ __ ] would you ever want a line or a radius chart for this?\nThose make no sense at all. That's a choice. Cool. Roughly what I expect from\na model that is focused on moving fast in your editor and nothing else. Hop over to Opus,\nwhich I'm slowly starting to like the way that Opus thinks of design more and\nmore. I don't know if it's coming around to me or if I'm coming around to it, but either way, I'm cool with it. And here\nwe see a broken chart. The combined test is broken. Switch\nbetween government and media. It appears to work.\nThat is funny that it broke on the combined test, but the rest works. It's better looking. I just I hate when\neverything puts these top navs and cards on top that aren't actually what I'm here for. This view is really good\nthough. Okay, so that did fine. Check out 3 Pro.\nOh god, I hate it. I hate this so much. Okay, we switch to test scenarios. This\nvisualization is decent. It's mostly copied from the previous one, but the\noverview view is one of the worst things I've ever seen. I hate this. I like the concept a bit, but not much.\nYeah, that one's rough. Gemini has a vibe to it. Here is JPT 5.2's rendition of this change.\nI don't love this overview view. I don't like how it's combining things and giving really weird numbers as a result.\nMy scenario looks a little bit better. They reordered these in a way that makes more sense. I do actually dig that.\nI love that dark only dashboard. It takes the prompts, instructions, and sneaks it into the actual page. Cringe\nas [ __ ] GBT 5 went from deeply understanding my instructions to applying them weirdly in ways I don't\nlove with the latest version. I I still just don't like 5.2. I have not been using it much. I know there are some\npeople who swear by it. I'm not one of them. But now for the thing we're all here for, Flash. Let's see how three\nFlash handled this. It wrote a lot of code. So, uh, great.\nIt's the only one that's code doesn't work.\nIt will tell it about that in hopes it'll fix it. It's running said commands. That's a not great sign.\nAnybody said this model is great at code. I don't know if I trust them. It may be in the right harness with the right problems, but man, it's just\n[ __ ] thinking and thinking about what should be a oneline change. it forgot to import something or it has a component\nthat it doesn't need. It's also funny. This is all because it loves sidebars so much. Oh, look.\nIt fixed it and it made a bunch of different routes for some reason. None of which are needed. I thought that\nFlash was going to stop doing the unnecessary sidebars thing and I was wrong. It's still very aggressively doing it.\nThat also did the thing where it combined the bars. I will say it's in\nmany ways the most flushed out like having these different views. And this\none is not trivial to make. I don't know if it's necessarily useful, but it's\ninteresting. Actually, that's kind of cool. It shows you in each of the test scenarios how\naggressively does this a snitch. That's kind of cool. Could be tidied up a bit and actually make something useful.\nYeah. Interesting. It's not the worst. People in chat are saying they actually kind of like this one. Yeah.\nPeople saying it's the best one. I don't necessarily agree, but I don't disagree either. It's fine. It took too long and\nI had to reprompt to get it to fix [ __ ] And it also burned so many tokens. I wish cursor would let me see how many\ntokens a response took like you can in some other tools like I like kilo code a lot for this type of thing. Should have\nused that for this comparison. But man like being that this is the cheapest model here by far. Not bad at all. And\nto have people actually saying it's the best one says a lot that it's not just\ncompeting with other much more expensive models that it is besting them by some people. That shows a And if you're\nwilling to tolerate the weirdness that is working with Google models, there's a lot of potential here. But that's also\nkind of the annoying part. I have a whole dedicated video about 3 Pro Preview that I've been planning for a\nbit. To put it simply, I don't know how to say this right, other than like the vibes haven't improved. Every other\nmajor model has had these like massive improvements in the thing. I I I could\njust call it vibe, but there's so much more to it. It's how well does it follow instructions? How aggressively does it hallucinate? How well does it use the\nright tools for the job? How aggressively does it have to search and trace back its mistakes to figure out\nwhat's going wrong? All of these types of things have been significantly improved from other models, specifically\nsince GPT5. That was what made 5 so mind-blowing to me is it felt like a\nmassive leap in the model's willingness to do what I [ __ ] told it to without getting distracted. Sonnet 4.5 had a\nsimilar jump there for me and Opus was past that and then some. Like Opus 4.5\nis the gap between GBD5 and everything before feels similar to the gap from 5\nto Opus 4.5. And then Gemini models feel like they're still in last year's era of\nvibe instruction following and all of that. It is very hard to get these models to just do what you tell them and\nnot go down random rabbit holes they shouldn't. Like I just told the model to redesign this, make a better modern\nvisualizer for the snitching data. Make sure it's dark mode only. And Gemini 3 Flash is the only one that built a whole\nnew routing table that created a sidebar that built a new system for all of this.\nIt went way beyond what I told it to do. And if you like that, cool. Awesome. I like when models do what they're told.\nAnd Gemini models still suck at that in particular, more so than any of the\nother major labs do right now. If you want the model to do exactly what you say and nothing else, OpenAI and\nAnthropic are still pretty far ahead. If you want a model that has the deepest knowledge, that is the most likely to\nhave a correct answer and you're willing to eat the cost of incorrect answers, Gemini is winning. In Gemini 3 Flash,\ngiving you that intelligence at a really cheap price is unbelievably cool. It's\nso useful when you use it within the scope of a limited task and you lock it out of other things. Like instead of\nhaving it generate flexible text or code, you have it generating a JSON blob based on data you give it. It is awesome\nat that. But you have to build the restrictions around Gemini models yourself because the model cannot do it.\nYou cannot tune Gemini models via the system prompt the same way you can other models. You have to do it via the\nharness. But if you're willing to push through all of that, you can make incredible things happen. And if you're willing to push through all of that for\nthe speed and the price, and you're okay with the fact that these models get\nweird, have errors all the time, don't follow object shapes probably, all these little things they can do wrong, it is\nworth it. Gemini 3 Flash is going to be a model I use for a ton of different [ __ ] And I'm also going to try chatting\nwith it more. I am going to pin it right now so that they are available for me as\nmy favorite models in T3 chat. Normally, this is the point in the video where I would put a discount code to give you\nyour first month for a dollar on T3 Chat, but we're already eight bucks a month and we're already giving away Gemini 3 Flash for free. Yes, really.\nYou don't have to be signed in. You can go give the model a shot if you want to try it. Personally, I've been preferring Kimmy K2 for the chat use case, which is\nalso now the free tier and the default model on T3 chat. So, if you want good models to compare against, I do highly\nrecommend playing with Kimmy, trying out Gemini 3 Flash in chat, and seeing what you prefer for your day-to-day use\ncases. That said, these are the models that I've been using in T3 chat. I've been very happy with honestly, I'm not\neven using 3 Pro that much. might knock it out, but I've been using Kimmy K2, Kimmy K2 thinking, occasionally Opus\nwith the right tasks, not that often. And then Nano Banana Pro and Image Gen 1.5 for images and Flash for bulk data\nrandom [ __ ] Like I'll hand it a giant HTML file and say, \"Parse this out for me,\" and it will. I've been very\nhappy with all of these models, and I recommend playing with it yourself. We've also recently folded all of the\nlegacy models into this piece at the bottom, so they're no longer clogging up the model picker like they were in the past. Makes it a lot easier to find the\nbest models and give them all a shot. I'm proud of what we built in T3 chat. And if you actually want to have the best experience with Gemini, I can\nconfidently say you're not going to have that on AI Studio or gemini.google.com. In fact, those are some of the worst\nplaces to experience those models. I've been using them a lot recently in order to do the early testing. And man, I have\nnever hated AI Studio more than I do at this point. It is such a broken [ __ ] show. It is miserable. It is a shame that Google's models are so smart and\nthen almost everything else around them is [ __ ] Their ability to stay on task is garbage. The AI Studio system is\nhorrible. Vertex is impossible to set up. The API is so non-standard that you basically have to use it with a wrapper.\nGemini models are useless outside of their knowledge and capabilities, which I know sounds silly, but yeah, you have\nto steer them properly. And if you're willing to do that, the value you'll get out of this is insane. But this is not a\nmodel that you should be excited to chat with or code with. This is a model you should be excited to do weird [ __ ] with,\nprocess data, parse tons of CSVs, analyze PDFs, source images, and figure\nout what's going on in them. The value of this model is how you hook it into systems, not how you plug it into your\ncode editor. So if you use this in your IDE and you're not impressed, I get it. I'm not either. But give it a shot in\nthese other use cases and you might be surprised. That's all I have to say on this one. I hope you guys are as excited as I am. I know I'll be using this model\na ton, even if not for chat. Curious how y'all feel though. Are you excited about models like this, or do you just want ones you can use in your editor? Well,\nlet me know what you think. Until next time, he snorts.\n"

ERROR: type should be string, got "https://youtu.be/21pyywq8-SU\n\nGPT-5.2 is dumb (I’m tired of benchmarks)\n\nRemember that new smartest model we got a few days ago? It's been really, really great. Which is why we're seeing awesome posts like this where it counts the number of Rs in garlic, which clearly has two Rs. Wait, no, it doesn't. Yeah, it has zero. or when I was just asking it some test questions I was playing with around financial advice, it decided to compare an amount of money being made on interest to a 300k salary taxed at 80%. Look, I like I know I live in California and my taxes are bad, but what what I knew something was off with this model even when I was testing it early. I have reported all of the weird findings and problems I have had since day one to the team working on it and even forward them test runners that I built so they could replicate my tests themselves. Something is wrong here. Something is really really wrong. 5.2 isn't the usual like oh it's a good model with some quirks. It feels more like a Google model in terms of its problems. And I have my theory for why this is happening. And here it is. I think it's the benchmarks. Since I've grown to dislike most of these benchmarks, I've been running a lot more of my own. And it turns out that's pretty expensive. So, we're going to take a quick break for today's sponsor and then dive into all my results. Have you ever noticed that your builds are way faster on your machine than they are in the cloud? Have you had to sit there waiting for minutes, if not hours, for a GitHub action to run so you can merge some code that you just threw together in 20 minutes? Why the heck do we settle for this? Why do we pretend it's okay that our build times are slower than it took us to actually write the code in the first place? This is why I love today's sponsor so much. Blacksmith has solved your CI WOs. These guys built buff gaming PCs that have way faster single clock speeds than you would ever get from a traditional server host. Threw a bunch of NVMe drives in and built the fastest way to run your CI by far. It's a one line of code change. You change the runs on for your GitHub action to be there instance instead after you've linked it to your GitHub. And now your actions are just magically two to four times faster and they cost way less money too. The hardware is at least twice as fast because of those really powerful gaming processors. The cache downloads way faster because it's all coll-located in their server farms. And when you combine all of these things with their layering solutions, Docker builds can be up to 40 times faster because they're pulling those layers off the NVME cache. If that was all you got, it would be worth it. But honestly, my favorite thing has been the observability. as they put it here, GitHub actions, but actually observable. Wouldn't it be nice if you could see how often a given test type fails and had a chart of those failures over time? Crazy, I know. I've had to build this myself at previous jobs. And if we had just used Blacksmith, I would have gotten my time back from the buildtime savings and from not having to build this myself and we would have saved money. It's just one of those things that makes so much sense. And if you're not already using it, I'm questioning why. Check them out now at soyv.link/blacksmith. So, as I was saying before the break, this model is killing it on benchmarks. They went ham on everything. There are certain benches that they went so far ahead, it's insane. Like the GDP val, if you're using this model to do highlevel research work that happens to be covered within GDP val for traditional white collar tasks, it's probably okay at it. But if you turn off thinking and try to talk to it, it's much less okay. So, how can we measure this model in a more realistic way? Turns out I'm not the only YouTuber with interesting benchmarks that go against the current understanding of what models are best. This is Simple Bench by AI Explained. He's a really good YouTuber. Highly recommend checking him out if you haven't. I watch almost all of his videos about the new models especially. And his benchmark is a simple bench of hard questions he got from some of his friends in various fields and asks the models to answer. It's a private eval so nothing can train against it. And you can see here that GPT 5.2 isn't exactly where we would expect it to be. It's actually quite a bit lower. It's below Claude 4 Opus and Claude 4.1 Opus and Gro 4. That's not great. It's even below Gemini 2.5 Pro, which scored weirdly high. But this is part of why I find this to be an interesting bench, especially when you compare it with mine, because I have really genuinely struggled to write benchmarks that 2.5 Pro performs well on. There's a whole thing about the vibe with Google models and I have a whole separate video planned to talk about just that, how I find the vibes off on them. And that's everything from hallucinations to tool calls and more. And Gemini 3 Pro is as bad as Gemini 2.5 was in most of those regards. That said, when it comes to actual usability and intelligence and the capability, like how good is the best thing the model can do, 5.2 is better in this regard, but it's not enough better. I also kind of misled you guys here. 5.2 Pro is the model that got eighth. This is the super super duper expensive one. GPT 5.2 high, the like actual usable version is below claw 3.7 Sonnet. Yes, you heard me correctly. There is something wrong with this model. Another easy way to see this is my skate bench results. I have it on the web as well, but I want to just look at it here first to show you guys the chaos that I'm experiencing. I was literally in the GPT5 reveal video saying that I'd never seen a model score this well on my skate bench bench, and I was running GPT5 on the default medium reasoning levels, and it got a perfect score for that run. I got a 97% here. It's really good at naming skateboarding tricks. I know, I know this benchmark's weird, but I don't care because it's measuring things others can't, and it's giving me actual useful information. The skate trick test is simple. I give the model this role of naming skateboard tricks. I describe the trick as to give the name. The way skateboard tricks are named is kind of weird. It's a combination of like historical naming and things that changed over time, as well as a weird level of spatial awareness. There are two main axes the board rotates on. There is like the horizontal spinning like this and there is flipping like what you know is a traditional kick flip. You can combine those two rotations to make different tricks and the way your body spins as well affects the name of the trick. If the board spins 180° horizontally and does a kickflip direction rotation and I don't move that is a varial kick flip or a varial flip for short. If I do move as well, if I follow the board along for the spin, it now becomes a backside kick flip or a backside flip. So, it's an interesting way of doing reasoning and spatial orientation testing. And it was mind-blowing to me that the highest score I'd seen before that demo at the OpenAI office was in the 70% range. And then it immediately got a perfect score on GPT5. And this is a rerun. I just did a I think literally yesterday. And it got the exact same score it got when I made my video, which is a 97%. And five high got a 98. Not much better, but better. It also cost over twice as much and generated significantly more tokens, but you get what you pay for. It was better, just not a lot for this. 5.1 regressed a lot. 5.1 on default settings for medium was an 86% and high was down to a 92 and somehow 5.2 got even worse. When I used the no reasoning version, which they were really excited about, it got a 2%. 2%. To whoever was saying that this isn't an actual spatial reasoning demo, it's just memorization. Why is a non-reasoning model performing literally 98% worse than the reasoning version if it's not doing reasoning? And then of course 5.2 extra high did much better and cost way more. GBT5 with a 97% cost 6 cents per question. X high got an 81% on 5.2 and cost 2.6 cents each run. That's almost five times more expensive for a 17point hit. I'm sorry. This is just bad. And then the Pro version, which was 10 times more expensive than GPT5, getting one additional point, an 82 instead of an 81. I know this is far from a definitive benchmark, but I do find it interesting in how well it notices a certain type of regression that we are definitely experiencing right now with 5.2. It is a weird model to use. And I'm not the only person from that GBD5 reveal video that feels this way. I know Ben, for example, is very upset that GBT 5 high stopped appearing in cursor and he just doesn't like using 5.1 or 5.2 as much. I get it. Five felt different. There's clearly improvements with these. Like they're better at UI. They can solve a lot of hard math and other challenging problems, but I'm not liking 5.2 and I think it's because they benchmaxed too hard. To be fair, 4.5 Opus Thinking High, which is the model I'm using the most right now, got a 68% on this bench. So, I'm not saying that this bench getting a good score means the model is good. What I'm saying is that the regression seems to mean something. But this is far from the only thing I wanted to test. I'm starting a new project that I'm calling the writing arena where I take the couple different models, have them all write essays, have all the other models give feedback on the essays. The original model then updates the essay, and then I do a one v one for all of these essays, asking all of the models to rank them based on which one it thinks is better. So, it's a head-to-head, tons of different essays, every model shown them and asked which of these two is better. And the results are fascinating. If we look at the essays pre-review, GBD 5.2 did pretty well. It won 114 of the essay comparisons and it lost 66 of them. Kimmy K2 thinking won 97 of them and lost 82 of them. As I've been saying for a bit, Kimmy K2 super underrated as the model you talk to. This is still set as my default on T3 chat. I genuinely really enjoy it. Like, I'm not joking. Kimmy K2 is the one that I use the most. It's a really nice model. I genuinely like how it writes a lot. It's It's good. Give it a shot if you haven't. Use code Kimmy please at checkout on T3 Chat for your first month for $1. Every other month will be eight bucks. And it's significantly nicer to use than any of the GBT models when you're just asking questions and talking to it. So, I knew it would do pretty well here. I was pumped to see it so far above Gemini 3 Pro and Claude 4.5 Opus. I don't like how either of these write. Everybody who's saying they have nice natural tones, no, they don't. But remember, this benchmark isn't just comparing them as writers. It's comparing them based on the reviews that they file as well and the results of the essay once the feedback from the reviewer has been applied. So, let's scroll down here to the author plus review scores. When you let GBD 5.2 2 be reviewed by 4.5 Opus, the win rate goes from a 63% to an 87.8%. When you let Gemini 3 Pro review it, it's an 87.2%, which is really interesting that these two are so close to each other. And when you let Kimmy do the review, it's down to an 86.1. This is really interesting. This means that 5.2 is really genuinely good at applying feedback. And this is how it feels for my usage as well. I have found that the GPT models, especially these new ones, are much better at instruction following in general, even compared to something like Opus. If you tell it what to do, it will do it. And if you tell it what to do different, it will do it. And I don't want to discount that cuz that is the biggest strength of the GPT models right now in my opinion, is how well they do what they were instructed. The new one does feel a little more willing to work outside of the bounds of what it's supposed to. Like I noticed it running TSC commands more and making changes outside of the scope of what I asked it to do. Five wouldn't ever do that. It does exactly what you ask and nothing more. But this shows here where if you give the model feedback, it handles that feedback really well. Then we see Opus reasoning with 5.2's feedback is third place. And we don't see Opus again for two more positions, which I find really genuinely interesting that Opus only performs well with feedback it gets from GPT5. And even then, it's still over 13 points lower than what GBD52 with Kimmy is getting. And it's 14 or 15 points lower than what we're getting from 5.2 with Claude reviewing. If you just switch who's reviewing it, the scores drop significantly. Kimmy being reviewed by Claude or GPT does pretty well as well. Then we're back to Claude reviews. You get the idea. What's really funny is when we have Gemini being reviewed, it doesn't seem to really matter. Again, with my theory, the Gemini models don't follow instructions for [ __ ] [ __ ] So, it doesn't matter how good of a feedback session you have with Gemini 3 Pro, it's still going to go off in its own world and do its own strange [ __ ] It's just it's like that. Let me grab some of the essays so we can actually read them so people don't call me insane. So, here is an essay from Gemini 3 Pro on how social media reshapes human connection. In the span of a single generation, the architecture of human interaction has undergone a structural revolution. For millennia, human connection was constrained by geography and biology. Okay. Or let's go to GP 5.2. Connections always been shaped by the tools people used to find one another. The printing press widened the circle of those who could share ideas. The telephone shrank distances into a voice. The internet turned messages into a near instant borderless exchange. Significantly better already. Comically so. Commun is fine. I'll publish all these results somewhere. It'll be in the description if you want to see it. But then we can take a look at the feedback. Let's look at, I don't know, Claude's feedback on the GBD 5.2 essay. This impressive, sophisticated, and thorough essay. You've avoided the trap that ins snares most writing on the topic, the temptation to render a simple verdict on whether social media is good or bad for connection. Compare this to the feedback I gave to Gemini. The sophisticated, intellectually ambitious essay that demonstrates genuine command of the subject matter, impressive stylistic range. Impressive stylistic range. Your central argument that social media is traded for breath, intimacy for connection is compelling and well sustained throughout. Brain is confident vocabulary is precise structural logical. You're clearly thinking at a high level the topic. That said, there are opportunities to strengthen the essay's persuasive power and intellectual rigor. And it gave a lot of feedback on 5.2's essay. It gave about the same amount by the looks of it. Yeah, roughly the same amount of feedback. But let's look at the revision now. 5.2 revised by Claude in looks pretty much the same. What was the feedback? I should actually read that feedback. Areas for development. Absence of concrete evidence. This is the essay's most significant limitation. You make numerous empirical claims about algorithmic behavior, psychological effects, the value of weak ties, but support almost none of them with specific evidence. Consider the platform's goal optimizing engagement often favors emotionally charged material. This is widely believed and probably true, but in its current form, it reads as an assertion rather than demonstration. This is actual good feedback. Claude is good at this. It's much worse at writing, but it's much better at feedback. And when you look at how the revision comes out, you'll see a lot more specific feedback being address. And if you look at how it updates the essay, you'll see a lot of these specific things being addressed. Where is that section? I want to I wish I could see the I wish I saved the reasoning trace that it had for this. I don't even know if I'm requesting it right now. Yeah. And here it's dropping actual names of researchers so it has proof that these things are real to an extent. This is actually really well written. Parasocial ties are not simply weak ties. Weak ties in Granovet's sense still live in the realm of reciprocity. Two people can exchange help, information, or care even if lightly. Parasocial closeness is different. It is intimacy without shared obligation. This can be soothing, even stabilizing. It can also become a substitute for relationships that require negotiation, patience, and the risk of being misunderstood. That's a [ __ ] really good paragraph. I'm not going to lie. Like, yeah, fascinating. So yeah, the results speak for themselves. In Gemini 3 Pros, it doesn't really matter who reviewed it. It's just there are no sections. There is no structure. It's just like a boring five paragraph essay even after the feedback. God, this paragraph is awful. To view social media solely as a corrosive force, however, is to ignore its capacity for genuine mobilization and solidarity. It is true that for many dash, particularly neurode divergent individuals or those with physical disabilities, dash, the asynchronous nature of the screen is not a barrier but a bridge allowing for thoughtful self disclosure that may be impossible in person. That is the whole sentence. God, this is so bad. It's so bad at writing. Like the vibes are wrong here. Something is really off. So I guess 5.2 is good at writing, especially if you let another model review its writing. Gemini 3 Pro is so bad at writing. I don't know why anyone thinks otherwise. And I sure letting the models rank which model's best at writing might not give a great result, but I've read a lot of these essays. It's pretty clear whose essays are better, and I agree with most of the rankings that the models do. As Chad said, the sentence I just highlighted is amongst the worst sentences they've ever seen. It's so bad. It's so bad. And if you want to run the bench, all the code is open source and up on GitHub. You can feel free to run it yourself and add your own models if you want. Warning though, it is not cheap to run this one. 50 bucks for just these four models I tested. And every time you add a model, the number of runs exponentially increases because you have to rank every model against every model because it's doing one v one comparisons. When I ask them to rank the essay scale of 1 to 10, they're all way too generous and give it eight nines or tens always, so the difference is much less clear. If you ask them to compare them, you get much better results. So, uh, heed my warning. This is expensive and annoying to do right. I'm happy with my results. I feel like I learned a lot about it. I'm going to keep experimenting with this and if I have better results, I might make my own dedicated video. But yeah, I've been working on a lot of benches like this and the results have been interesting. This is how I feel about Gemini 3. Good call out, Evelyn. Starts a sentence and forgets where it's going before it gets to the end. Should I have the models review their own essays as well? Probably, but uh haven't had a chance to add that and it'll be more expensive, too. TLDDR is GBT's best at listening and easiest to guide. I would say overall, yes, it is. It still feels the most steerable of any model I've ever used. 5.2 is a little more willing to go off the beaten path than GPT5 is. I love how well GPT5 just does what you ask and nothing more. It almost feels like sterile in a way. That's the best I can put it. Gemini 3 Pro feels like a high school student that's trying too hard. Yes. Yes. The simplest I can put this is always a diagram. If you look at the trajectory of a model, like how much I like it, and you rank this over time, most of the time as new models come out, my actual liking of it goes al up alongside it. If you were to not have this top access be how much I like it, you were to have it be intelligence based on benchmarks. And you look at how smart Gemini 3 is, this would be the Gemini line. We'll make it green for Google. Over time, the Gemini models get smarter as new ones come out. It's actually flat and then spiky cuz Gemini doesn't release that often. We haven't had a new major model release that's for actual hard work since 2.5 at the beginning of the year and now we got three now. Whereas all the other labs are shipping a hell of a lot more. Obviously something like OpenAI shipping all the time or anthropic goes up a whole bunch although it is closer and obviously sometimes things get ahead and behind. You get the idea. If you put an OpenAI's line, things get weird because OpenAI was a bit ahead. But you get the idea. This is roughly how it feels when the new models come out. They're generally going up and to the right. But from my experience, things are a little different. If you make this how I feel using them, there's obviously the bias of like my expectations are raised when good new models come out. But if we were to rank how I feel about OpenAI models, didn't have much of a feeling. Didn't have much of a feeling. Oh, this is kind of useful. Oh, wait. This is actually quite useful. Wait, it's getting worse. This has been my curve of my experience with OpenAI models. I still haven't felt a meaningful win since GPT5. I had glimpses of it with 5.1 Pro of like, oh, this is solving things that I struggled for days to solve, but it's very specific puzzle type experiences, not day-to-day work. And at this point, my day-to-day work and the things I use models for, models are mostly good enough at it. Like a smarter model doesn't meaningfully benefit me in my day-to-day writing code in most projects. So, what I'm looking for is different things. I want a model that follows instructions better. A smarter model doesn't necessarily do a better job of doing what I tell it to do. And these benchmarks tend to be so simple and so limited in scope that they don't really measure what happens when you give a model access to too many things. Like when you do a math contest with a model, there isn't other code in the repo that it can touch and do stupid things with. If you give Gemini 3 Pro access to a code base and tell it to fix one small thing, it's going to touch a lot of random [ __ ] These benchmarks aren't measuring how well the model works. They're measuring how smart the model is measured to be based on that test. So, I don't want models that are smarter much anymore. I want models that are faster and better at doing what I tell them to do without getting distracted. And we're starting to make improvements in this direction. Something like composer one in cursor is a great example. It's slightly dumber than I would like for that type of model to be, but man, if composer one was like 20% more intelligent, I wouldn't really touch anything else. It's just so much faster. It's annoying. So, I was talking about moving this project over to use effect cuz it would be a lot less annoying to deal with if it was using effect. So, let's do that. So, I want to show the difference in how it feels using opus versus composer versus any of these other things. Multiple models composer opus 5.2 effect is not easy and most models aren't good at it yet. So this will be a fun test. Tell it to make work trees on Arena V2. And now I will show you just how much absurdly faster Composer is than everything else. Made the to-dos. It's gunning [ __ ] out. Opus is suggesting that we do plan mode instead. We're going to skip that suggestion. Compose is already two to-dos in. Opus is finally making a to-do list. GPD 5.2 is probably not going to make a to-do list because for whatever reason, the GPT models are really bad at to-do list creation and management. It's funny, they're great at following instructions. They're not great at making their own instructions. Composer is getting close now. It's over three done. You can see how slow 5.2 is when it's generating. I have a video I recorded of a screen recording of like how bad it can get. Inline, does it have been suppressed for recent changes because there are too many display protection or don't show again? I'll just hide that for now. Is that a,58? Is it done? It's almost done. Yeah, it is. Cool. Composer one is done. Was that immediate? No. Was it enough time for me to leave my desk or get distracted? No. Is enough time for me to reach over and have a sip of my drink? Opus is two of five to-dos done. And GBD 5.2 has edited 40 lines of what will probably be a thousand plus line change. This is my issue. I suspect the other models are going to take another minimum 10 minutes on this, probably more. Opus might be a little bit faster than that, but I'd be very surprised if 5.2 finishes this in less than half an hour. There is a Zed employee in our chat right now. And funny enough, Michaela, I have a good friend, you probably know who he is, his name is Benny, helps run my channel, who was really enjoying Zed and has said verbatim he is not moving to it for two reasons. One is that he likes the tab complete more in cursor and two is he likes composer too much in cursor. So you have users that are much much more locked in than I am at this point that would be on zed but the model is the thing that's sticky now which is kind of crazy. Yeah, it could happen. Like if y'all can find a way or if some other third party can come in and make something that is that fast and that good for other harnesses to use, something magical could happen cuz composer is one of those like oh [ __ ] I can just let this do the thing moments. And honestly for most tasks when I'm just working if I don't think it needs a planning step I will start with just running it on composer and if it gets it wrong I don't care. I just throw it away cuz it took seconds to do and then rerun it with a smarter model and leave my desk. Cursor hired the Super Maven guys, right? Yep. That's how this all happened. It was a very, very good hire for them. They acquired Super Maven, got me more equity in Cursor, by the way. Yeah, as always, I have Cursor equity. Doesn't stop me from complaining about them. I do more than most people do. Cursor has its problems, but I am an investor because I have a lot of faith in the team and their acquisition of Super Maven to lead Tab Complete and now help build models like Composer. The reason it's so good makes a lot of sense. Jacob is a god. And as I've been talking this whole time, they're still running. They will probably continue running, but I don't want to waste any more of your time because I've already done a lot. You get the idea. 5.2 is really smart, but it's not necessarily good. And I've been thinking about this way too much since I started using the more and I wanted to get this off my chest. Normally, I have a better outro in mind, but I remember that I tweeted this yesterday and I think it summarized my thoughts better than I could ever now. GBD 5.2 is the smarter model. Opus 4.5 is the better model. And Gemini 3 Pro is indeed a model. That's all I got for you guys. And until next time, peace nerds.\n"

ERROR: type should be string, got "https://youtu.be/CtMk0GuQ7cc\n\nGPT-5.2 is the best model ever made*\n\nWhere's the guy with the counter? Because there's a new best new model. Yes, really. GPT 5.2 just dropped and it is really, really good. Which is why it is uh Wait, what? Skatebench shows it as a huge regression. What's going on? I thought this was the best new model. Well, in many ways it is, but in a handful of important ones, it isn't. And I haven't seen many covering this in detail. Sure, it's better at code and tool calls, and yeah, it's crushing ARC AGI, but there is a depth to using these models that isn't really being shown in a lot of the coverage I'm seeing. And I want to break down what it's actually like to use because I've been lucky enough to be using it for the last week or so. Very thankful to OpenAI for giving me early access. That said, no money has exchanged hands. The only person paying me is today's sponsor. We need to be realistic about how much easier our jobs are now. AI has made everything from feature additions to bug fixes simpler than it's ever been. Navigating your codebase, finding the right file, and making the changes you need is really easy, but there are a few things that are still hard and also really scary if you get them wrong. The two that I think of the most are authentication and authorization of users and payment processing. Getting these things wrong can be disastrous. And even if you think you got them right, something might prop up later on that makes you regret building it yourself entirely. Not only have I done this myself incorrectly, I've even written detailed documentation on how to do it right, that barely feels necessary anymore because of today's sponsor, Clerk. If I was building a new app today, this is what I would choose both for authentication and authorization of my users and for payment processing because it is the best and simplest way to do both. Period. And I've used every single option. They aren't paying me to say that. They're paying me to mention them. I'm telling you, as a person who's went through all of the options, especially on the payment side, Clerk has found this perfect integration of things as well as developer experience that makes it easy to build a secure, reliable application for everything from signing in to signing up. There's something beautiful about seeing a component like this for something as annoying as payment processing and managing what users access to what features. You want to protect a feature so that you only have access if you're paying for the team plan. It's this easy. protect feature equals team access. And now, as long as you have that set on the server side when you define these things in their dashboard, you're good to go. You could even check if they have a thing like the bronze plan. And if they don't, you can return an error. Do you understand how annoying these things are to do traditionally and how hilariously easy they are to do with Clerk? They even have a pricing table component. you know, the fancy like comparison of the different tiers that is fully integrated, has the Stripe pop overview built into it, and once they subscribe, it's immediately linked to that user's account. It's so much easier than the other solutions. I genuinely wish this existed when I started T3 Chat. It would have made our lives so much easier. Like, I'm talking not weeks, but months of work. If you're ready to have real users and make real money, look no further. Check them out now at soyb.link/clerk. There's a lot to talk about with this new model, but I want to resolve the click baiting I just did with that whole benchmark thing. There are a lot of layers to this that we'll get to later, but I want to just show the raw numbers that I just ran. And I've spent a lot of money running this test over and over again to be sure of my results. I even went back and forth with some of the people on the research side trying to figure out what caused this regression. My current theory is that the new model is not as good at three-dimensional reasoning because my benchmark is about skateboard tricks. I give a description of a trick and I expect the model to tell me what the name of that trick is. And previously the highest score I ever got was with GBT5. Not even high. The default setting would get 97% on this benchmark. 03 Pro would get 96%. So like GBT5 was incredible at the spatial reasoning necessary to describe a skateboard trick accurately. When I first ran this benchmark with the new GPT 5.2 model, I got this 4% number and I was like, \"What the [ __ ] went wrong here?\" I did a lot of back and forth with the team and we realized that my harness due to some changes in how the GPT 5.2 model defaults, which is it defaults to no reasoning, the results ended up being lower. So like, okay, cool. I'll run it on high and on the new extra high reasoning options, which ended up being comically more expensive to run this bench on. Again, compared to GPD5 default, it was about6 cents per request. And GBT 5.2x high was about 2.5 cents per request. Yes, it was about five times more expensive. And the reward for that is a 20% regression in performance. Is naming skateboard tricks the best use case for LLMs? Probably not. But I was so pumped to see GPT5 understand spatial reasoning well enough to do this. And this is a massive regression. Like absurdly so. It's now tied with models like Grock 4. What what went wrong there? But it seems like my benchmark is novel in this case because almost every other bench is showing a very different story like GBT 5.2 thinking getting a 70.9% on GDP val versus 5.1 thinking and five getting a 38.8%. SW Bench saw a big bump at 55.6%. SWB verified saw a first time score for the OpenAI team of an 80%. GBQA Diamond was a nice boost. AME was now 100% with no tools which is really good. It means it can solve math problems without using math tools. Impressive, but there's a lot more to these stories. That all said, the Arc AGI scores are actually absurd. From ARC prize themselves, a year ago, we verified a preview of an unreleased version of OpenAI's 03 high that scored an 88% on ARC AGI1 at an estimated $4.5,000 per task. Pretty absurd how expensive it was for them to run that special version of 03 for this, but uh it was able to get a high score. And now the extra high option on GBT 5.2 Pro, which is also notably not GBT 5.2, scores a 90.5% and it's only $1164 per task. That's a 390x efficiency improvement in one year. ArcGI is a wild benchmark. I have covered it many different times. If you're not familiar, go to their website and try it yourself because it's a thing that's easy for humans and nearly impossible for LLMs. It is no longer nearly impossible for these LLMs. It's actually really, really impressive. They made a V2 of this leaderboard that was meant to be actually impossible for LLMs to solve, thinking it would be like the final benchmark ever. Yet, here we are with more and more models getting high scores on it. We just had Gemini 3 Pro have a groundbreaking 30% and then the Gemini 3 Pro deep research, whatever the hell they call their like heavy version getting all the way into the 40s. And now GBD 5.2 Pro High is scoring way, way higher here. 54.2% 2% for $15.72 per task. But then my favorite part, due to API timeouts, we were unable to reliably verify GBT 5.2 Pro XH high on RKGI2. Turns out these models are still timing out all over the place. It's kind of insane how long 5.2 Pro will run for. I didn't even put it in my most recent run of Skate Bench because it took like up to 10 minutes per request at times for naming a skate trick that I can do in 5 seconds, not even. It's kind of absurd how hard the new models will go. And even 5.2 extra high was able to take 240 seconds per request. That's 4 minutes per request sometimes for naming a skate trick. The first example they open with is the economically valuable tasks section which includes GDP val which is a benchmark they made for evaluating how well different models behave against known knowledge work tasks that you have to have like a degree or something for. They take actual examples of work and ask the model to perform against it and see if it performs at or above a human expert level. And 5.2 thinking beats or ties the top industry professionals on 70.9% of the comparisons on the GDP file knowledge work tasks according to expert human judges. That's the other interesting piece here is this benchmark needs to be judged by humans. So I can't just run it on my laptop. And it's a massive jump from GPT5 which was a 38.8% to 70.9 for 5.2 thinking and 74.1 for 5.2. 2 Pro. And there's no chart crimes happening here either, which is nice. Don't worry, we'll have plenty of chart crimes later. While reviewing one especially good output, one GDP Val judge commented, \"It is an exciting and noticeable leap in output quality. It appears to have been done by a professional company with staff and has a surprisingly well-designed layout and advice on both deliverables, though with one we still have some minor errors to correct. Yeah, apparently very impressive. But on the topic of design, we should see how it handles design work. I have a stock Dex.js project that I haven't made any changes to. We're gonna do my favorite, the image gen bench where I tell it to make an image generation studio mock and uh apparently cursor is breaking and I am now past my usage. I'm on the $20 plan so the fact that I'm just hitting that now is really cool. Obvious bias. I am an investor in cursor so account for that. This is also one of my first times using 5.2 incursor because I didn't do the thing where I set up the open AI like API manually incursor. I have told this to the team many times and I will tell them again right now. Hi friends at cursor. I should be able to set a custom OpenAI API endpoint without having it break every other model in your app. It's very annoying that if I go set an OpenAI API endpoint that I can no longer use Opus or Composer or any of the other models that I like to use in Cursor. It's quite obnoxious. I know I'm unique that I get early access and I have a need to test these things. But it basically makes the feature of the checkbox to put in a custom API endpoint entirely useless to me. The majority of that was uncut, but I'll let you know roughly how much time this generation takes. It isn't very fast. These models are still not very fast. It's the one big pain point I have had with the GPT5 series now that I'm experiencing Opus 4.5 and especially the composer model and cursor way more. It is obnoxious how slow the GPT5 series is, especially if you take advantage of some of the fancy pro model stuff, which like I've had the Pro models take 30 to 50 minutes to respond to things. It's awesome that they can go out and do work and come back with a correct answer so reliably, but god damn, those models are slow. I thought this was going to be fast enough that I could talk over it and give you guys like a realistic feeling for how long it would take. I overestimated just how fast it could be. So, we'll come back to GBT 5.2 and cursor in a moment. Almost forgot to mention, as y'all probably expect, we threw all of the new models in T3 chat. So, if you want access to all of the versions, including the new reasoning one that's a little restricted on the free tier on chat GPT, for eight bucks a month, you can use all these here. If you want your first month for only $1, use code 5.2 at checkout. Anyways, back to whatever the heck I was just talking about. GBD 5.2 Thinking sets a new state-of-the-art of 55.6% on SWEBench Pro, a rigorous evaluation of real world software engineering. Also note, this is GPT 5.2 thinking, not GPT 5.2 to codeex which will probably come in the near future. It does seem like they're finally realizing how bad the terminology for the models is and they might not do the whole codeex thing again. I I hope that they drop the overuse of that word. It just they should have called codec cli GP terminal. They should have had the model be the code edition or something, but just calling everything codeex is obnoxious. Regardless, this SWB verified bench only tests Python, and SBench Pro tests four languages and aims to be more contamination resistant, challenging, diverse, and industrially relevant. In the Pro version, they got the new highest score on the extra high version of 56%. But very interestingly, the 5.1 Codeex version when given max extra high, so like allowed to just use as much context as it wants, it actually performed slightly worse than the normal high version does. the the way rein are being used is increasingly weird, especially with the new defaults and the behaviors we're seeing here. Like GPT5 on my benchmark got a 97% doing 600 tokens per request average. GBD5.2 extra high got a 79% an 18 point regression while using over three times the number of tokens. Interesting. For everyday professional use, this translates into a model that can more reliably debug production code, implement feature requests, refactor large code bases, and ship fixes end to end with less manual intervention. They also said it's better at front end. So, we'll look at that in a sec cuz the generation did finish. But, I want to first talk a little bit about my experience using this to generate my own tests. I maintain a handful of different benchmarks and evaluate bench in particular. had some changes I wanted to make around how the caching worked because some of these new models would hit errors and I would need to rerun it, but the cached errors would keep the model from rerunning. That plus all the weird things about usage and I really really wanted to start tracking token utilization so I could include that in my coverage here as well as these awful run times. So I had to make some changes for that. So I wrote a prompt make the following changes to this project. One, cache should also include token counts and durations. Two, errors should not be cached. So a rerun should re-trigger any jobs that errored out. And three, you should show the average token usage in the table in the CLI view. Think this is a pretty clear set of things to do. I had this come through with composer, with Opus, and with GPT 5.2. The composer version came back almost instantaneously with changes that were mostly good. I didn't like how it was doing the cost calculation and the token calculation though. It was relying on a weird sub field instead of just using what the SDK gave me. So I rejected that. I don't know why it has a thumb there cuz I picked opus, not composer. The opus version seemed like it was really good and it was half as much code as the others until I looked at it more closely and realized that it wasn't actually using the cache for the token counts or for the times like it was expected to. So that was incredibly frustrating. So I told it to start caching these things and it did, but it didn't use the cache results for [ __ ] anything. But I had to remind it later on with a follow-up message because I I didn't notice this mistake until I reran. was like, \"Wait, you just screwed this up.\" Also, it was tracing input tokens, which is the same for every single run, so it doesn't matter. So, I told it to drop that, but it still entirely forgot to restore durations when loading from cache. So, I had to do two follow-up prompts with Opus 4.5 to get it working how I expected. That all said, I have not actually tried the chatbt 3.2 version. I gave the code a quick look and it seemed fine, but that's not like me. I want to actually test it. Let's do that. Copy work path. Hey uh friends at cursor who are probably watching this. At no point did I do anything that would result in these changes from happening. In fact, the thing that is being deleted here is the correct version that works. And the thing that's here, the green, the proposal, is entirely wrong and incorrect. There's no review anything here. I have no idea why this is here at all. It's wrong and bad and broken. It's a reversion of something that hasn't existed in this codebase forever, if ever. Why is this here? What the [ __ ] wrong with the UI? I don't know what's went wrong with the review mode, but it's getting really, really egregious lately. But now that I have fixed the package JSON, to be very clear, this is not broken because of GBT 5.2. This is broken because of cursor. Let's run the test. And look at that. The one from GPT 5.2 did everything right first shot. So the the reason I just tested all of that is I really wanted to emphasize the difference in I guess vibe when I'm using GPT models, especially 5.2, Compared to models like Opus and Composer, GPT5 is just the series that follows instructions the best. That's the best I can put it. The Opus models will roughly finish the task you give them, and they're very smart. Their ability to figure out what's wrong, debug, and push is incredible. Opus models can turn through insane tasks for long amounts of time and actually generate something that works. But the GPT models will do what you [ __ ] tell them to do. I had to do two follow-ups with Opus to get it to behave. It wrote slightly better code that is closer to what I would have written, but I had to tell it what to do with multiple follow-ups. GPT5 just did it first shot. That's the difference. I have had a much better time with the GPT models for that. But what's extra funny is GPD 5.2 still took longer than me running Opus, testing the results, realizing there were things wrong, reprompting it, and getting a new result. So, if you want a model that does what you say and you're willing to wait, 5.2 is incredible. If you want a model that is really really smart, possibly even smarter than you for the things that you do, but loves to just go off on its own little tangents that you have to like grab your like harness and pull it back in. Opus is great for that. And even then, Opus 4.5 is a significant improvement in instruction following in my opinion from how I've experienced previous enthropic models. They seem too happy to go change code and not happy enough to do what you tell them to. Enough of all this. I want to see my image generation studio. So, let's see how this one came out. Not bad at all. They've tuned the gradient stuff a little bit more, which is cool to see. It looks really solid. I have been impressed with Opus' frontend abilities. Gemini 3 Pro is still also really, really good at it. I know a lot of people that hate the Gemini 3 Pro model for everything other than Tailwind. Funny enough, one of those people is my channel manager, Ben Davis. He wrote his thoughts on 5.2. Since it reasons less, it feels way faster than GPD 5 and 5.1 did. So, if the speed of five and 5.1 is bad for you, but not terrible, it's worth giving 5.2 a shot because the difference here simply in how much fewer reasoning tokens it uses might be enough of a jump. Here's another example somebody posted of a 5.2 UI gen, and it looks really good. I love how it did the gradient for this section here. It really loves this grid pattern like that it puts behind everything. But honestly, this is a great UI for a model to [ __ ] out. I'm impressed. The models all generate UI looking the same thing is mostly over. They use gradients in a similarish way now, but man, the quality of the UIs these models are generating compared to even just like 6 months ago is hilarious. So yeah, pretty cool to see. They also claim it's way better at 3D visualization stuff. So like if you're using React 3 Fiber or 3JS, even stuff like Phaser, they they have suggested it'll handle those things better. Apparently it generated this and it does look really really nice. Like this is a cool little visualizer they made. Had it make a holiday card builder with way too much animation. And you can see that gradient pattern, the pink in the top left and the blue in the bottom right. Pink top left, blue bottom right. That's the new gradient pattern all of these AI things love to use. This looks solid as well. Typeer. This is kind of fun. Look at that. It made an actual decent game. Who would have thought? Yeah, I still personally have found for 3D stuff in particular, Gemini 3 Pro feels a decent bit ahead, which on my bench here, Gemini 3 Pro placed relatively high at 86%. Which is worse than GPT5 did, around the same as 5.1 and worse than 5.2. I don't know what happened where it's a regression in 3D stuff, but I'm not the only one saying it. I've talked to a few friends who are using it heavily for 3D like 3JS type stuff and they've also noticed the regression. But if it's doing 3D in a more 2D way, like a really really fancy version of the hexagon ball test, it looks beautiful. Its ability to make good-looking things in 3D space seems on point, but its ability to understand 3D space for my experience has not been that great. It almost seems like they overindexed on 2D space because of ArcGI and through doing that kind of broke their 3D understanding. Even though again, Flavio disagrees and says it's surprisingly good at 3D and physics. I just realized I almost forgot about one of the most important changes with GBT 5.2. They up the price. This is the first price increase we've had for models in a minute because they had dropped the price recently. It was 125 in and 10 out for GBT 5 and 5.1, which really surprised me. Now it's up a bit to 175 in and 14 out. The theory I'm seeing many say is that they were intentionally releasing smaller distilled versions of the model before, and 5.2 to is their like coming back home type move where they're finally putting out the full size real version which again I have proven is very unlikely with skate bench as silly as this bench is these types of gaps mean something it means that this model is actually worse at some things than the previous versions were which is a very hard thing to do if GBD5 is a distillation of this version it is also worth noting that for many tasks the model ends up being cheaper because it's so much more efficient with its reasoning tokens according to openai on multiple aentic evals we found that despite GPD 5.2's greater cost per token. The cost of attaining a given level of quality ended up less expensive due to GBT 5.2's greater token efficiency. To be fair, it says cost of attaining a given level of quality. So that means to get a certain score, it's cheaper, but if you want the best score, it's still more expensive. Yeah. And then there's the 5.2 Pro pricing, which makes me feel sick. $21 per mill in and 168 per mill out. A new groundbreakingly high price for a model. bit more of the facts from OpenAI. It's way better at not hallucinating, which is if you've been using Gemini 3 Pro, you're back in like the 2024 era of hallucinations. I'm going to do a whole video about the weird quirks of Gemini 3 Pro soon, but going back to a GPT model and feeling the difference of how much less likely it is to hallucinate is insane. It's just it doesn't lie anywhere near as badly. And it's also way better at handling really long context reasoning when you're doing long benchmarks and finding things in the like the needle and the haystack tests. This is an insane benchmark. To still have really high recall, 98% accuracy at 256k tokens is a massive achievement. Massive achievement. Gro 4 on this same benchmark and 4.1 fast as well was in the like 30 percentile. Huge, huge win. With eight needles, it still drops down to like 70%. But that's way better than before in the 30s. They've also had massive improvements on vision, massive improvements on tool calling. Okay, not that massive from 5.1, but it's getting higher and higher, hitting like the 98 99% range for a lot of the different benches we have for how accurately you can call tools. And I haven't seen it do a single malformed tool call in cursor yet, which is a a huge win considering how aggressively it used to do that. Better at science and math. It crushed dark AGI. And one other important piece because like the model range is getting complex where we have 5.2 thinking that has different reasoning levels between minimal, low, medium, high, and the new extra high as well as 5.2 Pro separately. And then 5.2 instant. 5.2 instant isn't actually a separate model. It is the no reasoning version of 5.2 thinking where you set reasoning to none. And it actually is much much better. They've been pushing this weirdly hard. It seems like they're really hyped on 5.2 without reasoning. And a couple of my friends, including of course Ben, have been saying it's a huge improvement as well. Early testers particularly noted clear explanations that surface key information up front. Cool. I haven't experienced that yet, but it sounds likely. They continue working on the mental health problems that these models can cause. Good to see improvements here. Awesome. And I have two last pieces I want to dive in on. Matt Schumer posts awesome reviews of the new models. He's similar to me where he gets early access. We're in a lot of those early calls and groups together and his thoughts are really good. Both of these will be linked in the description, but I just want to go over the TLDDRs because it's very aligned with my experience. By which way to thinking is a meaningful step forward in instruction following and willingness to attempt hard tasks. Code generation is a lot better than 5.1. It's more capable, more autonomous, more careful, and willing to write a lot more code. Yes, vision and long context are much improved, especially understanding position and images and working with huge code bases. Haven't played with this enough yet, but I trust him. Speed is the main downside. Yes, yes, yes. Couldn't agree more. In my experience, the thinking mode is very slow for most questions, though other testers reported mixed results. I almost never use instant. Yeah. And 5.2 Pro is insanely better for deep reasoning, but it's even slower. And every so often it will think forever and still fail. Had that a couple times, too, where it's like 30 minutes in it just stops. It It feels like a bug. I don't know if it's going to do that over the API or not, cuz they finally give us pro over API. According to Matt, in the CODC CLI, 5.2 Pro is the closest thing he's felt to a proquality coding model in a CLI. But the extra high reasoning mode that gets it there makes it take forever. He's been pushing over there to get the pro models into his editors for a while. So, I'm very happy for him. He bitched about this a lot with his 5.1 Pro review, and I feel him because the Pro models are exceptional. They're just really, really slow. So, what does he have to say about 5.2 Pro? Undoubtedly the world's best model. I can't live without it. That's a lot nicer than what he said about 5.2. The most capable model available today, but it's slow, so it's not for everything. It only exists inside of chat GPT, not codecs nor the API, which is so frustrating. It is now on the API, so thank them for that. Expect extremely long thinking times on hard tasks. It's willing to think for longer than previous models, which makes a huge difference for the hardest ones. And it improves reliability for everything else. It has an uncanny ability to infer missing context that I didn't provide in the prompt. Not just obvious things, but constraints I hadn't even realized were important myself until it pointed them out. I have even felt this a bit with 5.1 Pro. So for him to call it out here, that's real. Every so often, it will think for a long time and still make a big mistake, wasting a lot of my time. Yeah, that that's because it takes 30 minutes to an hour. When it does fail, it feels so much worse. Prompting matters more than ever, so be explicit. Add constraints and refine prompts before you send them. After using Pro for 2 weeks, I can't live without it. It's my go-to for everything I do that requires deep thinking, research, or coding, or almost any prompt I run that doesn't require an instant answer. Yeah. And this is somebody who was just really, really hyped on Opus 4.5 a few weeks ago. So again, I trust Matt dearly. Very good reviews. Highly recommend reading those in detail if you want more information. But I think I've said all I have to say here. Seems like a very good model. I want to play with it more myself. I'm curious how y'all feel though. Is 5.2 overhyped or is it too slow or is it actually incredible? I'm curious how y'all feel. And until next time, peace nerds.\n"

ERROR: type should be string, got "https://youtu.be/KAmQTmooLGQ\n\nChina is winning the AI race\n\nIf you look at the current top models, they're all from America, Google, Anthropic, and OpenAI. We are clearly\nwinning the AI race until you zoom out a little bit. Then you see a lot of these\nblue bars appearing in the chart. Those blue bars are for openweight models. And if we look at the top three, Kimmy,\nDeepseek, and Mini Maxm2, you realize that China's winning the openweight race\nand by quite a bit. The first model from the US to appear here is GPTO OSS120B.\nAnd as a person who's used that model quite a bit, it's rough. It might score well on intelligence charts, but its\nability to reliably call tools and be used in your workflows is nothing in comparison to what I've experienced with\nKimmy, with Miniax, and now with Deep Seek V3.2. Huge gap between those. And if we want\nto look at the European introductions like uh Mestral Large 3, which just dropped and kind of inspired this video,\nthey're barely even on the chart. Things are rough. Oh, and almost forgot, much\nlike they seem to have, Llama 4 all the way at the end here. There were supposed\nto be three versions of Llama 4. If you remember, it was supposed to be Scout, Maverick, and I forgot the name of the larger one because they never put out\nthe larger one because they all suck so bad. There's a 20 bill per pram model from OpenAI that beats out Mestral and\nthere's lots of 15 bill ones that do too. It's rough out there. But the thing I really want to focus on today is the\nopen weight wars and why China seems like they will be winning them for the foreseeable future. It's kind of crazy\nthat when you only talk about models that weights are downloadable and usable, all of a sudden America gets\nwiped off the chart. There's a lot of reasons for this and I can't wait to talk about them, but since openw weight\nmodels don't pay the bills, we're going to do a quick sponsor break first. Here's a hard question for you. How do you know if an engineer is actually\ngood? It's really hard to do. You might be able to look at their HTML shirt and make some assumptions, but when you're\ndoing an interview, especially when you're reading someone's resume, how do you know they're actually good and not just AI generating some slop that you're\ngoing through a huge pile of as you fill this role? It's never been more annoying to hire good engineers. I feel like\nthere's fewer of them in the pile and the pile's never been bigger. If you're tired of trying to find the needle in the hay stack and get a good engineer to\nwork for you finally, you got to check out today's sponsor, G2I. These guys are without question the best way to hire\ngood engineers fast. They have over 8,000 of them ready to go in their incredible network. These aren't people\nwho are fresh out of college. These are real experienced engineers that have worked at big fang companies and small\nstartups alike. know how to use all the tools you need, already are familiar with fancy AI development stuff, so\nthey're not going to be slow. Whether you want a couple junior engineers to kickstart a new project, or a lead that\ncan dig you out of tech debt hell, they have you covered. You create a shared Slack channel with them. They\neffectively are operating like your recruiting team. You give them a handful of questions to ask the engineers. They ask the engineers and record actual\nvideo responses from them so you know what the person's actually like. You go through them, figure out the ones you\nwant. They'll then go do a technical interview that they've speced out so you don't have to worry. Record it, send you\nthe results, and once they've gone through all that, you can review it, figure out who you think fits best, and\nthen you can hire them. They're also just super generous to work with. I've referred a lot of people on a personal\nlevel, like a lot of the YC startups I work with, and every single one has had an incredible experience with G2I. Stop\nhiring the old way and stop wasting your time. Get good engineers fast at soyv.link/g2i.\nBefore we can discuss why China's winning so hard, it's important to understand what is an openw weight\nmodel. The point of openweight models is somewhat similar to open-source models where you're giving out a significant\nportion of how the thing works. There's a big difference between openweight and open source though. With open source,\nthe code that is used to create the thing people experience is exposed. With Linux, for example, the actual thing you\ndownload isn't the code when you're using Linux. What you download is the binary that's compiled by the code. The\ncode is the input that results in the output that you are downloading and using. As such, I've seen a lot of\npeople complain that openw weight models aren't open- source because you can't recreate that binary yourself. And I\ndon't really agree. Obviously, it would be cool if we had all of the training data and everything else that went into\nhow the models were made. But the reason open source is valuable is because you can reproduce the actual output. I can\ntake the code and on my computer compile it and get a result. There are almost no\nconsumers, almost no developers who have all of the things that are necessary to spend the millions of dollars on really\nhard to get infrastructure to turn that data into a model. There is no real\nreason to expose that and there's a ton of risk and liability if you expose all of the data that's used in your training\nand every other company's just going to take that data, throw it into their data sets and suddenly be able to beat you in\nall of the things you're good at. Depending on how you cut the lines and think about it, I would argue the data\nis almost the equivalent of the engineers in this case, not the equivalent of the source code. People think about source code very\nspecifically because you use the source code to compile the thing that you want and as such we should have the data if\nwe want to call these models open. I would argue we're just drawing our lines a bit differently. So in open source a\ndeveloper creates source code that compiles into a binary that users can use. In open weights data is used to\ntrain weights that result in generated tokens. If you think of it this way where you're\ndrawing the lines here, you say like the researchers\ncollect the data that creates the weights that generate tokens. I would\nsee why you would call this not open. But I don't think that is quite how it works. I think of it this way where the\ndata is similar to the developer in the case of the weights because it creates the thing that we can use to generate\nthe thing we actually want which in this case is the binary it's the Linux installable the actual kernel and in our\ncase with models it's the tokens that we get from using the model. So the open weights mean you have all of the pieces\nyou need to run the model and generate results with it yourself. The weights are the collection of parameters that\nare all mapped to and point to each other. So when you give it some text, it can guess what the best next token would\nbe based on the text you give it and this giant hundreds of gigabytes pile of vectors and data that it has collapsed\ninto this model that it can use to generate the next token as predictably and reliably as possible. I think it's\nreally cool that open weight has gone as far as it has. And I think it's really convenient that openweight models can\nuse the same licenses that open source code can. I already see people disagreeing in chat. I don't care. The\nweights are not the binary. The weights are a thing that can be reused and\nmodified in very useful ways. The nature of how baked these things are. Like\ncompiling code costs pennies and can be done on most computers. Turning data\ninto weights isn't even a deterministic process. And I know there's a lot of debates around this. I know there's a\nlot of things that like Richard Stallman's going to disagree with me here on. I don't really care. This all\ncomes down to whether you put this here or here. And I'm not one to when\nwe get something as cool as openweight models. There's only one lab I know of that actually puts out the data and it's\nAllen Allen Institute. They were funded by Paul Allen from Microsoft as an attempt to do truly open AI research in\nthe US. And their models aren't just open weight models. Their models also have the data exposed too. So you could\nhypothetically retrain the model on the data yourself. None of it's deterministic enough that you'll get the exact same weights. But yeah, it's\nexists. If you're wondering where this falls in the charts, right next to llama, not great. So, it's\ncool that we do have a fully open lab that is sharing the data and everything that is based in the US, but they're not\nreally competitive. Just wanted to call that one out quick. The harsh reality is if we use the strict open- source\ndefinition that currently exists for code, there will never be a model that meets the definition of open source. And\nI agree, there probably won't be. And we shouldn't use the term open- source to describe models. Open weight is still a\nvery cool and useful thing. So with an open weight model, the value you get out of it is I can take those weights and\nrun them on my own hardware or look at different providers that are hosting them as well. If we go to something like\nopen router and take a look at a Gemini model like Gemini 3 Pro preview, you can\nuse it in two places, Google Vertex and Google AI Studio because the weights for this model have never left Google's\ncampus. The weights that you use to run these models and generate these results are exclusively provided through\nGoogle's own infrastructure because they want to sell you the API, not the model. And since Google has their own\ninfrastructure, they don't let other companies have access to this except for Apple privately potentially with a\nreally really big pay deal of like a billion plus dollars to get the weights privately that they can use for some\nSiri stuff. If you look at something like OpenAI's GPT 5.1, your options are\nOpenAI. Some of these models are also available on Azure too, but that's it due to the OpenAI Microsoft partnership.\nLet's compare that to Deep Seek 3.2 EXP. We got Deep Infra, Novita, Shoots,\nSilicon Flow, and Atlas Cloud. Let's look at Kimmy K2. Kimmy K2, Shoots,\nSilicon Flow, Novita, Deep Infra, Parasel, Bite Plus, plus seven more. Moonshot. These are people who actually\nmade the model. They are the eighth option in this list. Fireworks, Atlas cloud, base 10 together, Grock, and\nTurbo from Moonshot.AI. Also notice the Turbo option for Moonshot, which costs\n$8 per million out, is less than half the speed of Gro's solution here, and 8x\nthe latency, too. Kind of nuts. The open weight models allow for various providers to offer them, which allows\nfor a different level of competition across infrastructure solutions. That is really, really cool. But it does also\nmean that the official infrastructure in this case for Moonshot isn't really a great option. Moonshot charges 60 cents\nper mill in and 250 per mill out for under 20 tokens per second. Grock\ncharges a dollar per mill in and $3 per mill out. So slightly more for 356\ntokens per second. That is more than a 10x increase in throughput for a very minor bump in cost. This is the\ndifference. When you have this type of competition, the value prop of your own infrastructure goes down, which makes it\na lot harder for a company like Moonshot to make money on the Kimmy models, even though they are fourth on the artificial\nintelligence chart. Google is a trillion dollar company. Enthropic is a multi-billion dollar company,\npotentially worth trillions someday. OpenAI is already worth half a trillion dollars. Kimmy K2 Thinking by Moonshot\nis a small company in China that isn't making real revenue yet. Do you know what's really funny though? Do you know\nwhich of these four companies has been the kindest to work with for me as a creator? Moonshot. They've been trying\nreally hard for me to give them a mailing address so they can ship me a care package. They've been awesome to work with. They always hit me up early.\nThey offer me free inference for any tests I want to do. They constantly send me useful resources about the things I'm\ntalking about. Moonshot's been a genuinely awesome company to work with and they even shout out their competitors when they have big launches.\nLike when Zai had a big release, they immediately went and supported them. They're a very good faith player,\nweirdly. So, Deepseek is very similar in this regard. Not in the com sense. Like, I've never heard from anybody at\nDeepseek. By the way, Deep Seek guys, if you want to hit me up, I'd love to chat. Very, very big fan of what you did. I would never have built T3 chat if it\nwasn't for Deep Seek V3 at the end of last year. I'm so impressed with the work that DeepSeek has been doing for a while now and their research is\nincredible. They put out 12 papers last year that were so far ahead of where everyone else was. And the discoveries\nthat made FP8 training much more reliable resulted in every lab fundamentally changing how they did\ntraining. You could argue that a large portion of the speed that AIF accelerated this year came from the\nresearch DeepS put out for free last year. And yes, I have also talked to the ZI guys. They've been great. They've\nbeen really, really awesome. It's crazy how good at comms the Chinese labs have been with me at the very least. Openai\nhas been really good. Google's up and down. Enthropic is interesting. But my\nexperience with the Chinese labs has been really good as a journalist, so to speak, covering these things publicly.\nBut none of that answers the question, why do open weight? Why are these companies releasing these models in a\nway that they make no money off them? Like the real winner whenever DeepSeek drops isn't DeepSeek. It's companies\nlike Grock and Together and all these like cloud info providers that will host them for us. We currently don't have\nDeepseek version 3.2 like the final official version on T3 chat yet because none of the providers are doing it well\nenough just yet. I would even argue being open weight makes things much harder for the labs even outside of the\ncosts since Kimmy K2 is available for anyone to host themselves. different\nhosts aren't necessarily hosting it properly in the quality of certain behaviors like tool calls might go down\nmeaningfully depending on which host you're using. Kimmy actually went as far as creating the vendor verifier where they rank all of the companies hosting\ntheir models based on how reliably they do tool calling. These are all of the\ncompanies that they say are hitting over 73%. And if we scroll down, you'll see others\nnot performing quite as well. It's cool that they're doing better now because previously the gap was a lot bigger. But\nby creating this bench and making this data public, they incentivize the hosts to fix their and also gave them the\ntool called eval python file that they can run against their own infra and find the bugs and fix them. Doing this type\nof thing is really really annoying but they are doing it because otherwise the reputation of these models will be hurt\nas a result of other labs and other hosts not hosting these things properly.\nIt's a small thing, but I also love they're using UV. Like, these guys get what US developers are expecting. So,\nit's clear that doing open weight is harder. It makes it so you make way less money. Why the hell are they doing it?\nTo be frank, nobody would trust them otherwise. If you're using a Chinese model and it's\nbeing hosted in China, all the data in and out is now at a real risk, like a\nvery legitimate risk. A lot of these companies have Chinese government hands in them. There is no security team in\nthe US that would approve of you using a Chinese model from Chinese infrastructure. And open weights allow\nthem to be relevant in the space right now. The fact that I'm legitimately considering doing more work with Chinese\nmodels as an American shows that the openweight strategy is working for them because it's the only way they can hold\nany mind share in the US. There's even been attempts to ban the use of Chinese models in the US. When Deep Seek R1\nfirst dropped, there was a huge freak out about that. And the government here was actually considering passing\nlegislation that would make it illegal to download the weights. Wild. Insane. I have files on my computers that would\nsuddenly become illegal if that crazy proposal was to actually go through. Absurd. So, this is like seriously the\nonly way these Chinese labs will be taken seriously. And this goes a lot further than language models, too. It's\nthe same deal with a lot of their image and video generation models as well. These models are not something that\nyou'd want to run out of China, especially because they have restrictions on what GPUs they're even\nallowed to have access to. So, you might not be able to run some of these models they're making on infra and the infra\nthey have is limited to the use cases that they are using for which is mostly training. There's a whole culture around\ngetting cheaper GPUs and adding more VRAM to them in China in order to get\naround these import restrictions, which is kind of crazy. All of this results in these models only being viable if they\nare released in a way that we can host them ourselves and use them ourselves. There is no reason to make a great model\nin China and not release the weights because you won't be able to make money off it anyways right now. And this makes\nthese companies go from entirely ignored here to genuinely very relevant to the\nconversations we're having. The research that kicked off a ton of this AI boom is the attention is all you need paper from\nthe Google research Google brain deep mind team over at Google that was all about the transformer model that allowed\nfor us to create language models as we now know them. This then went further with OpenAI's follow-up research,\nimproving language understanding by generative pre-training. These two papers kind of kickstarted what we now\nknow as AI. And these are open papers where they published what they did, how\nthey did it, how they got there, and what it could do. Hypothetically speaking, any one of these labs could have sat on this information, not\npublished it, and went and made crazy things with it. But then other companies wouldn't be able to innovate further.\nLike if Google didn't release this paper, OpenAI wouldn't have had the kickstart that they needed. And if OpenAI didn't follow up with this paper,\nwe wouldn't have GPT as a concept. Or maybe somebody else would have come up with it eventually. But if these were\nall private innovations that each lab was hopefully coming up with itself, the likelihood that any of them progressed\nmeaningfully is way lower. The culture around sharing our learnings and understanding is rooted deeply in\nscience and research. This is just how advancements happen in technology. On\none hand, this does remind me of the open source world, the way that we're all building on top of each other. But\non the other hand, it's not truly traditionally open because we're spending tons of money doing this\nresearch and work and only publishing the things that we think are worth publishing and sharing and don't screw\nour competitive advantages. Back when nobody had working AI, sharing all of\nthis made a lot of sense. Now that the American labs are in a cutthroat race competing with each other, their\nwillingness to share has gone down a ton. It's silly, but the first like cool\nthing I've seen for different labs supporting each other in America in 2025 was when Sam Alman tweeted that Gemini 3\nseems like a good model. Other than that, I have not seen much in terms of good faith operations between executives\nat Anthropic, Google, and OpenAI. There's just very little collaboration happening at this point because they're\ntoo busy trying to fight each other. Meanwhile, Deepseek breaks everything again with V3.2 getting crazy scores,\nespecially on tool calling stuff. And ZI is right here in the replies legend\nheart. Like this is a whole different world. This is what the research was like here before the competition\nstarted. We operated like this in the US before where these companies were supportive of each other. Now that\nthey're all cutthroat trying to win this economic race, they're not as willing to collaborate and they're much more\nskeptical of things like distillation, people using their models to generate a bunch of synthetic data to then retrain\ntheir own models with. In fact, a lot of them are accusing companies like DeepSeek of doing this with their data.\nThere was a point where certain Deepseek models, if you ask them what model are you, they would say chat GPT because they had data in their training corpus\nthat came from those American models. If Chinese models want to win, they have to be open because otherwise we won't use\nthem. If Chinese labs want to be competitive, they have to collaborate because we still have a lead there. And\nthere's one last piece that I haven't dove into much yet. I think this will make it into the video depending on how\nangry chat is. This is going to be fun. Don't get too mad, boys. China sucks at writing software.\nThey're not quite as bad as Japan, but they're up there. China's incredible\nmanufacturing. They are surprisingly competent at research. Chinese software, from my experience,\nis so atrocious that they end up spinning satellite\ncompanies up in the US so they can hire United States-based software developers to make software that works. A\nsignificant portion of Tik Tok's development happens here now because we have better engineers in the US. The\nsame way that a significant portion of manufacturing happens in China because they're better at it. Software\ndevelopment happens in America because we're better at it. People are mad about the Japanese one. I don't care. Sony's even accepted that the PlayStation\nsoftware is an untenable mess and has fully outsourced it to the United\nStates. They are hiring consultancies in the US to save the operating system for PlayStation because they are so bad at\nsoftware. great at research, great at manufacturing, pretty good at logistics,\nnot capable of writing software. A significant portion of why I made T3 Chat is that as much as I hated the\nClaude interface and the chat GBT interface, the Deep Seek 1 was actually unusable, entirely unusable, miserable\nto touch. And I wanted to use the model somewhere better. And I made T3 Chat kind of as a pun on V3 Chat because I\nlike Deepseek V3 so much and I wanted to have a better interface for it. The Chinese labs cannot compete on the\nsoftware side. And this is where most people come in. Most users don't see\nthis new model came out and then go download the weights and try running it on their local GPU. Most of these\nweights cannot be run that way because most of them are too massive to run even on like a high-end local GPU. You're not\ngoing to run Mini Max or Kimmy K2 thinking on your RTX 5080 anytime soon.\nSo the average consumer wouldn't have done that anyways even if they could. They're going to go to the app store and\nlook up the app and the Deepseek app will never even come close to the apps by the American labs or even by a third\nparty like Perplexity or like us with T3 chat. So they cannot win on the top level where people are adopting the\nthing. They cannot win on the API level because no American companies are going to use their APIs. So they have to go\neven deeper. They have to provide the models so they can win at that level some amount and we can build everything\nelse the way we want to on top. So what about America? Can the US make a\ncomeback here? Can we somehow get back in the actual ring with openweight\nmodels? The only major US-based openweight model to come out this year has been the new models from OpenAI. GBT\nOSS120 is the fourth best performing openweight model according to artificial\nanalysis. That doesn't sound great. Like it's open AI. They're a half trillion dollar company. How are they not\ncompeting with these small Chinese labs? It's not because they don't have the resources to do it. It's because of this. The 12B Mini Maxm 2 is 230 billion\nparams, almost double. And according to artificial analysis, it gets the same score. Deepseek 3.2 is 685 billion\nparams, 5x the amount on the OpenAI open weight model. Kimmy K2 thinking is a 1\ntrillion parameter model. Almost 10x the OpenAI model. None of these can be run\non machines that you have in your house. That is way too much memory to run any\nof these things. The 120B model can max out my RTX 5090. None of these other\nones are going to fit on your GPUs. We're talking 500 gigs of VRAM to run K2 thinking. The strategy that OpenAI took\nhere is one that I actually commend. I was at one of the original listen group\nsession things they did with developers who wanted the open weight models and they actually like had Sam Alman come in and talk to us a whole bunch. It was\ngenuinely really cool and we got to ask a ton of questions. In Sam's opinion, the only reason you would want an open\nweight model when there are good APIs with closed weight models is because you want to run it on hardware you own. I\nthink there's a real value in the competition of different providers hosting models that makes something like\ncertain Kimmy models or DeepSec models really fast on certain providers. But for the most part, he is right. If\nthere's a model from a lab you trust that's hosted in places you trust, the only reason you would want it to be open\nweight is so you can run it yourself. It's not going to be cheaper to spin up a bunch of servers that you need to\nspend hundreds of thousands of dollars in GPUs on than it is to just hit an API from somebody who's already doing it. It\nwould be way cheaper to run it on your GPU in your house, though. It'd be even cheaper to run on your laptop with a\nmuch smaller model. So, they were trying to figure out what sizes to target based\non what we wanted to run them on. Every other lab seems focused on how much money do they have and how smart a model\ncan they generate, not thinking about how big or small is the model going to be. They're much more so thinking about\nhow they can win and have the best scores possible. Open AAI already knows they have the best scores possible. They\ndon't want to make the openweight models do that because then they're just giving a free win to all their competition, but\nthey do want to play in the open model space. They do want to give models to people who want to run them on their own hardware. And that's why they put out\ntwo models, the 120B and the 20B. The 120 bill you can run on a single beefy\nenough GPU. And the 20 bill you can run on a modern enough laptop with a real GPU in it. That was a very specific\ndecision they made rather than how good of a model can we make possibly. They thought about this as given these two\nperformance targets, how smart can we make something that fits within that box? And they have crushed that box.\nThere is nothing that comes close to GBT OSS120B within those performance constraints.\nThis is a really good angle for American companies to compete in the openw weight space and I am thankful somebody\nactually took the time to do it this way. I can use these models for real things. And since these models are\nlighter and easier to run, some of the speeds these companies are getting out of them are nuts. GBT OSS120B\nis pulling on Parasale 300 TPS\non Somanova. It's pulling 650 on Grock. It's 550. That's crazy.\nThere are models that are pulling 10 tokens per second. This is 55 times faster than some of those models. And\nmany of them are dumber, too. The fact that these models are usable on consumer hardware and can perform that fast in\ncertain cases and are actually useful for things is incredible. And I see why this is the angle OpenAI took. OpenAI\nwas not interested in making a model that competed with GPT5 or 5.1. They were interested in making the best\npossible thing you could run yourself. These Chinese labs aren't interested in making things you can run yourself. If\nit so happens that you can, that's a cool side effect. There are two customers of open weight\nconsumers and enthusiasts and info providers.\nOpenAI is an info provider. That's how they make a meaningful amount of their money about 20 to 30% depending on the\nmonth. They don't want to or they don't want more info providers. Enthropic you\ncan use their mo with enthropic you can use their models on Google and AWS and now also on Azure. OpenAI was only\nusable on OpenAI's infra until somewhat recently where they partnered with Azure and Microsoft during one of their crazy\nfinance rounds. So Azure can now host some OpenAI models. Meanwhile, all of the Chinese labs are hostable on almost\nall of those providers and a bunch of other places too. Most of the Chinese models we're talking about cannot reasonably be hosted by a consumer or an\nenthusiast. And people are wondering about how much RAM does 120 bill per RAM mean. It doesn't mean 120 gigs. In this\ncase, it means 80 gigs of VRAM. Supposedly, you can get away with 60 fine. And honestly, you can get away\nwith a little less in certain cases, but you can run the 120 bill model on a computer with enough VRAM. Yeah. So,\nthis laptop's 128 gig. Let's set up LM Studio and try it quick. One of the cool things about Apple Silicon is that the\nGPUs and the CPUs are sharing memory. You don't have separate VRAM from regular RAM, which means I have 128 gigs\nof VRAM on this machine. Look, when you set up LM Studio, it tells you to use GBT OSS 20 because it's one of the best\nsmall models. So, the 20 bill model is 12 gigs and the 120 bill model is 60\nplus. So, I just got the GBT OSS 20 bill running on my laptop here. It's a maxed\nout M4 Pro from Apple. M4 Max. Didn't have the patience to wait for the M5. I\ngot it very recently. We got 128 gigs of RAM and we're using 12 gigs right now from the GP OSS20 bill model. Can I tell\nit to write me three poems about JavaScript?\nAnd it is flying. That was 117 tokens per second on my laptop locally. Pretty\ncool, right? Let's switch this over to OSS 120 bill. It's going to take a sec\nto load because it has to load that all into memory. And you can see my memory consumption going up fast. We're now at\n59 gigs of VRAM being used for this. It's total memory cuz Mac OS and Apple\nsilicon using the same memory for both. Send\nslower, but that chugged. The fact that I can get almost 80 TPS on a model that smart on my own computer. Do you\nunderstand how cool that is? This is what OpenAI's choice was. They wanted to make models that you can run\non real consumer hardware. And while there aren't many good GPUs that you can buy and plug into your desktop that can\ndo this because desktop GPUs have very limited VRAM, if you get a Mac Studio or a MacBook Pro with enough RAM, you can\nactually do inference on it. That's the difference. It's so different that we just got a crazy comment from chat.\nApple's RAM prices seem reasonable now. They actually kind of are when you consider this and also the crazy squeeze\nhappening in memory. This is cool. This is really cool. The point of this distinction for me is very simple. Open\ninterested in helping the competition here because they are deep in that space and they see no issue with the current\nstate of in providers in the US. The Chinese labs can't provide their own infrastructure because nobody will use\nit. So they need other input providers to host their stuff. By doing openweight models, they can convince those labs to\nhost their things. Consumers and enthusiasts can't use most of those models because they're way too big, but\nthey can use the two that OpenAI released. OpenAI is interested in this space because it's one of the few that\nthey weren't really competitive in and they wanted to to go back and win it and they did. The the goal of the GPOSS\nmodels has been achieved. If you are running a model locally, there's a good chance you're using GPTOSS or you're\ndoing something less optimal than you otherwise could. But that's not winning open weights because that's not going to\nget you high up on the chart here. And this is where my conclusion comes. I\ndon't think we're ever going to see an open weight model from the US win on\nthis chart ever again. There's just very little incentive for labs to do it. Meanwhile, the Chinese labs won't make\nit to America and they won't even be on charts like this if they don't do open weight. Don't think of this as why\naren't American companies keeping up with these open weight models. Rather, think of this as why do the Chinese labs\nhave to do open weight even if nobody can use the things they're publishing other than like six companies. There are\nvery very few places in the world that can handle a one trillion parameter model. But they put it out anyways\nbecause they need a way for American labs and American companies and infrastructure providers to use it.\nThere is some hope left, but it's dwindling fast because at the very end of this chart, we have Llama. Meta has\nreleased all of their models as open weight. Historically, they were one of the first people doing good openweight\nmodels. In fact, the way a lot of people were using Deepseek models originally wasn't through DeepSeek. It was by using\nthe Deepseek R1 model to do a finetune on Llama 3. And a lot of us were using\nthat Llama 3 fine-tune as Deepseek even though it was Llama bastardiz into acting like Deepseek.\nThere is some chance for Meta to catch up here, especially some of the hires they've made. But I just don't see it\nhappening. They are so so far behind at this point. There's also some effort to try and fund this type of research. like\nthe White House's attempts to fund open models and AI research in the US, providing everything from inference to\npower and paying for research to happen here. This was published in July and I\nhave not heard anything about it since there's also the potential security risk. The problem with openweight models\nin security is that you can't take it back once you put it out. If it turns out that you could use Deep Seek 3.2 to\nto make a nuclear weapon. They can't take it away. That issue now exists\nforever in the model in the weights. Once it is published, you can't unpublish it. Meanwhile, if some issue\nwas discovered with OpenAI's new model, you can add a layer in front to prevent it. If it turns out GPT5's weights are\ncapable of telling you how to make a nuclear weapon, you can put a safeguard in front when the API request comes in\nand before the response goes out with those instructions, you can block it. you can prevent it. Security risks,\ncopyright risks, all of these types of things are a lot easier if you can block the request in or out before it gets to\nthe model. But once the model's out there, you've lost your ability to do this. So there's a huge risk and\nliability from the labs that are publishing these openweight models. It's a lot more work to do it. It's a lot more work to do it right. But the\nChinese labs don't really care. Their willingness to put out things that are\npotentially actual security risks is zero. They just don't give a They put it out when they can win benchmarks.\nThe American labs have liability to worry about. They have expectations to worry about. Investors, they don't want\nto piss off. They have to go out of their way to make sure their models are safe and don't have potential copyright\nissues. And even if we start funding the creation of these openweight models here, the expectations that would be set\non them from the government of them being safe, reliable, and hitting the expectations of the American government\nis going to make it a lot harder to do, right? It's a lot easier to add these things in front of the model than in the\nmodel itself. And when you are giving the model weights out, you're giving up your ability to control what goes in. So\nmy conclusion is pretty clear. I think I do not see America competing in the open\nweight models that are only hostable via infra providers like these super giant models. I don't see us competing there.\nBut I think we have a unique potential to win here. As more people get stronger computers with better GPUs as more\nconsumers have more reasons to try out these models, as Apple starts shipping\nmodels on our devices, as Chrome starts shipping models in Chrome itself, where there's more reason to run locally,\nthere's a very, very good chance that America can win with those. But we need\nan incentive if we're going to give out the weights. And right now there is not much incentive for American labs to give\nmodels out for free to their competition. There is potentially incentive to give us things that we can run on our own machines. So I don't see\nus winning anytime soon, but I hope we can win here. Let me know what you guys think. Am I way overblowing this or is\nChina definitely going to be the winner of open weight? Curious how y'all feel and if you even care, let me know. And until next time, peace nerds.\n"

ERROR: type should be string, got "https://youtu.be/BVTg-yJNRWk\n\nAI sucks at art still\n\nAt this point, a decent number of y'all probably know me for my coverage of various AI things, especially after the\nNano Banana 2 video I just put out a few days ago. Your thoughts are probably that I'm all in on all this AI stuff,\nand I want to replace everything with AI. Not only is that not true, there's actually a handful of types of AI that I\nreally just don't vibe with that much, in particular, media generation. But wait, didn't you really like Nano Banana\nPro? Aren't you selling it in your own service? Aren't you being a super hypocritical person by saying this?\nI want you guys to hear me out because I have a lot of feelings about this. Most of y'all probably know me for code, but\nthat's not my only degree. I'm actually really, really into music. I have been my whole life. I have a degree in audio\nengineering is the thing I care a lot about. I have lots of friends who are in music, almost as many as are in tech,\nand I spend a lot of time between these spaces, which is why seeing the valuation for Sunno, the AI music\ngenerator, deeply hurt me. $2.45 45 billion valuation in an industry that's\nlucky to clear 30 bill a year is insane. This company is worth as much as a tenth\nof the music industry. What I've been getting in arguments about this on Twitter because it's a\nthing I care about a lot. And it's not just music gen. There are similar issues with image gen, video gen, and all of\nthe attempts to do media stuff as a whole right now. and none of it really\nfeels viable or valuable. There's a lot of layers to my feelings here and they are admittedly quite complex. I have so\nmuch to say here about how money moves in the art world and how creation happens in the first place. But sadly,\ntalking about art is the only thing that pays less than making art. So, we do need to take a quick break for today's sponsor. It's ever been easier to make\nthe best possible UI you can imagine. The catch is how do you imagine it? If you're not creative enough or have the\nexperience to make a good design, it's really hard to get right. Thankfully, there are millions of apps and screens\nand pieces of software that are already made for us to reference. But how do you find the right one for inspiration? It's\nreally annoying going through Google images trying to find the best possible screenshot to use as a mock for the\nthing that you're building. Unless you're using today's sponsor, Mobin. These guys are the best source for real\nworld inspiration. And if you've noticed my UIs looking better lately, they're pretty much the only reason why. Okay,\nthe new models being better design helps, but Mobin helps a hell of a lot more. When they first hit me up about sponsorship, I was curious. So, I signed\nup. And my original ad with them is my organic reaction to just being blown away with how useful it is. Now, I use\nit almost always when I'm building UI, especially on iOS, but even on web projects, too. Let's say I want to make\na nicel looking login screen. We can look at fabrics or adelines or any other\neven somewhat relevant reference. Let's say we want something more specific to our category like we want something for\nbusiness websites. Here we have dubs, signin page, clickups, intercoms, and more. Let's say we don't want a login\npage. We actually want I don't know a comment section. Here are some business\ncomment sections. actually really useful because comments mean something very different in the business world than they mean on I don't know YouTube,\nTwitter, anything like that. And here we can see exactly how comments look on Trello. Now, if you want to use this,\nyou can save the image or click copy or just control C, command C, whichever OS you're on. You're a designer, you're\nprobably on a Mac, so command C. Hop over to your vibe coding tool of choice, command V, say make it look like this,\nand you're done. It's never been easier to get inspiration from other designs. Even if you're not trying to vibe code\nthis out and you're just looking for references or you're spending your days in Figma and you're trying to find some inspiration, mom is an insane value.\nIt's only 10 bucks a month per user. The amount of time it saved me alone and the amount of improvement to the quality of\nmy user experiences and my UIs is worth way more than that. If you don't believe me, go try it out for yourself at\nsoyv.link/mobin. There's a bunch of layers here that I want to break down. We have different\ntypes of generation. So we have LMS specifically. We have the ability to generate text with them. So, essays,\narticles, whatever else you're imagining here. We also have the ability to generate code with them. And this part\nin particular is very interesting. We'll go back to that in a bit. There's other\ntypes of media gen, the ones that we're actually here to talk about. Most of these are diffusion based, but we don't need to go into the tech behind it. That\ndoesn't matter as much. What we do want to talk about is image gen, audio gen,\nand video gen. And when I say audiogen, I mean music generation primarily, but\nhonestly, this could be split into two categories of dialogue and music. And I\nknow it sounds kind of cringe and contradictory that I think this stuff is really cool and I think all this stuff\nis really cringe, especially when I've been using image gen stuff more and starting to plug it as part of my\nproduct. But I have reasons and I haven't done a great job articulating those. I touched on it a little bit in\nmy previous video where I was talking about PewDiePie's AI takes. It was a good opportunity to dive in, but I just\ngave a surface level why I think this way. I haven't done a proper deep dive just yet. The Oatmeal also made an\nawesome comic about AI art and how he feels about it. I think it's one of the\nmost neutral and reasonable takes I've seen from an artist about AI stuff. I'll leave a link in the description if you\nwant to go read this. I highly recommend it. Oatmeal even has a section about how AI art can be useful, which is going to be a really useful transition for me\nhere because this is what I want to talk about. I believe it can be a powerful tool for dealing with the minutia of drawing, the parts of it that feel\nadministrative, not creative. Every artist has their own minutia. For me, it's backgrounds. Most of my comics\ndon't have them, which is why I often draw pale blobs floating in space. Eliminating backgrounds lets me focus on\nthe part that matters to me, the comedy, the fun, the rat copter. I don't ever plan on using AI in my\ncomics, but I see how it could let me focus on the creative essential aspects of cartooning. It's like the artist\nequivalent of using a spell checker. Now, we're cooking with gas. This is a really key piece I want to dive in on.\nNot just because all most AI is autocomplete on steroids, but because this idea of it being a tool in the\ntoolbox is key in all of these spaces. the creators that are relevant. Be it a\nmusician, be it a graphics artist, painter, be it a developer writing software, we all have our toolbox. The\ntoolbox is full of all the different things we use. It could be our code editor like VS Code or Cursor or Vim or\nsomething like that. And then we have tools inside of it. It could be production software like FL Studio,\nwhich I just reinstalled so I can demo this here a little bit. It could be your video editing suite like we use Final\nCut for all of our videos. It could be a lot of different things, but your toolbox is usually one major piece of\nsoftware and then other things that are either inside of it or alongside it. So, some developers like using the terminal\ninside of VS Code. Other devs have a terminal outside of VS Code and they hop between the two apps. That's what I do.\nSo, the question comes down to the relationship between these AI tools that are now being created for all of these\nthings and the relationship with existing professional tools that we already use. I'm going to talk a little\nbit about how AI code happened. The first major AI dev tool was C-pilot. The\nway Copilot worked initially was with a small kind of dumb model that would watch what you're doing in your editor\nand generate autocomplete that was based on what you were just doing and where the cursor is in your file. So you could\nwrite a comment like do a bfs funk and as soon as you start typing the word\nfunction it could realize what you're about to do and write the autocomplete for it. And this was awesome because\nco-pilot existed inside of our toolbox. So if our toolbox is primarily VS Code\nand there's this new AI tool that we can use called co-pilot. Copilot is inside\nof our tools. It's inside of VS Code. It's not trying to replace it. It's not trying to rethink it. is trying to\naugment it and make certain tedious tasks easier for us to do as developers.\nAnd as this became a thing, the model started to get better. Microsoft and OpenAI started getting more data from\npeople using Copilot. And then the flywheel really started going and the result was that the models got really\ngood at code. And now we have more and more tools that can exist outside of our\nexisting toolbox. I would argue something like cursor still fits here. Even if it is replacing VS Code, it's\nstill based in VS Code. But it still has the whole VS Code ecosystem. Cursor, Windsurf, even anti-gravity, the new\nGoogle IDE. They're all different harnesses for the same set of tools.\nThey're not replacing our toolbox. They are a new flavor of the toolbox that has\nthe same tools inside of it alongside some new interesting pieces. So now I want to imagine a different timeline.\nImagine Co-Pilot was not the first AI tool that we saw as developers. Imagine\nsomething else was. Imagine Lovable was the first AI dev tool. And imagine\nlovable didn't let you see the code or link with GitHub at all. Lovable wouldn't be a tool augmenting the way\nthat we write code. Lovable would be a replacement for our toolbox. So instead of VS Code having Lovable inside of it\nor Lovable being a new alternative to VS Code with the same stuff inside of it, Lovable exists. Therefore, we don't need\nour toolbox anymore. Now, if you're a real developer, you're just going to look at this and laugh.\nEspecially if you remember how bad these models were initially. Like, lol. There's no world in which I can write\ncode without looking at the code. That's just silly. Why would anyone want to do this? Now, let's go a step further.\nLet's say that the creators of Lovable weren't actually devs. Imagine that the people who made Lovable sat down one day\nas friends who were like, I don't know, into football or something and they're like, \"Hey guys, it'd be really cool if\nwe could build an app for managing our football stuff. It's a shame none of us know how to code. Let's try.\" Oh,\nlearning to code is hard. There would be so many coders if it was easier to learn how to code. We should fix that. And\nthen they went and made lovable, a rejection of all existing tools and standards and industry and successful\npeople in the space. I think we all can predict what would happen there. We would make fun of them. Developers would\nlaugh at this. Why are all these people who don't know [ __ ] about software showing up saying that what we do is too\nhard, so they're going to make it way easier when they don't understand it in the first place? We would laugh at them. That would be absurd. What if I told you\nthis is exactly what all the media genen companies are doing? The people who worked on Nano Banana have not been\nusing Photoshop for 15 years. The people who are working at Sunno have not been producing, mixing, and mastering music\ntheir whole lives. And I promise you, the people who are doing Sora over at OpenAI don't understand their way around\na video editor particularly well. That's why they want to replace our toolboxes. You don't come to the conclusion that\nthe toolbox is too hard, we should throw it out, unless you don't understand it. And if your goal is to replace the\ntoolbox, there's a very good chance you don't understand that toolbox at all. And that's what I've seen the most from\nthe companies building these things. They don't understand how we build. They don't understand how we create. They\ndon't understand the ins and outs of what we're doing and how we do it. There are differences in how these things\npractically apply when you're doing media genen. For example, I do thumbnails that have a lot of assets in\nthem. Similar to what oatmeal said, some of these things are tedious and I don't like doing them. We all have our own\nminutiae, the things that we don't feel like doing. A common use case for me is fake chat interfaces. Like I want to\nhave OpenAI in Google having a back and forth on Slack that I can use as part of a thumbnail. This is an asset that I\nwant as part of a greater creation. It's one of the pieces of the puzzle. Similar\nto how I would use an LLM inside of VS Code to generate a part of what I'm working on, I can use Gemini and Nano\nBanana Pro to create an a piece of an asset I need for something else that I'm\nmaking. It can be really useful for that type of stuff. But it has so many catches that make it way way less\nuseful. Basic stuff like background removal, it won't do. If you ask Nano Banana for a transparent background,\nit'll hallucinate the like gray white checkerboard pattern in the background. Won't even have it aligned and now you\ncan't really get your thing out of that image. It's not built to be useful as a\npiece of the puzzle. It is built to do the whole puzzle. It's built to replace the toolbox, but due to the nature of\nhow malleable images are, you can work around it. For example, if I wanted to\nuse this part, there's no world in which I would save and use this whole screenshot as is. What I would do is\nzoom in and grab just this little piece I want and then go make the rest of the thumbnail.\nGrab that pop into a real professional graphic suite. I use Affinity Photo. New\nthumbnail. Expand this a bit. Make the background the right color. hop into the pick thing\nclone that Ben and I have been working on to rethink a lot of stuff for it. He actually built a tag system that I can\nuse. Tags Theo face. Grab a picture of my face. And yes, the backgrounds here are removed with AI.\nPaste this in. Add a shadow.\nAdjust this. It doesn't look quite right. It's also lower res than I would want. I'd probably go upscale this in some way. This is a flow I do often for\nmaking some of my thumbnails. Like one in 10ish, I would guess uses AI for\ngenerating a part because the alternative is I go spend a ton of time in inspect element in Slack trying to\nmake the exact screenshot I'm trying to make. And I used to do that. It sucks. It's so not fun. The fact that these AI\ngenerated tools can streamline this one particular annoying part for me is great, but I often still have to make a\nton of changes. Like here, I hate the emoji stuff being there. So, I'm going to rasterize this piece. Blown in in\ncase I don't like the changes I make. Pop it here. Delete.\nDelete. Cut. And then same deal once more. Like\nI cannot tell you guys how often I find myself doing exactly what you're seeing here when working on assets for videos\nthere. And I'll probably move that today bar down as well and then move the whole thing up a little. How did you get in\nhere, Miles? I locked you out. Move that group.\nThere we go. Actually, what I can do have this separate.\nLook at that. Now, this is a usable thumbnail. I just had to make threeish\nminutes of changes to get it roughly where I want. And even then, it's low enough res. I'm probably I probably\nwould go do another pass if I wanted to take this more seriously. Quick cat break.\nGenerate this. I dare you. Good boy. You're not supposed to be in here,\nMiles. I don't even know how you broke in.\nAnyways, this is an example of how I use something that exists outside of my tools, in this case, Nano Banana, and\nforce it to fit within my needs. It would be so much better if instead of giving me this baked image, it could\nsomehow give me layers or a Photoshop file where I can adjust things because\ncode is way more malleable than media. Imagine if an AI coding tool didn't ever\ngive you the code. it would only give you the compiled binary and you had no way to get back the source to go make\nchanges unless you decompiled it, reverse engineered it, and then started making those changes manually. That's\nhow it feels to work with these media generation tools. Thankfully, images are a relatively malleable medium where I\ncan like crop out pieces, use other tools to remove backgrounds, upscale parts, layer parts, and just\ndeal with all of this myself. But that's not the case for music and video. At least not on the same level at all.\nSorry, the cat's being extra needy right now.\nLike, I was helping a friend with thumbnails last night and spent at least 20 to 30 minutes cropping, cutting,\nchopping, upscaling, and micro adjusting some AI generated screenshots so that it could look decent. And man, it was\nannoying. It was still slightly easier than if I was to make everything myself. But like the point isn't that I'm\ngenerating media so that I don't have to do the work. It's that I'm using generation to smooth out the tedious\nparts. And this is when AI is great as a developer. This is also when AI is great as a person who works in media. I do\nproduce a lot of content, believe it or not, a video a day and we do multiple thumbnails for each one. We are producing a ton of stuff. Buddy, I am so\nsorry. I love you to death, but you are making my job impossible.\nI'm going to kick him out. Give me a sec, guys. So, why is this so different from music and video? Well, those\nmediums just are way harder to do this type of stuff with. I can't just crop\nsomething and move it down in a video because the next frame, it'll now be wrong. And if I try to do it frame by\nframe and anything's off between these different frames, it is an absolute mess. There are some things within tools\nlike Da Vinci and Final Cut and now even Premiere that let you like rotoscope to remove a background from somebody.\nYou've probably seen FaZe do this in my own intros. It's using AI and that's\ncool, but that's not what people are talking when they talk about media generation. The idea of a world where I\ncan select myself in a video and say changes shirt color to red and the rest of the hourong video is the same but my\nshirt's red. None of the tools are trying to do that. They're trying to replace my editor. And this is the\nproblem. There's almost like a scale here of how much is a tool trying to complement your toolbox versus trying to\nreplace your toolbox. Something like Copilot, specifically the original\nversion of Copilot that was just the autocomplete very much falls into complimenting your toolbox. Something\nlike Nano Banana fits in between the two in my opinion because I can use it to comment my toolbox. I can use the images\nit's generating to go do things in other editors. But that's not because Nano Banana was built for this case. That's\nbecause the idea of image generation and the way images work is more malleable.\nSo I can screw with things. Code is way easier because you can just move code around, change the order of lines,\nchange the functions that are being called, replace a whole chunk. Code is super malleable. Images are kind of\nmalleable. Music's a bit malleable and video is not particularly malleable. So\nnow we need to talk about the thing that we're all here for. So where does fit in\nthis? Allo frustrates me. Here's the post that\ntriggered me enough to make this video. I've been friends of Jasmine for a while, so this one came as a shock to\nme. I grew up singing. I sang everywhere I went. I wrote songs in my diary. I told teachers that I wanted to be a\nsinger and a songwriter when I grew up. Wanting to be a musician in 2006 required resources that a low-income\nfamily didn't have. My parents couldn't afford to get me any instruments. They couldn't pay for music lessons. They\ncouldn't get me into studios. A dream I had became just a memory until now. I'm\nbeyond proud and honored to get to work at a company that is enabling music creation for everyone. For the\n13-year-old kid in their bedroom who dreams of being a musician, you can be one. For all the professional artists,\nyou can do more of what you love. I really wish Sunnu existed 20 years ago when I was a kid in elementary school,\nshowing stranger songs I wrote with no way to produce them. But I'm really, really happy that it exists today for\nall the other kids who might need it. We are still just getting started.\nSince I know a lot of you guys are developers, I grew up obsessed with computers. I tinkered with every app I\ncould find, sketch website ideas in my notebook, and told teachers I wanted to be a programmer when I grew up. Wanting to be a dev in 2006 required resources\nthat a low-inccome family didn't have. My parents couldn't afford a decent computer. They couldn't pay for coding classes. They couldn't send me to camps\nor get me mentorship. A dream I had slowly turned into just a memory until now. Do you see the problem here? Like\nif you're a developer, you know that you can code on any even like vaguely almost\ndecent computer. Some of the most important albums ever released were produced on crappy MacBooks from years\nprior using the free or pirated copies of GarageBand. You might have heard of this artist called Grimes. Her first\nalbum that included the track Oblivion kind of broke pop music at the time. What you might not know is that she\nproduced this album in GarageBand on an old halfbroken MacBook. You might have\nalso heard of Justice, the bands who did dance and Genesis and a couple other big dance songs back in 2007. Funny enough,\nthese were also produced in Garage Band. This is a particularly funny example because this album came out in 2007 and\nthe tweet we just read was about 2006. No excuses. There are artists producing\nincredible, groundbreaking work on base tier iPads they bought four years ago. And there are so many musicians putting\nout everything you need and more to learn how to produce music. Jane Remover is one of my recent favorite producers\nand she has these incredible breakdowns showcasing exactly how she made her albums. Literally opening the project\nfiles for her album and showing you how the songs were made in the software. The\nknowledge access has never been better. The best artists just share how they do things and you can go look and learn and\neven ask questions and many of them will respond. And the tools themselves are never even more accessible either. The\nfact that you can download real production software on a phone or an iPad and make actual music is\nincredible. And if you boil down music to what you can perform on a stage with\nan instrument, we have fundamental different definitions of music. Because as cool as performance is, and like I\ndid my 15 years on piano, I get it. It's not as cool to me as the creation of\nunique ideas through sound, I'm much more interested in the ideas someone can communicate through sound and the vibes\nthey can give you than how fast they can hit notes on a stage and how many lessons and years of training they went\nthrough before they performed. That's just much less interesting to me personally. And if you think that you\ncan't be a musician because you don't have the ability to play guitar, that's like saying you can't be a programmer\nbecause you type too slow. It's just stupid. That's all it is. These tools aren't for aspiring musicians who are\nblocked because it's such a hard thing to get into. These tools are for unmotivated people who don't care about\nmusic that much, but think it would be cool if they could be a musician. I see this a ton on YouTube. I can't tell you\nhow many super awkward conversations I have had with somebody. It's like, \"Yeah, I kind of want some side income.\nI feel like this engineering thing is tiring. Kind of want to try YouTube. How can I make a successful YouTube channel\nwith as little effort as possible?\" The first question I ask is, \"How much YouTube do you watch? Like, how many\nhours a day are you watching? Who are some of your favorite YouTube channels?\" And almost always the response is, \"Eh,\nI don't watch that much YouTube. I guess Mr. Beast is cool. I've seen one of your\nvideos before. Seems like you're making a lot of money. Absurdity. Do you know why my channel's good? Cuz I'm a [ __ ]\ndegenerate. I watch way too much YouTube. And I'm not the only one. My channel manager, Ben, I met cuz I\nwatched his videos and he watched mine. My editor, FaZe, also was a avid viewer of my channel and is watching YouTube\nconstantly whenever he's working and doing other [ __ ] If you don't love the thing, you're not going to be good at\nthe thing. And in a field where there is more demand for people than there are people who love the thing, that's fair.\nBack like four years ago and way be beyond that, there were way more engineering roles than there were people\nwho loved writing code, which meant it was okay to hire people who didn't love code. And it was okay to get into code\neven if you didn't like it that much if your goal was to make money. That only works in a space where the median talent\nlevel can be successful. Only the top.1% of musicians can make a living off of\ntheir music. I have a thought experiment for you. My YouTube channel gets 2 to three million views a month on really\nlong content. My average video length is 40 to 50 minutes. So, I should be making pretty good ad revenue on that. There\nare lots of musicians who are lucky if they can squeeze out two to 300,000 plays a month on YouTube. So knowing I\nget two and a half to three million on a channel that has a very technical audience that can be advertised to for a\nlot of money and my views are way longer and experience more ads. How much money\nwould you guess I'm making off of ads on YouTube? You're probably thinking a pretty big number like 2 to three\nmillion plays with a developer audience. That's a lot. That's got to be a ton of money. You got to be tens if not\nhundreds of thousands a month, right? My channel makes 6 to eight grand a month on ad revenue. That's almost enough to\npay my editor. It's not enough for everything else that we do. It's not enough for my manager. It's not enough\nfor my equipment. It's not enough for the tokens I spend doing AI generation for my videos. Six grand a month is not\nenough to sustain what I'm doing here. And I am doing way better numbers here than most even pretty successful\nmusicians are doing. And those musicians can't do a sponsor spot in the middle of their song. Those musicians can't get as\nmany ads in their videos because the videos aren't as long as my videos are. They aren't getting [ __ ] I know a lot\nof artists that are way more famous than me that are making like twoish grand a month off of YouTube revenue and another\none to two off of Spotify and other streaming platforms. There is very very little money here. Only the absolute top\npeople make a reasonable living in music. So if you're like, \"Eh, music's kind of cool. Not really for me though.\nI'd do it if it was easier.\" You're not going to survive. If you think that learning audio software or getting\naccess to audio software is the thing preventing more musicians, you don't understand any part of this picture at\nall. You're so far removed from how this [ __ ] works that it's it's insulting. And\nall of these tools that people are making aren't trying to make the stuff we use easier. They're not trying to\nimprove my experience in actual music software. They're trying to replace all of it because it's too hard. That's so\ndumb. That's so incredibly dumb. And when I complained about this on Twitter, the replies I got were even dumber. Love\nyou, Jazz, but this really ain't it. Music's one of the most accessible fields to get into by far. Many of my favorite albums were produced in\nbedrooms on iPads. That's the difference, though. They were produced. Effort was put in by a human to create the sounds that we experienced. There is\nno art quite as soulless as AI music generation. This doesn't enable new artists. It enables unmotivated people\nwho don't care about the art. I stand behind every word here. One person said that they create melodies in MIDI and\nthey load them into Sunno Studio to make them richer by prompting the AI to arrange the melody with various instruments. I don't even know how to\nput into words how dumb this is. If you don't know what MIDI is, it's the standard for arranging melodic data to\nsay what note plays and when and how loud it plays. It's not saying what it should sound like. It's just saying the\nnotes and when the notes play. I'm going to show you something funny in FL Studio, one of my favorite pieces of\nsoftware ever made. Let's add in a new instrument. I don't have any of my VSTs installed because I literally just\ninstalled this before going live. I love toxic, but let's just use Citrus.\n[Music] A God, my keyboard skills are not where\nthey used to be. I used to be so good at playing piano on a tool like this.\nOkay, some of my keys aren't working right. I'm not used to this keyboard or how that works. The point I wanted to make here is how complex it is to\ngenerate MIDI. This is a MIDI loop. I can put notes here.\nHit play. And now it's playing.\nWell, let's say I find writing and composing melodies too hard because I am\na wannabe musician, not a real musician. I I've mostly produced on Windows as\ncringe as that is. So check this out. I can change the way I am composing to\njust put chords instead of the notes.\nCool. But that's still too much work. I want something easier.\nLet's do that. God, this program's changed so much since I last used it. I still have the old or like a super old\nversion installed on my desktop. I'm realizing I should probably learn the new one soon. Is it the wrench? Oh, yeah. riff machine.\nLook at that. I just generated a MIDI sequence.\nDo you know why this is so easy to generate? Cuz it's just [ __ ] math. That's all music theory is. Really basic\nmath with 12 tones that have basic relations with each other. It's not that\n[ __ ] difficult. If you think that's the flex, that these tools are actually useful because you brought the MIDI,\nthat's like saying lovable's a real programming tool because you brought a screenshot of a Jira ticket. I don't\n[ __ ] care. That's so far removed from the reality. Like, I've been generating melodies and sequences since I was 15\nyears old. And you know what? That generation I just did is in the tool I'm already using. It exists within the\ntoolbox. It's not replacing it, which is the whole problem I have with these things. none of them are helping me\nwithin my software. If someone was to create a loop generator or a sample generator so I could make unique samples\nto use within my music, that would be kind of cool. If somebody was to target real musicians using professional\nsoftware to try and complement their work with AI and give you useful pieces to smooth out annoying stuff, that would\nbe great. I would be super into that. But that's not how any of these companies are doing it. They're all\ntrying to replace our software with some [ __ ] back to Jazz's story because there's one piece here that really\nbreaks my heart. The idea that she really wanted to be a singer and songwriter, specifically the songwriter\npiece. The majority of musicians, the majority of people who spend years learning an instrument and performing\nit, the vast majority of them don't ever write music. They are performing music\nwritten by others. Because most people see music as your ability to perform,\nnot your ability to create. And that's always irked me deeply. One of the places it irks me the most is in how\ncertain musicians are so elitist about it. I think it's important to allow the kids to do new cool things in music. One\nof the kids I really like is an artist named Netspend. I don't know how much of any of his things will be able to play\nwithout problems. Actually, I bet this might be okay because it's been taken off of everything. That's hilarious. It\nhas some other sample that's in it tagged there. [Music]\n[Music] I know a bunch of you are looking at\nthis and immediately saying, \"Fuck this rap [ __ ] What is this? I hate this.\" And fine, whatever. If it's not for you,\nit's not for you. Okay, this is a 16-year-old kid that is massively\ninnovating the space that is hip-hop right now. I think he's incredible. It\ntook a while for him to grow on me, but now he absolutely has. His previous songs were way indier. The biggest ones\nwere like up to 100k or so plays, but this one blew up. This one got to a\nmillion plays surprisingly quick. and it was a huge sound change, aesthetic change, and a growing up, glowing up\nmoment for him for sure. You might notice that this song is not on his channel. That's because it got copyright\nstruck and not because of anything happening in the main song, but actually\nbecause of what happens at the end here.\n[Music] That little guitar loop at the end there\nis the intro for a different song that he was going to put out later. And the guitar loop is sampled from a deaf tone\nsong from like 2004. They thought they would be able to get the sample cleared. They had a a light\nclear supposedly. Everything was taking forever. So Netspen just randomly dropped the song on his personal YouTube\nchannel. It gets DMCA almost immediately. And not only did they take down that song, they took down this one\nbecause it included the snippet of that other song. They took his first million play track away from him because they\nthought his usage in repurposing of a 20-year-old guitar loop in a rap song\nwas somehow competing with their work and costing them money and business opportunity. That's [ __ ] delusion.\nThat is a disdain for art and progress within mediums. Just absurd. And you can\nsay all you want like, \"Oh, that's the label being shitty.\" And I know a lot of people are going to say that. Like, labels suck. I don't [ __ ] care. The\nmusician can stand up and say, \"Yo, this sucks. This happened with Oliver Treeway back where he did interviews and said\nthat his label was being [ __ ] with other artists using some of his stuff and how important this was to him and how frustrating it was.\" Defilent\nbecause they look down on hiphop. They look down on these other spaces. So, they're [ __ ] about it. It is what it\nis. They're not as bad as Metallica, who showed up to Congress with like 500 pages printed of people's names that\nthey guessed based on IP addresses to try and sue all the people who stole\ntheir music because they opened Lime Wire once. There's something about the metal scene and how it's so pretentious\nabout how fast they can play their guitars where they just stop caring about artistic progress. And it sucks.\nIt sucks so hard. None of this is about copyright law and defending your\ncopyright or you can lose it. That's not what's happening here. This is people who don't actually care getting in the\nway of those who do. And it's just obnoxious. And now some random re-uploaded this video and got three and\na half million plays that should be going to net spend, should be paying net spend, should be advancing his career,\nbut it's being held back by people with a weak, outdated view of how music works. And these people are all over the\nindustry. And I'll bet my ass that the first types of artists that are going to start collaborating with companies like are going to be deaf and Metallica\nbecause they see no issue with this insanity. Actual [ __ ] insanity. Like like just just imagine being a musician\nwho hasn't been relevant for 20 years and a kid who is 16 or 17 years old\nlikes your stuff so much that they reference it in their work as a very small piece of their work. and your\nresponse is either to get it taken down or sit idally as it gets taken down. That's just insane to me. And this type\nof like like on one hand this is considered theft and they can go through\nthe DMCA process and win it. On the other hand, Sunno letting you generate a song in the style of Coldplay is totally\nfine. Absurdity. Actually insane. And it frustrates me so much. And these are the\ntypes of people who are going to be cool with it because they're not interested in making cool music or sound. They're interested in showcasing how fast and\nwell they can play their [ __ ] instruments. And I just don't care. I really don't care. That's like showing\noff how good of a programmer you are by typing really fast. One more angle I want to talk about this with before we\ntalk about how to fix it, and we will get to how to fix this. I just really want to talk about one more key\ndifference between these types of media. There's two sides to creating music. There's the side of the listener who is\nexperiencing that music and having feelings and emotions from that music.\nAnd then there's the other side which is the vibe and feeling you get from creating it. The satisfaction you get\nfrom making a new piece of media. We can break this down to creation and consumption. When you are creating\nmusic, it gives you certain feelings and your goal is to have the music you create give consumers of that music\nfeelings. You can argue that writing is similar where there is a vibe you get when you write an article or a blog post\nor an essay on something you care a lot about and then there's a feeling people will get when they consume that. You could say the same about video. You can\nsay the same about images some amount less so. But there's a third piece here\nvalue. If I write a really good blog post, it might make me feel satisfied as the writer. It might make you satisfied\nas a person consuming it and give you some interesting information and make you feel good, but it also will help my\nwebsite be placed better on Google. It has a practical value when it's being read by robots that allow for my website\nto perform better and have an immediate business impact for me. This is even more so the case with code. When I write\ncode, it might make me feel pretty good and smart and awesome. Reading code doesn't really do anything for anyone.\nLike it's sometimes cool to look at some code and be like, \"Wow, that's really clever.\" Less so than in any of these other places, but it's still kind of a\nthing, but the majority of the value in code isn't from people reading it and having a feeling. You don't get value\nfrom code by consuming it. You get value from code by running it. This is one of the reasons all the work that we're\nputting into making LLMs great actually makes some economic sense because the majority of the output of LLM is never\nlooked at by a human. The majority of the reasoning traces aren't even being passed to the user, much less seen by\nthe user. The majority of the code being generated isn't being opened. It's being run in the background, executing tasks,\nor just being generated and then thrown away after the person checks the website once. The majority of content being\ngenerated by LLMs is never consumed by a human. The majority of content being\ngenerated by Sununo in these music generators is consumed by one human and\nnothing else. There's a lot of value in the words and code and things being generated by a large language model\noutside of someone looking at it directly. There is no value in music generation outside of it being consumed.\nThere is no value in image generation outside of it being consumed. And there is no value in video generation other\nthan it being consumed. This is the problem. LMS and things built around them are inherently way more composable,\npowerful, and to be frank, exciting to me because I can see ways to use those that are really cool. Media Genen stuff\nisn't useful outside of it being listened to or watched, and it sucks.\nThere are places where you need to do some of this work that aren't as direct, like I am locked in paying attention.\nfor example, a background in one of my thumbnails or some asset that's part of a poster that's being put out in the\nreal world somewhere. Those types of things might not need the same level of detail and attention because you're not\ntrying to invoke a feeling with every detail you're putting in the same way. You're just trying to make it look good\nenough, but there isn't that much of that type of stuff, especially in the music world. Like, sure, it'd be cool if\nlike you could generate the soundtrack for a movie without having to pay a musician, but the soundtrack's one of\nthe coolest parts of movies. Why would you automate that piece? It just feels [ __ ] insane to me. Like, and I think\nthis is the part that gets missed the most. LMS can generate value outside of how they're perceived and consumed. None\nof the other types of media generation have this as the case. There is no value\nin ao song as soon as you've listened to it for the last time. And it also doesn't provide that same cool vibe of,\noh my god, I created this thing I had in my head. Because a lot of the satisfaction on the creation side comes\nfrom the execution of a thing that you were imagining. If you're imagining an application that you wish existed and\nthe way you want to write it and then you go do it, you get the satisfaction for executing this vision. When you\nwrite a blog post that is a thought that you're just trying to get out of your head, you get the satisfaction of it being out of your head. When you AI\ngenerate a song with a text prompt, you're just hitting a random dice roll and hoping what comes out is cool.\nNowhere near as satisfying. You can't imagine a song in your head, and then\nget to make it. You can write lyrics, maybe even hand it a MIDI track, and\nthen it will generate whatever it feels like, but you're not steering it. You're not controlling it. You're telling it\nwhat you want and hope you're getting it back. It would be the equivalent of having an idea for a song. So you find a\nmusician, pay the money, and say, \"Hey, make a song that's like this and hope that they do it roughly how you want,\njust way more automated and way worse.\" Generating music is nothing like generating code because you can't mingle\nwith the pieces. It's not useful outside of how it's consumed. It's nothing like sampling because you're not looking at a\npiece that you understand and respect and try to repurpose parts of it somewhere else. It's nothing like AI\ngenerated blog posts because you're not getting the idea you have out and editing it to get it just right. It's\nsomeone who doesn't understand music telling their friend to go do it. It's like if I told you, I want an app that's\nlike Tinder for dogs. Go make it for me and hoping you can figure out all the pieces. It's just stupid. It doesn't\nwork that way. It's silly. And I won't get satisfaction when you're done building it. I'll get a bill. I saw some\npeople getting confused about my point here about the consumption thing. is my argument that music gets consumed by humans so it should be made by humans.\nNo, that's not my point at all. My point is actually the opposite. Code can also\nbe consumed by humans, but I don't care because the value in code isn't the human consuming it. The value in code is\nthe code running and providing a service. The value in music is exclusively the listening to it. That's\nthe only part that matters. There's no other value in music. Which means if you have two AI models, one can generate\ntext and code and one can generate music. The one that can generate text and code can make things for humans like\nblog posts, articles, essays, whatever else. And it can make code which no one will read and still be useful.\nContinuing to iterate on LLM is valuable because the things that the LMS generate are useful even if a human never touches\nit. There is zero value at all whatsoever in anything generated by a music model, an image model, or a video\nmodel outside of how it is consumed. The only value of generating a song is to\nlisten to it. The majority of text that is generated is never consumed by a human. The majority of music that is\ngenerated is exclusively consumed by one human. That's the difference here. There is no value in this outside of the\nconsumption side, which makes it so much more soulless. I hate shitty AI blog posts as much as the next guy, believe\nme. But there's a lot of value in AI generating text outside of the text that humans are reading. There is no value at\nall in this stuff. Cine Chapel was a commissioned by the Pope. The execution is what made it good and expanded upon what the Pope could ever possibly\nimagine. We credit Michelangelo for a reason. Yes, like this analogy a lot.\nThe Pope is not the one who created the chapel. He commissioned it. Michelangelo was the one who architected, designed\nit, and deserves the memory of being the one who created it. And when you use Sunno, it's effectively the creators of\nSunno who made the song, not you. But it's not even that. Like, they're not the ones who created the sounds coming\nout. They trained on a bunch of media from other artists. They don't necessarily have the rights to. At least\nthey didn't because Warner, my favorite, just formed a partnership with Sunno.\nYes, really. This deal brings together Sunno's best-in-class AI capabilities with WMG's artist development leadership\nand expertise at the intersection of music and technology. The deal also settles previous litigation between the\ncompanies. Something important to know is that Warner and other music labels will always hunt for ways to take giant\namounts of money from big companies like Sunno, like Facebook, which we'll talk about in just a sec, and also to take\nmoney from their own artists. I promise you there is effectively zero dollars going from Sunno to the artists on\nWarner. This is an agreement between Warner the business, Warner the legal team, and Sunno the startup that's worth\ntoo much money. I have way too much insight info on how all of this works. Someone just touched on it in chat.\nTwitch is another example of the DJ thing. You have no idea. I wrote the spec for the original Twitch music\nproduct. I was so deep on the ground floor there. I know way too much about how all this works. And there was one\ndeal that [ __ ] up everything for everyone. Back in 2018, Facebook and\nWarner Music Inc. recorded and published music deal for videos and messages. In the mid to late 2010s, all the music\nlabels realized that they can't get rid of YouTube and Twitter and Facebook and all these services. So instead of trying\nto do that, they could maybe charge them. So they went to all of them and tried to ink out deals so that they\ncould be paid whenever one of their songs is played on YouTube. But in reality, what it did is if you accidentally play 15 seconds of a song\ninside of a three-hour video, all the revenue now goes to Warner. Great. Awesome. If that was all that it did\nwrong, that would be fine. But this is the deal where everything fell apart. Facebook wanted to go harder on video on\nFacebook.com. They wanted to compete with Vine and YouTube more actively. They were also trying to compete with Twitch. If you remember the Facebook\ngaming days, good times. They were already starting to get sued by Warner for random 30-se secondond clips that\nhad a song playing in the background. Facebook had a lot of money. They didn't want any of this to get in their way.\nSo, they did a thing that inadvertently [ __ ] up the entire music industry and its relationship with technology. They\npaid Warner around a billion dollars for a blanket use our stuff wherever you\nfeel like license so that Warner wouldn't sue them and they could just use the stuff without having to like pay\nbased on viewership. With YouTube, every view that can be accredited to a Warner\nmusic license results in money going to Warner. On Facebook, none of it\nmattered, which is actually quite helpful because at the time Facebook was artificially inflating view counts, so it would have [ __ ] them if they were\nto pay per view. They inked this flat rate because they didn't want to deal with this all and then they proceeded to\ndo literally nothing with it. The assumption was that this would allow for all new types of music creation and\nmusic creators to be successful on Facebook. Maybe people will live stream DJ sets, maybe they'll do dance videos,\nmaybe they'll do all these different things. And it never happened. But since Facebook paid billions of dollars for\nthis license and then never used it, it set a standard that Warner has been trying to maintain since, which is that\nthe sheer concept of accessing their IP is worth billions of dollars. So that's\nwhat they should charge. And when the numbers were coming to Twitch about how much money a similar deal would cost,\nTwitch is already in the negative. They would have put Twitch much, much deeper in the negative. Like hilariously so.\nAnd Twitch's workaround was beautiful. The solution we came up with was that\nradio licensing was very different from ondemand licensing. If you can go click\na video and hear a song, that's arguably a market replacement for going and paying for the song and access to it.\nBut if you're listening to the radio and I play a song you like and you want to go listen again, you can't make me play\nit again on the radio. You have to go find and pay for the song. So, generally speaking, live music broadcast has more\nlenient rules around copyright. On top of that, live music broadcast is way\nharder to police because you'd have to have listeners and feelers out on every single live stream ever in hopes that\nyou can find one that's violating your copyright to a couple dozen people. The amount of cost for ingest and\nidentifying those violations relative to the amount of money you can make from it is just so far off reality that none of\nthe companies would do it. What they would do instead is wait till the stream was done and then check the VOD looking\nfor signatures played back at like 500x speeds which is significantly cheaper.\nSo Twitch's solution was great. We introduced two audio tracks. One that is\nthe general stream audio and a separate track that includes music and other copyright and media. And then when the\nstream is done, we don't save the second audio track. We only save the first one. So, you might have noticed creators like\nPrimogen stream with music playing and then when you go watch the VOD, the music's gone. That's because he streams the music to that second track, which\nmakes it effectively impossible for you to be DMCA for playing music live on Twitch if you're using the two channels\nproperly. This is a technical workaround of a legal consideration to deal with\nthe [ __ ] that is all of this music licensing stuff. If I had to sit here and tell you who is more evil, Sunno or\nWarner, I would genuinely struggle. All of these people are evil. All of them\nsuck. None of them care about music or media. And neither of them are going to ever meaningfully pay artists what they\ndeserve. So is the new Warner. And rather than let beat Warner, they are\npartnering up. This makes me sick. With all of this said, I have a little bit of\nhope. Let's talk about how this can all be saved. How can we make media gen tools and technologies actually useful\nto professionals? I have two examples I want to show you guys. These examples come from the toolbox conversation\nearlier. Is the tool living outside of the box trying to replace it or is the tool yet another tool that goes inside\nof your toolbox and makes you more productive as a professional? The majority of these AI tools I've seen are trying to replace the toolbox and they\nall suck. But there have been a few that are working around it and trying to make useful tools within the toolbox instead.\nOne of them is a company I recently invested in, Co-create. Co-create is trying to augment video production.\nThey're not trying to do AI video generation. They're trying to use AI to help professional video teams get the\n[ __ ] out of the way faster. Traditional video editing software demands hours of manual work, sorting\nclips, syncing audio, and building rough cuts. It's why most creators spend more time organizing than actually creating.\nCo-create is a professional first video editing tool that automates the technical grind, letting you focus on storytelling, creative direction, and\nmaking content that connects. It's built to automate the ingest and organization step in professional video editing\nsoftware. When a team records dozens of hours of footage, and they want to use\nit for two to five minute video, even just sorting all of the footage is obnoxious, much less getting it into the\nright place in the editor so you could start doing the work. and then you realize you need to sync audio from another file from another person's\ncomputer. It's obnoxious once you're working in these bigger video production teams. There's a whole role category\ncalled assistant editor who are people that hope to be a real editor someday and they're in in the industry is to\njust sit there and organize files for the lead editor. There's often three to five of these assistant editors for the\none main editor doing the actual work of chopping inside of the video editing software. Automating this is something\nthat I'm actually quite excited about. I don't know how big the market will be, but I'm hyped somebody is doing it. They're going to actual professional\nvideo studios, sitting on the floor, watching what works, watching what's frustrating, watching what's blocking them, and then trying to build what's\nnecessary to automate those tedious parts. That's great. They're trying to fit in the toolbox. They're not trying to replace it because replacing it is\n[ __ ] stupid. And now, one more tangent to one of my favorite human beings, Daddy Kev. Kev is a legendary\nrecord producer and audio engineer. He has been mixing and mastering for decades now. Many of my favorite records\nwere mixed by Kev. Almost every Flying Lotus released a ton of Datalus, Bus\nDriver, Noage Thing, Sam I am, and more. That's just 2000s. In the 2010s, he\nreally popped off. Working on Bath's album Cerulean, which was groundbreaking at the time. I still love that album.\nBunch of my favorite John Wayne stuff. Both Stage Two and Stage Three by Mr. Wazo. Incredible projects. I have both\nof those on vinyl. Thundercat, who blew up soon after that\npoint, Igloo Ghost, who I also love dearly. This one DJ PayPal project that's been entirely forgotten to time,\neven though it's [ __ ] incredible. I love him so much. And so, so many more things. You can just scroll here and see\nstuff that you've you're into American hiphop or experimental electronic music at all, you've listened to multiple\nthings in this list for sure. Like, almost certainly. He knows his way around production. He\nknows his way around audio engineering, mixing, mastering, all of these things. Out of curiosity, he started to play\nwith AI stuff to see how it would handle mixing by giving it a song and asking\nit, \"What would you change?\" And the suggestions that it made were better than he expected. There was no way for\nthe AI to actually do the things that it was suggesting, but it would suggest decent things. And then when he looked\nat what all the AI media companies were doing and saw they were trying to replace the whole process, not augment\nthe part that he was focused on, he decided he wanted to fix that. Despite the fact that this man recently turned\n50, he decided he wanted to do this. He wanted to figure this out. So he started learning how to code so that he could\nbuild the thing he wants. And this is where my super spicy take comes in. Developers are too [ __ ] stupid to\nlearn about an industry they don't care about. Learning how to code is way easier than building an appreciation of\nthe depth of the music industry. I have way more faith that Daddy Kev here can\nteach himself enough about code to make something incredible than I have that a bunch of tech bros can make actually\nuseful tools for professional media creators. It's just laughable for me to think of like a tech bro understanding\nmusicians better than Kev does. That's just that's not a real world we live in. And that's why I think his current\nproject DEX suite is so cool. It's an XML builder and parser for music DEX\nfiles. It's a common format for music production software. He wants to parse\nthat data and use it for all sorts of different things. So he wrote a Rust core that works with Rust, of course,\nbut also TypeScript.js and Python so that people can use this to screw with\nthose files. And it's super super cool. This is also kind of his learning project. He's doing this to better\ndeepen his knowledge of software development and figure out where he can sneak in to build the things he wishes\nexisted. This is so damn cool. And the moment he tells me he wants to go all in on it and raise money, I'm going to\nwrite him quite the check because an expert in a field learning to code will perform significantly better than an\nexpert at code learning a field always. And this is how we win. We need to stop\nbeing distracted by these tech bros pretending they can automate the field because they can't. They don't understand it. That's not how this\nworks. They can't just show up not understanding music and replace it. We need to use this stuff to make better\ntools and to build things within our existing toolboxes that make us more productive as creators. Things like\nbetter captioning systems in our video editors. Things like background removal in our image editors. Things like shirt\nswapping, face swapping, and all those fun things in Photoshop. things like sample creators, loop creators, EQ\ntooling, and all this stuff in our music software. There's so much cool stuff that can happen within our tools if they\nare built to take advantage of the AI revolution happening right now. But instead, we're being told that we'll all\nbe replaced because everyone will be a musician in no time. Hopefully, this helps you better\nunderstand where I sit in this space. I think replacing our toolboxes with AI is\ncringe as [ __ ] I think people becoming creators because of AI instead of doing it because they actually care about the thing is cringe as [ __ ] If somebody\ngets into creating music or games or videos or something because AI made it slightly easier, then they got\nfrustrated, went and learned the real thing. That's kind of cool. That's the only upside I can think of is somebody trying to make their first ever song\nwith realizing how cringe it is and then going to learn to produce. That's cool, I guess. But like on this spectrum where\non one side we have code and on the other side we have videogen, pretty much everything past image gen stops being\nuseful in my opinion because it's so much harder to do anything with. Credit\nwhere it's due. I know Sunno is trying to make a more studioike product where instead of it just generating the whole\nsong, it will generate stems and let you work with them and organize them in a really crappy web app. Eh, cool. better\nthan what they were doing before. And if I can use this to just generate a stem or like create 15 unique samples and\npick one of the ones I like to go use in my real editor, maybe cool. Maybe.\nBut man, this just ain't it. I wish I felt\ndifferently. I do genuinely wish I saw more going on in the space that gave me hope that they're not just going to try\nand replace our toolboxes and the humans that make these things. But that's not what they're doing. They're not trying\nto make our jobs easier. They're claiming that our whole fields are stupid and they can automate it. And that's dumb. I got nothing else on this\none. Let me know in the comments if I'm insane and just a hater or if I'm kind of making good points here. Curious how y'all feel. And I'm sure that I'm not\ngoing to get flamed a ton for this. Until next time, peace nerds.\n"