data stringlengths 115 7.61k |
|---|
bmk#1476: just crawl the internet
Ravna#1831: i was just about to say that lol, gzipped context = free context enlargement
bmk#1476: for images, executables
bmk#1476: if it has bytes it goes in the pile
Daj#7482: "GPT Neo write Paperclip_Maximizer.exe"
bmk#1476: i cant wait for models that powerful lol
bmk#1476: unfort... |
SmartManoj#1319: Joined the server.
brianweet#9814: Joined the server.
Daj#7482: Hey @SmartManoj @brianweet ! Welcome to the Git Branch Jungle! See the channel topic for info and don't hesitate to ask questions!
Brian#0686: Joined the server.
Daj#7482: Hey @Brian ! Welcome to the Bug Farm! See the channel topic for inf... |
kindiana#1016: just due to kernel launching overhead and cache effects
kindiana#1016: theoretically its the same number of flops if you do bs=1 * 100 or bs=100
kindiana#1016: but the second will run a lot faster in practice
Commutative Conjecture#6969: i thought the main benefit of batch was parallelization
Commutative... |
CheshireCat#1425: Joined the server.
cogscides#0387: Joined the server.
Sid#2121: Hey @cogscides ! Welcome to the loss valley! Please check the channel description for more info on the project, and reach out if you have any questions 🙂
shawwn#3694: poor @CheshireCat wasn't welcomed
Sid#2121: i didn't seem to be able t... |
Daj#7482: yea but we need to haha
DR.PROACT#2111: I. Is there any way to get access to gpt 3 other than open ai's invite? II. Is anyone here interested in testing ideas related to research (I'm in medical academia).
Sid#2121: not the official gpt3, that's why we're trying to replicate it
Sid#2121: we'll be open sourcin... |
Louis#0144: I’m rly pro keeping GPT away from patient dada
Louis#0144: Data
archivus#7382: @DR.PROACT you an actual medical doctor?
DR.PROACT#2111: Yes. All data has patient informed consent for research use. The data is mainly symptom frequency of certain diseases. But the idea of having gpt-3 enabled would be maybe f... |
@Daj yeah tables are important. But other results are usually written in the manuscript.
DR.PROACT#2111: So it wouldn't be too crazy to think it may be able to surprise us just with that data.
Daj#7482: I'm sure it could surprise us, it's probably just very non trivial to figure out what of its output is useful and wha... |
Daj#7482: haha
bmk#1476: > **A2E is like lipofuscin in that both are substances that accumulate in cells impeding cell function**, but whereas lipofuscin accumulates as an intracellular byproduct of normal cellular metabolism, A2E accumulates as an extracellular product of light exposure.
bmk#1476: Copy pasting some an... |
archivus#7382: https://cdn.discordapp.com/attachments/729741769738158194/737722220251709480/twitter_Ed_W6H1WAAALk92.jpg
archivus#7382: Here is the input (the highlighted part only)
archivus#7382: And here was the generated part https://cdn.discordapp.com/attachments/729741769738158194/737722322198462484/twitter_Ed_W6H... |
archivus#7382: Temperature 0.1 https://cdn.discordapp.com/attachments/729741769738158194/737724284545531956/Screen_Shot_2020-07-28_at_8.32.42_PM.png
DR.PROACT#2111: I don't want to be biased
archivus#7382: close to the minimum
DR.PROACT#2111: But that statement just gave me a new idea
DR.PROACT#2111: On my research
arc... |
archivus#7382: Last proof you can make GPT-3 say anything https://cdn.discordapp.com/attachments/729741769738158194/737725473995620475/Screen_Shot_2020-07-28_at_8.37.23_PM.png
Sid#2121: also @DR.PROACT that is really cool, i wish i knew more about the topic. please keep us updated!
archivus#7382: Wikipedia has loads of... |
bmk#1476: (/s)
Deleted User#0000: i think roberta is not auto-regressive
DR.PROACT#2111: Not one person has suggested mosquito lifespan as a limiting important factor for outbreaks
bmk#1476: https://twitter.com/gwern/status/1284276584573214721
Deleted User#0000: @DR.PROACT we are still at the stage where we are trying ... |
bmk#1476: https://twitter.com/nicklovescode/status/1284050958977130497
bmk#1476: "You need to tell it what the AI is and is not capable. " i think this is the key
archivus#7382: let's try it out
archivus#7382: anyone want to give me QA I should add as a prompt?
Deleted User#0000: @bmk yea, that was a surprising finding... |
DR.PROACT#2111: Not novel per se. But novel within the context of mayaro virus.
DR.PROACT#2111: It's as if you had a specialist on your side.
archivus#7382: Fuck https://cdn.discordapp.com/attachments/729741769738158194/737734006321446953/Screen_Shot_2020-07-28_at_9.11.24_PM.png
DR.PROACT#2111: Lol
DR.PROACT#2111: It's... |
archivus#7382: Last one I promise - sorry I polluted general ❤️
archivus#7382: https://cdn.discordapp.com/attachments/729741769738158194/737735254114173089/Screen_Shot_2020-07-28_at_9.16.07_PM.png
bmk#1476: @DR.PROACT what is the dataset youre thinking of?
Deleted User#0000: @archivus need to be able to fine-tune GPT-... |
DR.PROACT#2111: Just do up-to-date and dynamed offline
Deleted User#0000: @DR.PROACT nice! didn't know about the torrents, thanks for the tip
DR.PROACT#2111: > @DR.PROACT nice! didn't know about the torrents, thanks for the tip
@Deleted User sure np
DR.PROACT#2111: I'm a Dr so let me know if I can help in that regard.
... |
DR.PROACT#2111: From the demos I've seen
DR.PROACT#2111: I can imagine so many applications
DR.PROACT#2111: And I'm not even in tech
DR.PROACT#2111: Pretty sure someone is already way ahead
Deleted User#0000: yea, it's still early. there's a lot of limitations to the tech still
Deleted User#0000: but the promise is rea... |
zphang#7252: "Luther" for short
shawwn#3694: who decided it was a good idea...? well, if you guys like it. *shrug*
Daj#7482: It was the most voted on and the main contributers were unanimous
Daj#7482: Sure just 3 to 2 votes but not my fault people didn't vote lol
Sid#2121: I vote we have a mobile name - keep changing o... |
Sid#2121: next week: CHONK.AI
Daj#7482: tbh I don't care about publicity _at all_
Sid#2121: i am willing to bet that not one single person has said 'libreAI' with their actual mouth
Daj#7482: And if anything, the name is mysterious and intriguing lol
shawwn#3694: lol fair enough
shawwn#3694: you'd be wrong SId
Sid#2121... |
zitterbewegung#4846: i think i did
zitterbewegung#4846: netai.pw
zitterbewegung#4846: maybe ill use netai for my podcast
zitterbewegung#4846: with me and my deep learning project
AI_WAIFU#2844: Joined the server.
archivus#7382: http://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html
archivus#7382: Relevant to ou... |
Sid#2121: also, he wants everyone to know
bmk#1476: Skip connections are just a special case of highway networks
Commutative Conjecture#6969: ic
Sid#2121: @Commutative Conjecture if you're interested in peeping the minor ML drama lol https://www.youtube.com/watch?v=HGYYEUSm-0Q&t=5786s skip to 1:03:00
bmk#1476: i feel s... |
gs://neo-datasets/openwebtext-documents/
gs://neo-datasets/openwebtext-fixed/
gs://neo-datasets/openwebtext-new/
gs://neo-datasets/openwebtext/```
bmk#1476: haha
Sid#2121: i guess the original?
Sid#2121: or fixed??
Sid#2121: ??
Sid#2121: lol
bmk#1476: isnt that the one with 1024 instead of 1025
Sid#2121: idk, best wait... |
bmk#1476: o
bmk#1476: ^ u dropped this
bmk#1476: wait
bmk#1476: both spellings are acceptable?
bmk#1476: huh
bmk#1476: nvm
Sid#2121: i was also gonna go for the o then double checked my spelling and saw no o
Sid#2121: i prefer the o
Sid#2121: adding it in
Commutative Conjecture#6969: @Sid / @bmk
https://en.wikipedia.o... |
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/737930826389520405/unknown.png
bmk#1476: these two diagrams capture the entire idea
bmk#1476: for more details see the paper
bmk#1476: you literally just take the state and add it back to itself again
bmk#1476: after a few layers
Commutative Conjectur... |
Watching it
(I always feel special when watching an unlisted video with 10 views)
bmk#1476: clicking through it very quick and that's a lot of other content too
Nerdimite#3840: > Hey @Nerdimite ! Welcome to the group composed entirely of one GPT-10 discord bot trying to ouroboros itself into existence! Check the channe... |
AI_WAIFU#2844: reverse mode automatic differentiation
AI_WAIFU#2844: basically
AI_WAIFU#2844: so yes
AI_WAIFU#2844: the algorithm was reinvented like a dozen times so it got a bunch of different names
Commutative Conjecture#6969: ic
Commutative Conjecture#6969: thx 🙂
adalbertobrant#7154: Joined the server.
Commutative... |
kindiana#1016: and you can attend to different relative positions with a linear combination of different frequencies
Sid#2121: Hey @adalbertobrant ! Welcome to the AI-lab-that-cannot-be-named! Check the channel description for more info 🙂
Commutative Conjecture#6969: @kindiana
I don't understand positional encoding
Co... |
Commutative Conjecture#6969: might be something vdumb, but where?
kindiana#1016: neural networks don't like inputs inputs that are not around mean 0 and std 1
Commutative Conjecture#6969: not sure why it matters, you can normalize it
kindiana#1016: it also doesn't play nice with dot product attention
kindiana#1016: so ... |
Sid#2121: https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/utils.py#L610 this is the thread you'll want to follow
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/737948719277801582/Screenshot_2020-07-29_at_10.24.36.png
Commutative Conjecture#6969: erg
kindiana#1016: looks ... |
DR.PROACT#2111: Morning guys
DR.PROACT#2111: Humanity is rooting for something productive to come out of here
DR.PROACT#2111: 👽
bmk#1476: Moin
TravellingSalesman#5222: What do you guys think about this recent post in LW about AI overhang https://www.lesswrong.com/posts/N6vZEnCn6A95Xn39p/are-we-in-an-ai-overhang
> An ... |
Sid#2121: one of our big TODOs on data is getting clean pdf to txt extraction
Sid#2121: we're also trying to figure out ways to get a better, cleaner dataset from CC data than OSCAR. But a lot of this stuff isn't so entry level.
bmk#1476: Another thing that needs attention is working out a good html->text pipeline
bmk#... |
bmk#1476: Can you post a short sample of the data as a gist @arfa
Alm#9130: the model would have to predict what happend in the video in order to predict the next reaction haha
Sid#2121: lmao
Sid#2121: https://cdn.discordapp.com/attachments/729741769738158194/738169751305584761/Screenshot_2020-07-30_at_01.02.46.png
Si... |
Deleted User#0000: so how can i help
Anders#1378: Joined the server.
aquajet#7800: Hello @Anders welcome to the Twitch chat automation service! Check the channel description for some info on the project, and please ask if you have any questions.
aquajet#7800: So rn for the model the largest goal is to make the model mo... |
aquajet#7800: It's not explicit funding afaik but the tpus are being used through their program
arfa#0882: But aside from that
arfa#0882: GCP bucket costs, VM costs, Hetzner costs, etc.
aquajet#7800: donations i think
Sid#2121: couple private donations. we plan on setting up a patreon at some point too
arfa#0882: So ba... |
Isaac McHorse#2007: WHY ARE YOU BEING DISTRACTED? YOU CAN'T GET ANYTHING DONE LIKE THAT.
arfa#0882: You guys end up with a model that you can't run inference on without free TPUs which will be in high demand after you get media attention and everyone learns about TFRC and signs up :KEKW:
bmk#1476: I mean dajs original ... |
bmk#1476: meh we dont need 999999 uptime
bmk#1476: i mean care to suggest a better solution that doesnt cost 100x more?
arfa#0882: Well, if you end up in the situation Tensorfork is in, you might have zero uptime for weeks on end
bmk#1476: ?
bmk#1476: what situation is that
arfa#0882: We haven't been able to create any... |
aquajet#7800: my context window is 2
arfa#0882: Basically they said "You guys still have quota, but we're near 100% capacity in eu-west-4" which doesn't answer why you guys can create TPU pods and we can't
bmk#1476: o.O
arfa#0882: And I think shawwn is afraid to push the issue because he thinks we might've been manuall... |
bmk#1476: also that reminds me, @aquajet we probably want the webserver set up such that we can seamlessly pivot to a new dataset (i.e after we get the whole multilingual pipeline worked out) while requiring as little action on the parts of server owners (without doing arb code exec ofc)
aquajet#7800: Yep
aquajet#7800:... |
aquajet#7800: Yikes
bmk#1476: looks like we need to start from scratch
bmk#1476: https://docs.google.com/spreadsheets/d/1k8G2M5RxEwMrQuOwj-Wx-XxFBdZqyGKFGM6nPDyr-MQ/edit?usp=sharing
bmk#1476: heres a spreadsheet
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/738646020077912094/unknown.png
bmk#1476... |
bmk#1476: awesome
bmk#1476: i also have to sleep soon
adamb#9760: well like i said, it's just sitting idle atm
adamb#9760: so how do i talk to about putting it to work?
Daj#7482: I think @bmk is still our data tsar atm
bmk#1476: ok so
bmk#1476: here's the thing
bmk#1476: we have a current data pipeline
bmk#1476: we're ... |
adamb#9760: ok
adamb#9760: so what's the workflow
bmk#1476: you clone a repo and run a docker command
bmk#1476: we're currently working on a webserver to coordinate work so maybe we want to wait for that first
adamb#9760: where do the results get sent?
adamb#9760: are there any creds i also need?
bmk#1476: so we dont h... |
adamb#9760: when i was working on this i was extremely interested in enabling workflows for huge models trained on huge data
adamb#9760: but no one was working on big enough models outside of google yet, lol
Daj#7482: Well, if it works, this would be a great usecase
adamb#9760: so i put it down once it was clear that t... |
Deleted User#0000: say the word 'cat' is given to the transformer
Deleted User#0000: first you break it into 'c', 'a', 't'
Deleted User#0000: each of those tokens have a unique id, say 3, 0, 21
Deleted User#0000: then you fetch rows 3, 0, 21 from the embedding matrix above
Deleted User#0000: that's it..
Deleted User#00... |
Louis#0144: Well bilstm is contextualized too
Louis#0144: That’s not a fair comparison
Deleted User#0000: ah yea, ELMO started it all i think
Louis#0144: ELMO has good contextualization
Deleted User#0000: or ULMfit, i forget which
Deleted User#0000: man, bilstm's, those were the days..
Louis#0144: I miss them
Louis#014... |
Deleted User#0000: good to know!
Louis#0144: Usually people use residual gated graph neural networks
Louis#0144: GCNs are good when you need to represent lots of sparse knowledge
Louis#0144: Like multi doc QA
Louis#0144: A transformer is just a dense GCN
zphang#7252: is that on the hotpotqa leaderboard? so many anonymo... |
Louis#0144: Can I show my paper if I get accepted
zphang#7252: which conf?
Louis#0144: EMNLP
zphang#7252: aha
star#5322: Does someone have a two sentence of what a GCN is?
zphang#7252: I'd be interested to read your stuff
Louis#0144: CNNs... but for graphs
Louis#0144: Literally tho
Louis#0144: Laplacian smoothing is co... |
Louis#0144: @arvindr9
Louis#0144: hi
arvindr9#4837: Lol hi
E McNeill#1259: Joined the server.
bhauth#7283: Joined the server.
Daj#7482: Hey @technologiclee @arvindr9 @E McNeill @bhauth ! Welcome to the Mesa Optimizer Removal Facility! Check the channel topic for info and don't hesitate to ask questions!
Commutative Con... |
Commutative Conjecture#6969: > why can't transformers learn grammar? it seems like a pretty ideal case for multi headed attention (i.e. one head attends to the tense of the sentence, one head attends to the subject etc), and the output layers can take information from that to synthesize the next most likely word(piece)... |
Louis#0144: For the reasons mentioned above
Louis#0144: You only really need local attention
Louis#0144: And you only really need a few general rules + some random memorized exceptions
Commutative Conjecture#6969: > Modern LMs struggle very very little with grammar
@Louis
By modern, do you mean, NNs based?
Commutative ... |
Louis#0144: Yeah
Louis#0144: I’m not convinced grammar is really an issue
Louis#0144: The main issue that faces LMs right now is causality imo
Louis#0144: And coherency
Louis#0144: Also bigger context windows would be nice
Louis#0144: Lmao
Commutative Conjecture#6969: > Yeah
@Louis
Link?
Louis#0144: https://www.researc... |
Louis#0144: That’s an issue with coherency imo
Louis#0144: But yeah
Louis#0144: No idea why
Commutative Conjecture#6969: > https://www.researchgate.net/publication/239556866_A_Rule-Based_Style_and_Grammar_Checker
@Louis
Can't see the abstract, but grammar checking has nothing to do with grammar learning
Daj#7482: I thi... |
Daj#7482: When have my custom introductions ever been inaccurate? hahah
Deleted User#0000: @Commutative Conjecture https://arxiv.org/abs/1906.04341 there's a lot of papers along this theme
Deleted User#0000: https://www.pnas.org/content/early/2020/06/02/1907367117
Deleted User#0000: Chris Manning himself
Commutative Co... |
bmk#1476: https://twitter.com/nabla_theta/status/1289977562882424838
star#5322: yeah the "whether it understands" point seems, completely an airball to me
bmk#1476: i know gary marcus of all people probably isnt going to change his mind but hey
star#5322: I liked gwern's point about like, sifting, idr which post it was... |
bmk#1476: absolutely yes
bmk#1476: i cite gwern everywhere lol
Daj#7482: It takes me a while, but yes haha
bmk#1476: gwern put this so much more elegantly than me lol
bmk#1476: for reference, here's a snippet from my current draft:
bmk#1476: https://cdn.discordapp.com/attachments/729741769738158194/739539314454560909/... |
thenightocean#6100: anyway, hello everyone!
Commutative Conjecture#6969: Hi
Commutative Conjecture#6969: @miguelos
Hi, were you on the LW Discord server at some point?
miguelos#7956: Yes
Commutative Conjecture#6969: ok
Commutative Conjecture#6969: I was wondering whether that was you when you rooted for the universal k... |
Noa Nabeshima#0290: Actually it's unclear how the model would learn
Noa Nabeshima#0290: er, what it would learn from the BPE-dropout
Noa Nabeshima#0290: to me, I haven't read the papers
Deleted User#0000: Joined the server.
Noa Nabeshima#0290: By restricting a model with BPE-dropout to small length tokens for generatio... |
Louis#0144: Oh
Commutative Conjecture#6969: rats
miguelos#7956: GPT-3 is so dumb that it doesn't know what GPT-3 is.
mshang#7454: Joined the server.
Kazumi#1297: I bet GPT-3 wasn't in the dataset that GPT-3 was trained on
thenightocean#6100: So if I understand correctly, you guys think gpt-neo might advance faster and ... |
Daj#7482: > Can GPT-3 detect that text is an output of GPT-3? I wonder how it would play out if people would scrape data from the internet where content marketers would be using GPT-3 to create articles. Feedback loop?
@Kacper Wikieł iirc OA did some research on this for GPT2, haven't really followed for GPT3 because I... |
Daj#7482: Hey @Aran Komatsuzaki ! Welcome to the FOOMigation Tent! (I'm sorry) Check the description for info and don't hesitate to ask questions!
Aran Komatsuzaki#5714: Joined the server.
AI_WAIFU#2844: I think we can scale far higher than 10T with appropriate architectural improvements. Gshard is a 0.6T model and it ... |
bmk#1476: Does aws do that?
bmk#1476: Huh
adamb#9760: lambda is an extremely impressive resource for burst CPU
adamb#9760: https://github.com/StanfordSNR/gg
bmk#1476: Well
AI_WAIFU#2844: No, is it a private repo?
bmk#1476: It's not so much burst
Daj#7482: Yes repo is private until release
adamb#9760: get 1000s of cores... |
AI_WAIFU#2844: https://github.com/AI-WAIFU
Daj#7482: invite sent
Daj#7482: Most Mesh tensorflow discussion happens in #gpt-neox-devs
bmk#1476: Also we need to ingress 3.5pb
bmk#1476: I don't think that's happening in an hour
AI_WAIFU#2844: Great! I'll start crawling around the code base. If I've got questions I'll brin... |
AI_WAIFU#2844: Another thing I was thinking of trying, but haven't had time to, is just dumb gradient boosting.
AI_WAIFU#2844: It's not parallel, but it does get around the problem of running out of ram.
AI_WAIFU#2844: You train one model, then you save the logits of that model on the training data, and train the next ... |
of model parallelism within each matrix multiply and model parallelism across the layers of the network. All models
were trained on V100 GPU’s on part of a high-bandwidth cluster provided by Microsoft."
bmk#1476: They were horribly vague about it as usual
bmk#1476: Yeah that
bmk#1476: Basically mtf as we have it right ... |
bmk#1476: so
bmk#1476: that's not *that* many weeks
miguelos#7956: You mean Neo GPT-3, or OpenAI releasing GPT-3?
bmk#1476: neogpt3
bmk#1476: or opengpt3
bmk#1476: or whatever the heck we decide to call it
bmk#1476: oa wont releast gpt3 lol
miguelos#7956: What's the next best easily accessible model?
bmk#1476: or if th... |
miguelos#7956: What's the biggest time/resource sink, in order?
miguelos#7956: 1. Training (6 months)
2. Collecting data (1 month)
3. Fixing code (1 month)
miguelos#7956: ?
bmk#1476: look if you have a really good idea you can try asking OA
bmk#1476: we're all volunteers we have no time frame
bmk#1476: we do stuff when... |
miguelos#7956: Do we have enough, bit in the wrong format (html)? Or do we need a lot of effort to crawl for more data?
miguelos#7956: I speak French.
bmk#1476: awesome
bmk#1476: the plan rn is:
bmk#1476: making multilingual dataset to benchmark htmltotext converters -> tuning htmltotext converters and using the benchm... |
bmk#1476: that would be very helpful
miguelos#7956: Cleaning data? What are we talking about? Removing some kind of sensitive data, or make it all into a uniform format that won't confuse it too much?
bmk#1476: the latter
miguelos#7956: I'm surprised getting a lot of training data isn't a solved problem.
bmk#1476: im s... |
miguelos#7956: I also found an overview of GPT-3 dataset. Are the stated tokens the number of words? Are we trying to match 1:1 what they had?
miguelos#7956: @Alm: I'm looking to do context-free natural language parsing.
star#5322: I don't think GPT-2 is that crazy expensive to run, maybe I'm off my ass but doesn't it ... |
star#5322: 60 or 70 years ago we thought you could build AI in a summer, but that turned out to be hubris.
miguelos#7956: Depends on what AI means.
miguelos#7956: But surely all of these people were trying to workaround the lack of compute. That's the only sane reason for symbolic AI.
miguelos#7956: But to imagine the ... |
aquajet#7800: Hello @SCnightmare! Welcome to the AGI Fire Alarm Factory! Check out the channel description and pinned posts for information about our current projects and feel free to ask any questions
Noa Nabeshima#0290: > But will it scale?
I'm actually curious about this
star#5322: I'm also pretty curious. Didn't me... |
2020-08-03: I spent $2
2020-08-04: I lost $8
Today's balance: $5
Yesterday's balance: ___
kindiana#1016: trained with the right data, certainly, unsure if openai's gpt3 weights/training data can do it though
miguelos#7956: It seems like it would understand that earned/made and spent/lost are somewhat equivalent. It ca... |
> 2020-08-02: I made $10
> 2020-08-03: I spent $2
> 2020-08-04: I lost $8
>
> Today's balance: $5
> Yesterday's balance: ___
@miguelos At 0 temperature GPT3 gives $10 dollars
miguelos#7956: @Noa Nabeshima 0 temperature?
Sid#2121: @miguelos temperature is a setting when you're sampling from the model. On a very basic l... |
Louis#0144: yeah most people use VAEs i guess
Louis#0144: but this is a question about AEs
Louis#0144: like boring naive AEs
star#5322: cause vanilla AEs aren't forced to use their latent space in a good way so there's kinda no way to sample
star#5322: sure, I guess that's right then?
star#5322: But not necessarily, li... |
star#5322: and it might be hard to incrementally co-evolve an encoder and a decoder that do the right thing
Daj#7482: I'm in general with you on never assuming simplicity in learned solutions
Daj#7482: But that paper was interesting
star#5322: I think double descent is one of the most tantalizing pieces of evidence sha... |
star#5322: yup
Daj#7482: My own completely baseless (but testable) theory is that it relates to the microjitters of loss between individual batches
star#5322: but I would hesitate to jump from very broad claims like "for some reason, we seem to repeatedly observe DL generalizing better than we'd possibly expect" to ver... |
Daj#7482: Exactly haha
star#5322: it seemed like there was already plenty of evidence his points don't have that kind of thought behind them, sadly
star#5322: and are more like reflexive reactions/dismissals
Daj#7482: It's fascinating how there's always someone to fill every intellectual sophism niche
star#5322: Also s... |
Daj#7482: Yea
star#5322: and models don't inference differently on different orders of data, right?
Daj#7482: I'm implying something weird is happening that happens across timesteps
Daj#7482: > and models don't inference differently on different orders of data, right?
@star But they will have different intermediate gra... |
star#5322: one way to model the effect of minibatch SGD vs. full dataset SGD is just that randomness is added to the gradient, which is a thing we have evidence is good anyway
star#5322: but we don't know if there's something more subtle going on or that's just it
bmk#1476: full dataset SGD is batch SGD
star#5322: yeah... |
Daj#7482: Memory constraints
bmk#1476: both of the above
star#5322: I don't understand why a microbatch is related to a memory constraint
bmk#1476: grad updates necessitate allreduces
bmk#1476: allreduces are evil
star#5322: and slow
star#5322: yeah I understand the expensive part
star#5322: just not the mem part
bmk#1... |
star#5322: yeah I don't think that more batch is always better
bmk#1476: nns are black magic
star#5322: but who knows
bmk#1476: maybe some noise good, too much/no noise bad
star#5322: maybe small model/dataset/??? needs more regularization and grad-noise is regularization
Daj#7482: I have some vague intuition that big ... |
star#5322: I agree shawwn
star#5322: Or, I think theoretical models would be *very very valuable* but we're nowhere near close to having a viable theoretical model for most of this
Daj#7482: If Newton could be an alchemist calculus-user during a pandemic than so can we!
star#5322: so we need lots more empirical data an... |
Daj#7482: I was in the MIRI camp of "understanding intelligence from the ground up" (apologies if this is not accurate), right until sometime between GPT2 and 3
star#5322: and we don't understand NNs well enough empirically to know how to make our theory better
bmk#1476: so we need to study nns empirically
Daj#7482: I ... |
shawwn#3694: And we only have v2 pods
Daj#7482: Yes I should taboo emergent lol
Daj#7482: Huh you're right shawn that should work
bmk#1476: i feel like gptx has a solid chance of actually being useful for agi
bmk#1476: so studying it in depth would be interesting
Daj#7482: > yeah I view the "understand NNs from the gro... |
star#5322: So while noting that the public MIRI canon is large, and I am not familiar with the majority of it, the "agent that takes only optimal actions" is not, to my knowledge, a primary object of study
star#5322: for the obvious reasons you say
bmk#1476: in all likelihood we'll build a very flawed paperclip maximiz... |
bmk#1476: (shitty pun)
Daj#7482: ahh
star#5322: > I think non-MIRI-style AGIs will be incredibly strong, but not interpretable
@Daj fwiw I don't really know what you mean by "MIRI-style AGIs" but usually that kind of phrasing stems from a pretty substantial misunderstanding of the MIRI project
(which, while factually t... |
star#5322: the straw man of the confused position is something more like "MIRI studies Decision Theory and Logical Counterfactuals because MIRI thinks that if we just put an AI together using enough Decision Theory Juice and had the right Counterfactual module and two other modules then the AI would 'work right'"
Daj#7... |
star#5322: Writing a compiler in Haskell is like, the most canonical and amazing usecase
star#5322: writing a parser, by itself, is also an extremely good project
Daj#7482: Haskell is one of those languages that is optimized for how smug you feel after you make something work
star#5322: hmmmmmmmmmmmmm
bmk#1476: hmm. i ... |
Daj#7482: I think getting why FP is fun is like asking why math is fun
bmk#1476: i have no idea where to even begin
star#5322: and I'll find a link in a sec
star#5322: https://www.cis.upenn.edu/~cis194/spring13/
Daj#7482: Oh yeah there's this great Haskell course where I built a parser
Daj#7482: That's the one!
star#53... |
Daj#7482: > cause all of like Coq LEAN Agda are based on that concept you know?
@star Yes I worded it poorly, that's why I switched to Lean
star#5322: linky? I don't think I've read that
Daj#7482: I started with Coq because it's the first one I knew that di dthat
star#5322: Oh gotcha
Daj#7482: Then I switched to Lean b... |
Daj#7482: I've heard good things about it but I'm not quite at the level to grok it yet
star#5322: like I said, probably would have happened eventually if I got better at things
star#5322: Idr Agda and LEAN being that . . . different, tbh?
star#5322: the very good book to learn agda with is https://plfa.github.io/
star... |
Daj#7482: I just have little excuses to use it
star#5322: but purity / immutability is pretty sweet
star#5322: very fair
Daj#7482: It seems AGI is going to be written in horrific Python spaghetti code
Daj#7482: and inscrutable C++ libraries beneath lol
star#5322: Agda has moments of brilliance though . . . I've been fo... |
Daj#7482: That's a pretty high risk bet that I think is >50%
star#5322: that is an even stronger claim
Daj#7482: yep I'd say I'm maybe like 60% on it
star#5322: TF didn't exist what, 7 years ago?
bmk#1476: Mary Garcus will be there to say that the paperclip machine isnt actually agi because all it does is make papercli... |
Daj#7482: Turns out all EY's concerns about firealarms were wrong, AGI's are very willing to tell us exactly what terrible things they plan
asparagui#6391: life goals
Daj#7482: "GPT5, how can I align you?"
"Here's a MIRI paper I calculate they would have discovered in 2200: ..."
bmk#1476: https://cdn.discordapp.com/at... |
AI_WAIFU#2844: The brown line looks a lot like double desent.
Daj#7482: I will definitely take a closer look at what you just said in the morning when I can think hah
Daj#7482: Thanks bmk for giving me the kind of high quality content my sleepy brain wanted to see
AI_WAIFU#2844: That sounds entirely reasonable.
zphang#... |
Aran Komatsuzaki#5714: Also, given how DeLighT works across hidden dimension, not length dimension, it's in a competition with MoE. So, there's also an uncertainty about whether it can meaningfully coexist with MoE.
Aran Komatsuzaki#5714: The author said it's possible to make it faster using custom CUDA, but I'm not su... |
Louis#0144: lmao
Louis#0144: @Aran Komatsuzaki
Louis#0144: Im w/ Riedl
Aran Komatsuzaki#5714: Actually not at GT this semester because of COVID-19. I'm having a leave of absence for this semester to stay in my home country.
Aran Komatsuzaki#5714: Riedl is probably the best prof for this kind of thing. He was the most k... |
Aran Komatsuzaki#5714: I like that
Aran Komatsuzaki#5714: I mean it makes things a bit more elegant
Louis#0144: I did topology too in my undergrad
Louis#0144: Pure math with a specialization in alg top
Louis#0144: Now i might be doing a fellowship at ETH with a few TDA people
Louis#0144: Rly excited
Aran Komatsuzaki#57... |
thenightocean#6100: quite a bold assumption
Louis#0144: Oh my god
Louis#0144: can we get a tinfoil hat emote@
Louis#0144: Please
Aran Komatsuzaki#5714: neural net, being an universal approximator, can imitate any continuous function, including the ones that appear in physics. so, you can find an analogy to whatever con... |
star#5322: lol
star#5322: Nate says that after the intro, there's two main sections: explaining HoTT like it's a type theory, and then explaining HoTT like it's a homotopy theory. So the difficulty curve drastically increases if you're a type theorist that doesn't know much homotopy, but starts out really hard and then... |
Louis#0144: ❤️
star#5322: I'm not learning it for any particular relation to ML. Just for interest as an FP & math nerd.
Louis#0144: FP?
Louis#0144: oh functional programming
Louis#0144: meh
Louis#0144: idk man, homotopy type theory just doesnt seem all too useful ngl
Louis#0144: I know saying that makes a bunch of alg... |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.