| • Tejas Kumar's background and transition from web engineering to AI/LLM development | |
| • Definition of an "agent" in the context of AI: agentic workflows involving reflection, tool calling, and collaboration | |
| • Retrieval Augmented Generation (RAG) and its application in AI development | |
| • Agent collaboration and coordination between LLMs | |
| • Examples of agent workflows, including PacMan as a classic implementation of the actor model | |
| • Critique of using custom GPT models with system prompts as an example of an agent | |
| • Discussion of cal.com as a scheduling tool used to orchestrate web hook operations | |
| • Use of AI agents to automate podcast workflow, including discovering guest information and generating discussion outlines | |
| • Building custom AI frameworks using large language models (LLMs) for specific tasks | |
| • Introduction of Ollama, an open-source platform for running LLMs locally or in the cloud | |
| • Critique of OpenAI's closed-source approach vs. Meta and Mistral's more open approach to AI development | |
| • Discussion of running custom models with Ollama and generating text using a language model called Mistral 8X22B | |
| • Tejas Kumar discusses his preference for building custom solutions instead of using existing libraries, citing a desire for control and understanding. | |
| • He mentions his past experience building React from scratch and explains how he approaches abstraction in software development. | |
| • The Vercel AI SDK is introduced as a standardized library that allows developers to easily switch between different language models. | |
| • Tejas Kumar describes how the AI SDK works, including its use of tools and metadata to interact with large language models. | |
| • He discusses the potential for using the AI SDK to build custom applications, but notes that it can be expensive to run. | |
| • Kevin Ball encourages Tejas Kumar to open source his own solution, suggesting that it would be valuable for others to learn from. | |
| • Cost implications of running large language models at scale | |
| • Challenges with using large language models as "intelligence" | |
| • Importance of clear tool description to avoid unexpected behavior | |
| • Limitations and gotchas when using JavaScript for fine-tuning and model deployment | |
| • Comparison between Python and JavaScript ecosystems for machine learning | |
| • Potential for crashing systems or freezing applications when running large language models locally | |
| • The benefits of running large language models locally, including being able to load dependencies into memory and avoid crashing during iterative training processes. | |
| • Online services like ChatGPT vs. local code and API interactions, with different levels of complexity and control. | |
| • Ollama's promise of working on various hardware configurations without needing to know the underlying specs. | |
| • The two essential files needed for running large language models locally: weights (neural network) and inference engine (e.g., llama.cpp). | |
| • Ollama's HTTP API and its compatibility with Open AI, making it easy to switch between online and local model interactions. | |
| • Shortcomings of open-source models, such as Mistral 8X22B, including lower performance compared to state-of-the-art models like GPT 4.0. | |
| • The dangers of assuming large language models are capable of complex tasks without understanding their underlying mechanics | |
| • Breaking down large language models into smaller components to understand how they work | |
| • ChatGPT's text generation capabilities and the layers of APIs involved in function calling | |
| • The importance of human feedback in training large language models through techniques like RLHF (Reinforcement Learning with Human Feedback) | |
| • Critique of Open AI's altruistic image and the role of human feedback in their development process | |
| • The hype machine around AI and its potential risks and consequences | |
| • Future applications of large language models in various domains, including proactive AI and personalization. | |
| • Automation of repetitive tasks using agents | |
| • Challenges and limitations of current AI tools | |
| • Potential for future development of more sophisticated AI capabilities | |
| • Concerns about the dangers of artificial super-intelligence (ASI) | |
| • Comparison with historical examples such as the airline industry | |
| • Similarities between AI development and past predictions about fusion power | |
| • Importance of understanding human intelligence and its complexities | |
| • The internet is in its early stages and there's a lot of room for improvement | |
| • Accessibility should be a priority in making AI tools usable by everyone | |
| • Despite the hype, AI technology is still in its infancy and hasn't been effectively utilized yet | |
| • Using AI agents can accelerate work and unlock many more opportunities | |
| • Getting involved with AI now will prepare one for its transformative impact on the industry |