#education
-
How AI Agent Memory Works: MemGPT & Memory Types | Stanford CS329Z Lecture 7 with Michael Ryan
Michael Ryan's lecture for Stanford CS329Z (Engineering AI Agents, recorded 14 October 2026) sorts agent memory into four types borrowed from the CoALA paper. Working memory is the context window, episodic memory holds raw logs of past trajectories and conversations, semantic memory holds facts about the user and the world, and procedural memory holds skills. Memory can live in the weights (parametric) or be retrieved into the context, and Ryan's rule of thumb is context for small sets of facts and fine-tuning or RL for skills.
-
Stanford CS229 I Machine Learning I Building Large Language Models (LLMs)
Yann Dubois (Stanford PhD student, co-author of Alpaca and AlpacaFarm) gives a 1h44m guest lecture in CS229 (Summer 2024, 2.8M views) on building a ChatGPT-style model. He splits the work into five components (architecture, loss, data, evaluation, systems) and skips architecture on purpose: academia works on it, but "what matters in practice is mostly the three other topics".
-
How I use LLMs
Karpathy's 2h11m general-audience sequel to Deep Dive into LLMs (2.7M views, Feb 2025): a hands-on tour of the LLM app ecosystem (ChatGPT as the feature-rich incumbent, plus Claude, Gemini, Grok, Perplexity, DeepSeek, Le Chat), organised by what goes into the context window.
-
Deep Dive into LLMs like ChatGPT
Karpathy's 3.5-hour general-audience deep dive (9.7M views, Feb 2025) covers the full ChatGPT pipeline: pretraining (internet → tokens → next-token prediction), supervised fine-tuning (conversations → an assistant persona), and reinforcement learning (practice problems → emergent reasoning). Throughout, he maps the stages onto how children learn from textbooks: exposition, worked examples, practice problems.