Andrej Karpathy's 3.5-hour general-audience walkthrough of how large language models are actually built, pretraining, fine-tuning, and reinforcement learning with intuitive explanations of hallucinations, tool use, and "jagged intelligence.
youtube.com· 3 hr 31 min read· 2025Save this - and 500 more like it - to your library.
Linkwise remembers everything you read, and lets you chat with it.
Get Linkwise — free on iPhone
StratecheryBen Thompson on why thin clients are winning in an AI world, and what the collapse of thick client economics means for hardware, software, and the future of personal computing.
All non-trivial abstractions, to some degree, are leaky.
Every layer we stack between ourselves and the metal eventually breaks. The fix isn't a better abstraction—it's knowing what's underneath when it does.
youtube.comDwarkesh sits down with two Anthropic researchers - one scaling RL, one doing mechanistic interpretability - to trace how modern LLMs actually think. The "auditing game" section, where the interp team catches a deliberately-planted evil model in 90 minutes, is worth the whole thing.