Ideas / From Intelligence to Agency
Intro to Large Language Models
Andrej Karpathy
Andrej Karpathy explains how language models learn, fail, and become something like a new operating system.
Karpathy explains training, transformers, scaling, alignment, tools, multimodality, reasoning, and security without requiring the audience to pretend these are simple subjects.
His description of the language model as a new kind of operating system supplies the bridge from specialized machine intuition to general-purpose computing through language.