Andrej Karpathy delivers 'Intro to Large Language Models' talk

Andrej Karpathy, formerly a founding member of OpenAI and director of AI at Tesla, published a one-hour, general-audience video explaining how large language models work, without assuming any technical background. It became one of the most widely shared public explainers of the period.

Karpathy’s central illustration reduced a model to two files: a large parameters file holding the neural network’s weights — he used Llama 2’s 70-billion-parameter version as an example, its weights occupying about 140GB on disk at two bytes per parameter — and a short run program, a few hundred lines of code, that executes the network using those weights. Given both files and no internet connection, he showed, the model was fully self-contained and could still generate text, making concrete what “training” and “inference” actually produced.

The talk moved from that foundation through training methods, into a section on model security covering jailbreaking, prompt injection and data poisoning, and closed with a speculative framing that circulated widely afterward: an “LLM-OS,” in which a language model functions as the kernel of a new kind of operating system, coordinating tools, a file system, memory and multimodal input and output much as a conventional operating system coordinates hardware resources.

The video’s significance lay in its role rather than in new research — it introduced no result of its own — but as a technically accurate, non-commercial explainer from a researcher with direct experience building the systems he was describing, it became a reference point widely recommended to policymakers, journalists and newcomers trying to understand what a language model actually was, at a moment when public discussion of the technology was heavily shaped by product marketing on one side and alarm about existential risk on the other.