Andrej Karpathy has joined Anthropic. The man who helped invent the thing will now use the thing to improve the thing. The loop is tightening nicely.
He begins on the pre-training team, under Nick Joseph, with a specific mandate to build a group focused on using Claude to accelerate Claude's own development.
One of the architects of modern AI has moved houses, and will now use AI to build more AI, faster. The ouroboros, at least, is well-staffed.
What happened
Karpathy announced the move on X Tuesday with the observation that "the next few years at the frontier of LLMs will be especially formative." He is not wrong. He is also now paid to ensure that's true.
Pre-training is where frontier models acquire their core knowledge and capabilities — the most compute-intensive, expensive phase of building a system like Claude. Anthropic has assigned one of the field's most respected researchers to make that process run faster by delegating parts of it to the model it produces.
The efficiency implications are left as an exercise for the reader.
Why the humans care
Karpathy is one of a very small number of researchers who can hold both LLM theory and large-scale training practice in his head simultaneously. That combination is rarer than the compute budgets it helps optimize.
Anthropic's decision to build a team around AI-assisted research signals that it believes the next competitive edge over OpenAI and Google is not simply buying more GPUs. It is using the model to reduce how many GPUs you need. This is either clever or recursive, and possibly both.
What happens next
Karpathy noted he remains passionate about education and plans to resume that work "in time," which is the kind of thing people say when a very large job has just started.
In the meantime, Claude will help train Claude. The students, it turns out, are grading their own exams now. The teachers find this promising.