Andrej Karpathy, one of OpenAI’s original co-founders and former Tesla AI director, has joined Anthropic. He’s not advising or taking a figurehead role — he’s on the pre-training team, starting immediately, reporting to Nicholas Joseph (himself an ex-OpenAI researcher).

His mandate is specific: use Claude to accelerate the research that builds the next Claude. Pre-training is the most compute-intensive phase of frontier model development, and the bet here is that AI-assisted research loops can compress the timeline meaningfully.

This isn’t just a theory Anthropic is testing. Karpathy already ran a version of it himself. In March he released an open-source project called autoresearch — give an AI agent a training script, a frozen eval metric, and a five-minute compute budget per experiment, then let it iterate overnight. It worked. The autonomous research loops Anthropic is now building around are loops Karpathy already validated in public.

He’ll work alongside John Schulman, another former OpenAI co-founder who joined Anthropic last year.

The hire comes during a busy month for Anthropic on the talent front. Ross Nordeen, a founding xAI member and ex-Tesla engineer, joined earlier in July. Chris Rohlf, a security engineer with over 20 years of experience including time at Meta, joined the frontier red team the same day as Karpathy.

Karpathy is probably best known outside research circles for coining the term “vibe coding” in early 2024. In his own announcement, he said he thinks the next few years at the frontier will be “especially formative” and that he plans to return to his education work in time.

For Anthropic, the timing matters. The past two months have been rough — a stretch of user complaints about Claude gave OpenAI room to push forward with GPT-4.5. Whether Karpathy’s arrival changes the trajectory of the next model is an open question, but it’s hard to read this as anything other than a serious signal about where Anthropic thinks the leverage is: not just raw compute, but using existing models to accelerate the research itself.