Andrej Karpathy is getting back to the grind. The AI researcher confirmed on X, May 19, that he has joined Anthropic and will begin work this week on the company's pre-training teamโ€”the group responsible for the heavy compute runs that produce base versions of Claude before any fine-tuning occurs. His mandate: build a team focused on using Claude itself to accelerate pretraining research.

The Self-Training Loop

"Excited to welcome Andrej to the Pretraining team," wrote Nicholas Joseph, Anthropic's head of pre-training and himself a former OpenAI employee. "He'll be building a team focused on using Claude to accelerate pretraining research itself." That framing describes something increasingly common at frontier labs: models being used as internal tools to design experiments, generate training code, and analyze data. Karpathy's hire formalizes this approach at Anthropic's most fundamental level.

A Career Defined by Early Moves

Karpathy's trajectory reads like a map of the last decade in AI. One of OpenAI's founding members in 2015, he departed three years later to lead Tesla's Autopilot and Full Self-Driving programsโ€”the rare researcher who left an AI lab for applied robotics. He returned to OpenAI in mid-2023 for a one-year stint focused on artificial general intelligence research before leaving in February 2024 to launch Eureka Labs, an AI-native education startup that Anthropic says is being paused rather than shut down.

The OpenAI Brain Drain Continues

Karpathy's move fits a pattern. Co-founders Dario and Daniela Amodei left OpenAI to start Anthropic, along with chief scientist Jared Kaplan and several lead authors of GPT-2 and GPT-3. The trickle of mid-career researchers following them has not slowed. Industry observers see Karpathy as another signal that the center of gravity for pre-training research is shifting westward on the mapโ€”or at least across the bay.

Beyond Research: The Educator

Karpathy brings more than lab credentials. His free YouTube lectures building neural networks from scratch have attracted tens of millions of views, making him one of the most effective science communicators in AI. He also popularized "vibe coding" earlier this yearโ€”a term that captured how developers are increasingly collaborating with LLMs to write software. Whether he resumes education work through Eureka Labs once his Anthropic tenure stabilizes remains an open question.

Key Takeaways

  • Karpathy will lead a team using Claude to accelerate the training runs that produce base models
  • The hire formalizes a growing industry trend: frontier labs using their own models as research tools
  • Eureka Labs is paused, not deadโ€”Karpathy plans to return to his education startup eventually
  • Anthropic continues to attract senior talent from OpenAI in an increasingly competitive landscape

The Bottom Line

Karpathy's move signals that Anthropic is serious about closing the loopโ€”using its own models to make better models. If the self-training approach pays off, it could reshape how frontier labs approach research, making the gap between leaders and laggards harder to close.