- Andrej Karpathy is joining Anthropic‘s pretraining team, the group that handles initial training of large AI models, per Axios.
- Karpathy will stand up his own pretraining team focused on using Claude to speed up pretraining research.
- He was an OpenAI founding-member, helped build Tesla Autopilot and Full Self-Driving, returned to OpenAI, then left for good in 2024.
- Most recently he ran Eureka Labs, an AI-education startup; he says he plans to return to that work at a later date.
What Happened
Prominent AI researcher Andrej Karpathy is joining Anthropic over a return to OpenAI, The Decoder reported on Monday, citing Axios. Karpathy is joining Anthropic’s pretraining team — the group that handles the initial training of large AI models. He will stand up his own pretraining team focused on using Claude to speed up pretraining research. The work is about building the strongest possible base model, which then gets fine-tuned with methods like reinforcement learning for specific tasks.
Why It Matters
Karpathy is one of the largest individual names in modern AI research. He was a founding-member at OpenAI in 2015, served as a core researcher there, left to lead Tesla’s Autopilot and Full Self-Driving teams, returned to OpenAI, and left for good in 2024 to launch Eureka Labs (an AI-education startup). His decision to rejoin frontier-LLM research at Anthropic rather than returning to OpenAI is a meaningful talent-and-narrative win for Anthropic and a corresponding loss for OpenAI.
In a post on X announcing the move, Karpathy said he’s excited to get back into research and development, calling the next few years at the frontier of large language models “especially formative.” He recently said he was “blown away” by the progress of agentic AI for coding after dismissing agentic capabilities just a few months before.
Technical Details
Pretraining is the most compute-intensive stage of large-language-model development. Anthropic’s pretraining team is responsible for the base capabilities of the Claude model family — Claude Opus, Sonnet, and Haiku tiers, plus the Claude Mythos Preview cyber-capability model. The hypothesis underpinning Karpathy’s new team is that AI models can help improve themselves at the pretraining stage — a key bet for whether AI progress compounds exponentially or hits a near-term ceiling.
Karpathy’s prior work includes seminal contributions to convolutional neural networks (Stanford CS231n), the original GPT-2 implementation, and a series of widely-followed YouTube videos walking through the math of modern transformer-based models. His Eureka Labs work will continue at a later date, per his own statement; he framed AI in education as a long-term commitment.
Who’s Affected
Anthropic gains a top-tier pretraining researcher and the narrative that the most-watched individual researcher returning to frontier LLM work chose Anthropic. OpenAI loses both the talent and the symbolic moment — Karpathy was on OpenAI’s founding cap-table-and-research team and his choice to skip a return is publicly readable. Tesla, where Karpathy previously led Autopilot, retains the work he completed there. The broader AI research community gains another visible signal that the centre of frontier-model gravity is fragmenting beyond OpenAI alone — Karpathy’s move follows Mira Murati (Thinking Machines Lab), Ilya Sutskever (Safe Superintelligence), Jan Leike (Anthropic), and Dario Amodei himself (founded Anthropic after OpenAI).
What’s Next
Anthropic has not disclosed timelines for Karpathy’s team build-out or specific pretraining research priorities. Expect Karpathy to publish progress updates through his usual public channels — X, his blog, and YouTube — over the next 6-12 months. His specific bet, that Claude can be used to accelerate pretraining research itself, will be testable as Anthropic’s next-generation models reach internal evaluation.
