Andrej Karpathy Returns To Big-Model Work At Anthropic

Anthropic taps the former OpenAI and Tesla AI lead to run pretraining research using its assistant; the company also adds cybersecurity veteran Chris Rohlf to its red team

Andrej Karpathy, a well-known figure in artificial intelligence who helped found OpenAI and later led AI development at Tesla, has started a new role at Anthropic focused on the pretraining phase for its Claude models. Karpathy announced his move on the social platform X this week and is joining a group led by Nick Joseph that oversees the large-scale training runs responsible for giving the models their baseline capabilities — a phase that industry observers say is both technically demanding and among the costliest parts of building advanced systems.

An Anthropic representative said Karpathy will stand up a team devoted to using Claude itself as a research tool to speed up and improve pretraining experiments. Analysts and former lab researchers see that as a strategic shift: rather than relying only on ever-larger compute budgets, companies are increasingly experimenting with model-assisted research to tighten development cycles. One analyst familiar with large-scale training estimated that pretraining can consume roughly one-fourth to one-third of a project’s total development outlay, making efficiency gains particularly valuable.

Karpathy’s résumé spans theory and production. He worked on deep learning and computer vision at OpenAI before leaving in 2017 for Tesla, where he oversaw Full Self-Driving and Autopilot programs through 2022. He returned to OpenAI for a year, then departed in 2024 to launch Eureka Labs, a startup aimed at bringing AI assistants into education. He has continued to teach through an online course and occasional video lectures, and has said publicly that he intends to return to educational projects when time allows. It is not yet clear whether he will remain active with Eureka Labs while at Anthropic.

Experts say Karpathy’s hire signals Anthropic’s intent to marry hands-on systems engineering with theoretical insight. “Someone who can translate mathematical ideas into scalable training pipelines is rare,” said an AI researcher unaffiliated with Anthropic. “Putting that skill set on pretraining — and using the models themselves to run experiments — could materially change how quickly models improve.” That approach also raises questions about reproducibility and oversight as teams lean on internal assistants for experimental design and debugging.

Anthropic also added Chris Rohlf to its frontier red team, a unit that probes advanced models for security weaknesses and misuse pathways. Rohlf brings more than two decades in cybersecurity, including a stint with Yahoo’s elite defensive group and several years at Meta, plus work at Georgetown’s Center for Security and Emerging Technology. In a post on X, Rohlf framed the move as an opportunity to apply AI to cyber defense, saying the technology could reshape how organizations detect and respond to threats.

Observers will be watching whether Anthropic’s new hires speed up model training or prompt tighter collaboration with external researchers. Requests for comment from Karpathy and Rohlf were not immediately returned.