A founding OpenAI researcher who had spent years teaching from the outside
Andrej Karpathy was a founding member of OpenAI before leaving in 2017 to lead Tesla's Autopilot and Full Self-Driving programs as senior director of AI. He returned to OpenAI for about a year in 2023 and 2024, then founded Eureka Labs, a startup focused on AI-driven education, which occupied him until this move. [3]
Between the Tesla and Anthropic chapters, Karpathy built a public following as an explainer of how large language models actually work, through lecture series, open-source teaching code, and long-form interviews, rather than through a lab role. His move to Anthropic put him back inside a frontier lab after that outside stretch. [3]
Pretraining, not an agents team
Karpathy joined Anthropic's pretraining team, the group responsible for the large training runs that give Claude its core capabilities, working under team lead Nick Joseph. His mandate is to build a new team focused on using Claude itself to accelerate pretraining research, rather than to build customer-facing agent products. [3]
Announcing the move, Karpathy wrote that he thought the next few years at the frontier of large language models would be especially formative and that he was excited to get back to research and development. He also said he remained passionate about education and planned to resume that work later. [3][4]
Set against his own decade-of-agents timeline
About seven months before this move, Karpathy told the Dwarkesh Podcast that he preferred calling the current period the decade of agents rather than the year of agents. He said the problems were tractable but still difficult, that agents were cognitively lacking on tasks requiring continual learning, multimodality, and computer use, and that working through those issues would take about a decade. [2]
Atlas interpretation: That timeline was about deployed agents doing open-ended work, not about pretraining. Karpathy's new role sits upstream of the gap he described: it targets the research process that produces frontier models, using an existing model as a tool inside that process, rather than shipping an agent meant to replace a knowledge worker. The move is consistent with someone who thinks the agent problem is still years from settled and who chose to work on the model-building pipeline instead of a product built on today's agents. [2][3]
Sources
- OpenAI co-founder Andrej Karpathy joins Anthropic
Axios · May 19, 2026
- Andrej Karpathy: It will take a decade to work through the issues with agents
Dwarkesh Podcast · Oct 17, 2025
- OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training team
TechCrunch · May 19, 2026
- Personal update: I've joined Anthropic
Andrej Karpathy · May 19, 2026