Why we're watching this: Karpathy is one of the most credible pre-training researchers alive. His mandate is to use Claude to accelerate the very research that produces Claude. That loop, AI improving its own training, is not a feature launch. It is a capability bet on where the frontier moves next.
Key Takeaways
- Andrej Karpathy joins Anthropic’s pre-training team this week, reporting to team lead Nick Joseph, to build a group focused on using Claude to accelerate pre-training research
- Karpathy co-founded OpenAI, led Tesla’s Full Self-Driving and Autopilot programs, and most recently founded Eureka Labs, an AI-in-education startup he says he plans to return to “in time”
- Separately, Chris Rohlf, a 20-year cybersecurity veteran from Meta and Yahoo’s “The Paranoids” team, has joined Anthropic’s Frontier Red Team
- Pre-training is the most compute-intensive phase of building a frontier model and the stage where core knowledge and capabilities are established
Andrej Karpathy has joined Anthropic, the AI safety company confirmed Monday, where he will lead a new team inside the pre-training organisation focused on using Claude to accelerate its own training research.
“I think the next few years at the frontier of LLMs will be especially formative,” Karpathy posted on X. “I am very excited to join the team here and get back to R&D.”
Personal update: I’ve joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.
– Andrej Karpathy (@karpathy) May 19, 2026
The hire is notable on two levels. First, Karpathy is one of a handful of researchers who can credibly operate at both the theoretical and large-scale infrastructure ends of pre-training. Second, his mandate is explicitly recursive: to deploy Claude as a tool for improving the research process that produces Claude.
Karpathy’s trajectory is well-documented in the field. He co-founded OpenAI, left in 2017 to lead Tesla’s Autopilot and Full Self-Driving programs, returned to OpenAI in 2023 for a year, then launched Eureka Labs, an education startup and the home of his Neural Networks: Zero to Hero course. He says he intends to return to that education work, though no timeline has been given and the status of Eureka Labs is unclear.
“I remain deeply passionate about education and plan to resume my work on it in time.” – Andrej Karpathy, joining Anthropic
Anthropic is also bringing on Chris Rohlf to its Frontier Red Team, which adversarially tests advanced AI models against severe threats. Rohlf spent six years at Meta’s security organisation and previously worked at Yahoo’s “The Paranoids,” one of the most respected internal cybersecurity teams in the industry. He was also a fellow at Georgetown’s Center for Security and Emerging Technology on the CyberAI project.
“We have a real opportunity in front of us to dramatically improve cyber security with AI,” Rohlf posted. “I can’t think of a better company or team to join at this critical moment in time.”
First week on the @AnthropicAI Frontier Red Team
The speed of AI progress is astounding. We have a real opportunity in front of us to dramatically improve cyber security with AI. I can’t think of a better company or team to join at this critical moment in time.
– chrisrohlf (@chrisrohlf) May 19, 2026
Recursive AI-assisted research is a compelling frame, but the outcomes are not guaranteed. The question is whether Claude can meaningfully accelerate the design of better pre-training runs, or whether human researcher intuition remains the binding constraint. Karpathy’s presence suggests Anthropic believes the former.
The two hires together signal where Anthropic is concentrating resources ahead of its next model generation: deeper pre-training capability and more rigorous adversarial safety testing, not just product surface area, a talent and capability consolidation Relve, an AI trends intelligence platform, is tracking as Anthropic’s pre-IPO push intensifies.
