Anthropic hires OpenAI co-founder Andrej Karpathy to lead pre-training

At a glance:

• Andrej Karpathy, a founding member of OpenAI and former Tesla AI chief, joins Anthropic to lead its pre-training team.
• He will build a team using Claude to accelerate the compute-intensive core model development process.
• Separately, Anthropic has hired cybersecurity veteran Chris Rohlf to its frontier red team.

A pivotal hire for Anthropic's model development

Anthropic has secured a high-profile AI researcher in Andrej Karpathy, who announced his new role via social media on Tuesday. Karpathy, a co-founder of OpenAI and former head of AI at Tesla, confirmed he has joined the company to work on pre-training, the foundational phase where large-scale training runs instill a model like Claude with its core knowledge and capabilities. He will report to Nick Joseph, head of pre-training at Anthropic. This move places one of the field's most respected practitioners—known for bridging theoretical deep learning and large-scale training practice—at the heart of Anthropic's most resource-intensive operation.

Pre-training is not only technically demanding but also extraordinarily expensive, consuming vast amounts of compute. By tasking Karpathy with building a dedicated team focused on using Claude to accelerate pre-training research, Anthropic is signaling a strategic bet. The company appears to believe that AI-augmented research, rather than a pure hardware arms race, is the key to maintaining competitiveness with rivals like OpenAI and Google. Karpathy's unique profile—with deep roots in both academia-style research and industrial-scale deployment at Tesla—makes him an ideal architect for this approach. His return to the frontier lab ecosystem, after a year back at OpenAI and a subsequent stint building his own education-focused AI startup, underscores the intense demand for top-tier talent in foundational model development.

Karpathy's career arc: from OpenAI's origins to Tesla and back

The researcher's career has traced a significant path through the AI industry. He was a founding member of OpenAI in 2015, focusing on deep learning and computer vision before departing in 2017 to lead Tesla's rapidly evolving Full Self-Driving (FSD) and Autopilot programs. His return to OpenAI in 2023 was a notable homecoming, but he left again in 2024 to launch Eureka Labs, a startup aimed at applying AI assistants to education. His continued passion for teaching—evident in his popular "Neural Networks: Zero to Hero" online course and YouTube lectures—remains a constant. While he stated his commitment to resuming educational work "in time," his immediate future is now firmly at Anthropic. The transition raises questions about the status of Eureka Labs, which has seen few public updates since its announcement.

A parallel push to harden AI against cyber threats

Anthropic's announcement of Karpathy was accompanied by the separate hiring of Chris Rohlf to its frontier red team. Rohlf brings over two decades of cybersecurity experience, having served at Yahoo's renowned "The Paranoids" security unit and spending six years at Meta. Most recently a fellow at Georgetown's Center for Security and Emerging Technology working on the CyberAI project, Rohlf will now focus on stress-testing Anthropic's advanced models against severe, real-world threats. His hiring complements Anthropic's longstanding "AI safety" mandate with deep, practical adversarial expertise. In his announcement, Rohlf highlighted the transformative opportunity to improve cybersecurity with AI, stating he couldn't imagine a better team or moment to join. Together, the appointments of Karpathy and Rohlf represent a dual investment: one in pushing the boundaries of model capability, the other in ensuring those powerful systems are resilient against malicious use.

What this means for the AI landscape

For Anthropic, acquiring Karpathy is a clear statement of intent to lead in the foundational science of model creation. It suggests the company is prioritizing algorithmic and methodological innovation—potentially in efficiency, data curation, or training stability—over simply outspending competitors on GPU clusters. For the industry, it reinforces the narrative that the race for generative AI supremacy is as much about attracting legendary researchers as it is about scaling infrastructure. For Karpathy, the move offers a return to pure R&D at a company whose constitutional AI philosophy aligns with his own research interests, while providing a vast computational sandbox to test his ideas. The coming months will reveal how quickly he can assemble his team and what novel pre-training techniques emerge from his lab, potentially setting a new benchmark for efficiency and capability in the next generation of large language models.

Looking ahead: Key developments to watch

Observers should monitor several developments stemming from this news. First, the formation and early output of Karpathy's internal team at Anthropic will be critical; any publications or technical talks from the group could signal new directions in pre-training optimization. Second, the operational status of Eureka Labs remains a point of curiosity—will Karpathy maintain the venture as a side project, or will Anthropic's demands require his full focus? Third, Rohlf's integration into the red team and any subsequent publications on AI security vulnerabilities will be closely watched by the policy and safety communities. Finally, the competitive response from OpenAI and Google will be telling; both firms may accelerate their own AI-assisted research initiatives or pursue parallel high-profile hires in a bid to maintain their perceived edges in the foundational model race.