Andrej Karpathy joins Anthropic after stints at OpenAI and Tesla

Why it matters

Frontier-lab pretraining expertise is scarce; Karpathy joining Anthropic’s pretraining team signals where the next gains in model quality may come from.

Andrej Karpathy is joining Anthropic (@claudeai). The move was flagged by Aligned News in a post on X, and Wikipedia lists Karpathy as part of Anthropic’s pretraining team in 2026. The post date Aligned News cites is May 19; RuntimeWire has not independently verified that announcement timing.

Karpathy is one of the most recognizable hands-on builders in modern AI. He co-founded OpenAI and later served as director of AI and Autopilot Vision at Tesla, roles that put him at the center of scaling deep learning systems for both research and production, according to Wikipedia. After leaving Tesla, he founded Eureka Labs in 2024 to focus on AI education, per the same page. That arc gives him unusually broad perspective across research labs, consumer-scale deployment, and teaching.

Why Anthropic, why now

Pretraining is the foundation-setting phase of large model development: it is where data pipelines, compute orchestration, and efficiency tricks compound into downstream capabilities. Joining Anthropic’s pretraining group signals Karpathy’s bet on where the next increments of model quality will be won. While his specific title and scope are not public in the sources cited, the pretraining remit sits close to the metal on data, scaling laws, and training systems.

For Anthropic, a seasoned operator who has shipped frontier-scale training runs at Tesla and worked on early OpenAI models brings practical instincts about tradeoffs under real constraints. For Karpathy, Anthropic offers a small cohort of peers tackling the same hard problems at the frontier. The circulation of senior technical talent among a handful of labs is not new, but each move subtly reweights where hard-won techniques and tastes for systems engineering accumulate.

The competitive backdrop

OpenAI, which Karpathy helped start, continues to ship ChatGPT and platform updates, reflecting a push to harden safety features and broaden consumer and developer utility. At Anthropic, the pretraining team is among the groups that determine the ceiling of each generation’s capability. Wikipedia’s accounting places Karpathy at all three phases of today’s AI stack: OpenAI’s general-purpose research, Tesla’s applied autonomy, and now Anthropic’s core training engine.

What we still do not know

• Announcement timing: Aligned News cites May 19 for Karpathy’s announcement, but that date is not independently corroborated in the materials provided here.
• Role specifics: beyond Wikipedia’s note that he is joining Anthropic’s pretraining team, no title or scope is detailed in the sources above.
• How this intersects with Eureka Labs: whether Karpathy continues any formal role there is not addressed in the sources.

What is clear is the throughline: Karpathy keeps choosing places where the frontier is defined by training data, compute budgets, and a culture that sweats systems details. Pretraining is where all three collide. If you are building in and around foundation models, expect more of the field’s tacit knowledge to concentrate inside a small set of teams that can run and learn from the biggest experiments.