Andrej Karpathy joins Anthropic’s pre‑training team

Andrej Karpathy announced on X that he began working at Anthropic on May 19, 2026, joining the company’s pre‑training team that builds the large‑scale runs powering Claude’s core knowledge [1]. The move places a rare bridge between LLM theory and massive compute practice directly into Anthropic’s R&D pipeline.

Karpathy, a founding research scientist at OpenAI, left the lab in 2017 to head Tesla’s Full‑Self‑Driving and Autopilot computer‑vision programs, stayed until 2022, and then returned to OpenAI for a brief stint before departing again in early 2024 to launch Eureka Labs, an AI‑focused education startup [1]. His new role puts him under team lead Nick Joseph, another ex‑OpenAI engineer who helped launch Anthropic, and tasks him with building a team that will use Claude itself to accelerate pre‑training research [1].

Anthropic’s spokesperson said the hire signals the company’s belief that AI‑assisted research, rather than raw compute alone, is the key to staying competitive with rivals such as OpenAI and Google [1]. By leveraging Karpathy’s expertise in scaling deep‑learning workloads, Anthropic hopes to improve the efficiency of Claude’s training cycles, a phase that consumes the bulk of the model’s development budget [1]. The company also announced the addition of Chris Rohlf to its frontier red‑team, underscoring a broader push to harden its models against emerging threats [1].

Industry observers have framed the hire as a major win in the ongoing talent war for top AI researchers [2]. Karpathy’s reputation for “vibe coding” and his prolific social‑media commentary have made him a high‑profile figure in the community, and his public endorsement of Anthropic could sway other engineers seeking impactful R&D environments [2]. While his education startup remains in limbo, Karpathy reiterated his passion for teaching and said he will resume that work when time permits [2].

The real question now is whether Karpathy’s presence will translate into measurable gains in Claude’s performance or cost‑efficiency, and how quickly Anthropic can leverage his experience to outpace OpenAI’s next‑generation models.