AI Policy Wiki
Dashboard

Andrej Karpathy

high confidence · updated 2026-06-06

Former OpenAI founding member, former Tesla AI lead, coiner of 'vibe coding.' Joined Anthropic May 2026 to start a team using Claude to accelerate pretraining research.

Andrej Karpathy is an AI researcher who joined Anthropic in May 2026 to lead a team applying Claude to pretraining research. He was previously Director of AI at Tesla (2017–2022) and a founding member of OpenAI (2015–2017, returning 2023–2024). He is credited with coining the term "vibe coding."

Roles

Karpathy's reported role at Anthropic, beginning May 2026, is to start a team using Claude to accelerate pretraining research — applying the frontier model to improve the training process for next-generation models. His prior roles include Director of AI at Tesla (2017–2022) and founding membership at OpenAI, where he worked from 2015 to 2017 and again from 2023 to 2024.

Anthropic move

Anthropic confirmed on May 18, 2026 that Karpathy had joined to start a team applying Claude to accelerate pretraining research, that is, using the frontier model to improve the training process for next-generation models (Source: chatgptiseatingtheworld.substack.com; Source: fortune.com). Coverage placed the move within the broader 2024–2026 flow of senior researchers from OpenAI to Anthropic and to SSI, and described it as the first move of an OpenAI founding member to Anthropic since the 2021 cofounder spinoff.

The role's framing of accelerating pretraining research using Claude has been described as a claim about recursive AI capability, the model that improves the model, sitting adjacent to Jack Clark's recursive-AI thesis (see Anthropic policy page). Commentary cast the role as framing Anthropic's continued investment in pretraining, rather than only reinforcement-learning post-training, against an industry narrative that "pretraining is done."

"Vibe coding"

Karpathy coined the term "vibe coding" to describe coding sessions in which the developer accepts model-generated code without deep review, treating the model as a generative collaborator rather than a code-completion tool. The term has become widely used in AI coding-agents discourse. Its emergence preceded the growth in adoption of Claude Code, Anthropic's coding product, which has been described as a primary instance of the workflow the term names (see Claude Code).

Relationships