Andrej Karpathy has spent years telling developers that the way they write software is about to change, and he popularized the term that now defines much of the debate about AI’s role in programming. The concept he called vibe coding, the practice of letting an AI model generate software from natural-language descriptions, drew both enthusiasm and ridicule when he introduced it in 2025. He is now back at a laboratory, working on the technology that could make the idea a permanent part of the profession.
Karpathy, who joined Anthropic this week to lead pre-training research, said in remarks accompanying the move that his passion for education remains undimmed and that he plans to return to AI teaching once he completes his initial research goals. The former Tesla autopilot chief has built a following of millions through his video lectures on neural networks, and his Eureka Labs, an education startup founded in 2024, was meant to bring AI tutoring to a wide audience. The Anthropic job, by his own account, is a detour, not a departure.
The move has been widely read as a statement about where the industry is heading. Karpathy left OpenAI in 2024, spent time building Eureka Labs, and has been the field’s most visible independent commentator since. His decision to return to a laboratory, and to one known for safety research rather than speed at all costs, was seen by industry observers as an endorsement of Anthropic’s approach to building AI systems that can be understood and controlled.
Anthropic’s researchers welcomed the hire with unusual public enthusiasm. Boris Cherny, who leads the company’s Claude Code project, the developer tool that has made Claude a fixture of software teams, said on a social platform that Karpathy’s experience in deep-learning architecture design would greatly advance the company’s work on interpretable AI systems. Cherny’s willingness to comment publicly on a new colleague, rare for the company’s technical leadership, signaled how much weight Anthropic places on Karpathy’s standing with developers.
The connection between the two men’s work is direct. Claude Code is the product that has taken vibe coding from a demonstration to a daily tool for millions of programmers, and Karpathy is now working on the models underneath it. His mandate, pre-training research aimed at recursive self-improvement, sits upstream of everything the developer tool does. If models can improve their own training, the tools built on them change in kind.
The safety dimension is what drew Karpathy, according to a person close to the company. Anthropic’s Claude models have consistently scored at the top of independent evaluations of safety and alignment, and the company has built its reputation on refusing to ship capabilities it cannot control. Karpathy has spoken openly about his own concerns, warning that the field’s trajectory could lead to systems whose behavior no one fully understands. Joining a laboratory built around that problem, rather than commenting on it from outside, is a decision that colleagues said was months in the making.
Neither side has disclosed the details of the arrangement. People familiar with the discussions said the initial focus will be on model safety mechanisms and on the developer toolchain, the two areas where Karpathy’s experience and Anthropic’s needs overlap most cleanly. The company has been expanding its pre-training team aggressively, and its data-center commitments suggest it intends to run frontier-scale training runs of its own, the kind of project that would use a researcher of Karpathy’s caliber.
For the wider industry, the hire is another data point in the talent reshuffling that has defined the past two years. Anthropic has pulled senior researchers from OpenAI, Google, and Meta, while losing few of its own, and it has turned that stability into a competitive advantage in recruiting. The addition of the field’s most recognizable teacher gives the company a public face that rivals, for all their resources, have struggled to match.
The education thread matters to how the hire is read. Karpathy’s videos have served as the entry point to deep learning for an entire generation of engineers, and his Eureka Labs project was designed to carry that mission beyond video lectures into interactive, AI-driven tutoring. His insistence, in remarks around the Anthropic announcement, that he will return to that work suggests he views his research stint as a way to deepen the understanding he eventually brings back to teaching. The industry’s most prominent educator is going back to the lab, in his telling, so that what he teaches next is worth teaching.
Anthropic’s culture has also changed in ways that suit him. The company has grown into a large organization with enterprise sales, government affairs, and a product portfolio, but its research division has retained the small-team structure that Karpathy has said he prefers. The laboratory’s willingness to let researchers publish openly, and its practice of giving senior hires substantial autonomy over their programs, were factors in his decision, people familiar with the process said.
Karpathy’s own plans, if he follows the pattern of his career, will keep changing. He has moved between industry and independence before, and his stated intention to return to education suggests he sees the Anthropic role as a chapter rather than a final destination. For now, the industry’s most prominent explainer has chosen to go quiet and build, and the models that emerge from his new team will say more than any of his videos.
