Preface

Andrej Karpathy, a prominent AI researcher known for his roles at OpenAI and Tesla, has taken a new position at Anthropic to work on pre-training for large language models. This move highlights a continuing trend in the AI industry where experienced researchers shuttle between leading labs, bringing both practical engineering experience and theoretical insight. The purpose of this article is to summarize the key facts of Karpathy’s new role, explain the significance of pre-training in modern model development, and note related hires that signal how Anthropic is strengthening both its research and security capabilities.

Lazy bag

Karpathy has joined Anthropic to lead work on pre-training, the compute-heavy phase that gives models their foundational abilities. Anthropic is forming a team to use Claude to accelerate research, emphasizing AI-assisted discovery over raw compute. The company also added cybersecurity veteran Chris Rohlf to its frontier red team.

Main Body

Andrej Karpathy, who previously co-founded and worked at OpenAI and later led AI efforts at Tesla, announced that he has joined Anthropic. His move was confirmed by a post on X, where he expressed enthusiasm for returning to research and development at the frontier of large language models (LLMs). Karpathy began work at Anthropic this week and is contributing to the pre-training team under the leadership of Nick Joseph.

Pre-training is a central phase in building advanced LLMs: it consists of large-scale training runs that imbue models with general knowledge and language capabilities. These runs are both expensive and computationally demanding, often representing a sizeable portion of the resources required to develop a frontier model. Anthropic described pre-training as responsible for giving Claude — the company’s flagship model — its core abilities.

Anthropic told media that Karpathy will form a team focused on using Claude itself to accelerate pre-training research. This approach signals a strategic emphasis on AI-assisted research workflows, where models help researchers iterate more quickly, rather than relying solely on the brute force of increased compute. By putting a researcher with Karpathy’s blend of theoretical knowledge and hands-on systems experience onto this task, Anthropic appears to be betting that smarter tooling and model-driven experimentation can be a competitive differentiator versus other industry leaders.

Karpathy’s background bridges academic deep learning research and applied, large-scale system building. After leaving OpenAI in 2017, he joined Tesla to lead Full Self-Driving (FSD) and Autopilot initiatives, managing teams and production-scale ML efforts. He returned to OpenAI for a year before departing in 2024 to found Eureka Labs, a startup aimed at applying AI assistants in education. Public updates about Eureka Labs have been sparse, and Karpathy has not clarified whether he will continue to be involved with the startup while at Anthropic.

In addition to his industry work, Karpathy is known for teaching. He created an online course titled "Neural Networks: Zero to Hero," which guides learners through building neural networks from scratch, and he maintains a YouTube channel where he posts lectures and explanations about LLMs and related topics. He has stated that he remains passionate about education and intends to return to those efforts when possible.

Separately, Anthropic announced the hire of Chris Rohlf to its frontier red team. The red team’s mission is to stress-test advanced AI models against strong adversarial and security threats. Rohlf brings more than two decades of cybersecurity experience, including time with Yahoo’s renowned cybersecurity group and subsequent years at Meta. He also served as a fellow at Georgetown’s Center for Security and Emerging Technology, working on CyberAI initiatives.

Rohlf commented that the industry now has a clear opportunity to improve cybersecurity through AI and that he sees Anthropic as a fitting place to pursue that goal. His addition to the red team complements the research hires by strengthening the company’s ability to evaluate and harden its models against misuse and technical vulnerabilities.

Taken together, these hires illustrate Anthropic’s dual focus: pushing forward model development while also investing in rigorous evaluation and security. Bringing in researchers who can connect theoretical advances to production-scale training pipelines helps accelerate the development cycle and can yield more efficient use of compute. Meanwhile, a seasoned red team helps ensure advances are accompanied by careful assessment of risk and mitigation.

For Anthropic, recruiting someone of Karpathy’s profile is a clear signal about priorities. It suggests the company values hands-on expertise in large-scale model training and an ability to operationalize research insights. At the same time, emphasizing AI-assisted pre-training suggests a strategy where the models themselves become tools for researchers, enabling faster hypothesis testing, automated analysis of training dynamics, and potentially cost reductions.

While Karpathy’s long-term plans regarding Eureka Labs are unclear, his move to Anthropic marks another step in the ongoing shuffle of top AI talent among industry-leading organizations. Observers will watch how his team uses Claude to influence pre-training workflows and whether AI-assisted research becomes a more dominant paradigm in the field.

Anthropic’s statements to media and Karpathy’s public posts provide the factual basis for this summary. TechCrunch and other outlets have sought comment from Karpathy for additional context about his role and future projects.

Key Insights Table