Andrej Karpathy Joins Anthropic’s Claude Pre-Training Team
A frontier AI training control room centered on Claude-scale pre-training, with a researcher silhouette facing live training curves and compute clusters, suggesting Karpathy’s return to large-scale R&D.📷 AI-generated image / TECH&SPACE
- ★Andrej Karpathy started this week on Anthropic’s Claude pre-training team.
- ★Pre-training remains one of the most expensive and compute-heavy phases of frontier model development.
- ★Anthropic also hired Chris Rohlf for its frontier red team focused on advanced security testing.
Andrej Karpathy has not joined Anthropic as a decorative name on an advisory page. According to TechCrunch, he has joined the pre-training team working on large-scale training for Claude, under team lead Nick Joseph. That is the hardest layer of industrial AI infrastructure: expensive clusters, long experiments, choices that cannot be cheaply reversed, and model behavior that is shaped well before fine-tuning and product packaging.
Karpathy’s background fits that layer unusually well. He co-founded OpenAI, led computer vision and AI work at Tesla around Full Self-Driving and Autopilot, then briefly returned to OpenAI in 2023 before leaving in 2024. In his public note, he said he had joined Anthropic because the next few years at the frontier of LLMs would be especially formative and that he was excited to return to research and development. He also said he remains deeply passionate about education and plans to resume that work in time, a clear reference to his education-focused project Eureka Labs.
One of the most visible researchers from the OpenAI and Tesla orbit is returning to large-scale model training as Anthropic also strengthens its frontier red team.
A security red-team lab view where model evaluation dashboards, threat probes and defensive cyber traces intersect, tying Rohlf’s hire to Anthropic’s safety testing layer.📷 AI-generated image / TECH&SPACE
This is not simply a personnel story. It is a signal about where frontier labs now expect advantage to emerge. The public sees chat interfaces, subscription tiers, demos and product launches. But the real separation between leading labs often starts in pre-training: data choices, run stability, scaling decisions, early evaluation, and the ability to turn a compute-heavy experiment into a model that is not just larger, but more useful and more predictable.
For Anthropic, the move says that the next phase of Claude will not depend only on integrations and distribution. It will depend on deep research discipline at the level where models are first formed. That matters because Anthropic is competing with OpenAI and Google in a market where technical gains quickly become commercial pressure, while failed training decisions can burn months and vast amounts of compute.
The second part of the same story is security. Anthropic also hired Chris Rohlf for its frontier red team, which stress-tests advanced models against severe threats. Rohlf has more than 20 years of cybersecurity experience, including work at Yahoo’s The Paranoids and six years at Meta. His statement that AI could dramatically improve cybersecurity is both an ambition and a warning: models that can write, analyze and plan more effectively can help defenders, but those same capabilities need to be tested before they become operational risks.
Karpathy’s arrival should therefore be read alongside Rohlf’s. One hire strengthens the lab on the model-building side; the other strengthens the team trying to understand how those models behave under pressure. For Anthropic, the priority stack is clear: scale Claude while showing that safety is not a late-stage footnote. For the rest of the market, the message is colder and more concrete: the race for people who can train and break frontier models is moving into a more serious phase.
