Andrej Karpathy started this week at Anthropic, joining a new pre-training research team focused on using Claude to accelerate Claude. The hire was confirmed Monday across TechCrunch, Bloomberg, and Anthropic itself. If you have lost count of which AI lab Karpathy works at, the current scoreboard reads: OpenAI (founding member, 2015 to 2017), Tesla (Autopilot lead, 2017 to 2022), OpenAI again (2023 to 2024), Eureka Labs (his own AI education startup, 2024 to now), and now Anthropic. The Eureka thing apparently continues in parallel.
Two interesting wrinkles. First, the assignment is pre-training specifically. Pre-training gets less press than alignment and post-training but is still the engine room of frontier models. The new team’s stated mission is using Claude itself to accelerate pre-training research. That is the recursive-self-improvement framing in its calm, project-management form: use the AI to make the AI better at making the AI. The framing has been around for years. The org chart now has it as a line item.
Second, this is the latest in what looks less like a coincidence and more like a migration pattern. The frontier labs have been trading senior people for two years, but the OpenAI-to-Anthropic direction has gotten conspicuously busier. Anthropic is also reportedly in talks for a round between $30 and $50 billion that would value the company near $950 billion, a number that would have read as an Onion headline three years ago and now reads as a Tuesday.
What this changes operationally is hard to say from the outside. Karpathy is one person, and pre-training is a team sport with a thousand engineers behind every breakthrough. What it changes culturally is more legible: another piece of the OpenAI origin myth has politely walked across the street to a competitor. The narrative gravity keeps pointing the same way.