Three defections in two years. Jan Leike left OpenAI for Anthropic in May 2024, writing a pointed public post about safety culture on his way out. John Schulman - co-inventor of RLHF, the training technique that made ChatGPT useful by teaching it to follow instructions - followed in August 2024. Now Andrej Karpathy, one of OpenAI's six original cofounders and arguably the most publicly followed AI researcher alive, has announced he's joining Anthropic's pretraining team.
Pretraining is where base model capability is built. It's the process of training a model on massive amounts of text and data before any fine-tuning or safety work - the step that determines how capable a model fundamentally is. Karpathy joining that team signals Anthropic is investing seriously in the core of model development, not just the safety and product layers it's historically been known for.
What Karpathy Brings
Karpathy is unusual among top researchers because he communicates clearly to non-experts. His YouTube channel has attracted hundreds of thousands of subscribers. His "Neural Networks: Zero to Hero" series has become a standard reference for developers learning how large language models work from scratch. He also coined the term "vibe coding" - now used industry-wide to describe AI-assisted development where you describe what you want in plain language.
Before co-founding OpenAI, he was Director of AI at Tesla, responsible for Autopilot. He rejoined OpenAI in 2023, left again in early 2024, and then founded Eureka Labs, an AI education company. That he's now taking a staff role at Anthropic rather than continuing independently is notable on its own.
What the Pattern Means
You can read three senior OpenAI departures to Anthropic in different ways. One interpretation: Anthropic is simply a well-funded competitor with strong talent attraction (over $7 billion raised). Another: something about OpenAI's current direction is pushing senior researchers toward the door.
Leike's exit was explicitly critical - he cited safety culture erosion in his farewell post. Schulman's was quieter. Karpathy's stated motivation hasn't been framed as a protest, but the accumulation builds a pattern regardless of individual explanations.
For Anthropic, the practical benefit is a researcher with deep pretraining experience at the organization that built GPT-3 and GPT-4. The models Karpathy worked on remain the benchmarks everything else is measured against. Having that institutional knowledge on Claude's development team is a genuine asset.
Pretraining improvements take months to show up in released models - you don't hire a researcher today and ship a smarter Claude next week. But the long-term signal matters. Three senior researchers in two years choosing Anthropic over staying at or returning to OpenAI says something about where serious people think the more interesting work is happening right now.