Summary
On May 19, 2026, Andrej Karpathy announced he is joining Anthropic’s pre-training team, reporting directly to Nick Joseph, Anthropic’s head of pre-training. His mandate is to build a new team focused on using Claude to accelerate pre-training research — applying AI to improve the training of future AI systems. This video from Wes Roth argues the move is better understood through the lens of Karpathy’s post-OpenAI independent work than as a simple talent acquisition story.
The centerpiece is Karpathy’s Auto Researcher project, sometimes called the “Karpathy loop” — a ~30-line agent built on his NanoGPT stack that proposes code changes to a training script, runs 5-minute experiments, evaluates validation loss, and commits improvements that pass. Over two continuous days of operation, the system ran 700 experiments, identified roughly 20 stackable improvements, and reduced GPT-2 training time from approximately 2 hours to 1.8 hours — an 11% speedup. The video draws direct parallels to Google DeepMind’s Alpha Evolve, which produced sub-1% efficiency improvements to Gemini chip manufacturing and Google’s Borg data center scheduler — gains still active in production.
The broader context is Anthropic co-founder Jack Clark’s public claim that there is a 60%+ probability of reaching fully AI-driven R&D with no human involvement by the end of 2028. The video frames Karpathy’s hire and his Auto Researcher work as direct evidence that the frontier labs are actively building toward that milestone, with automated ML research loops emerging as a core competitive moat.
📺 Source: Wes Roth · Published May 24, 2026
🏷️ Format: News Analysis
