Andrej Karpathy has announced that he is joining Anthropic.

The news broke on Tuesday, dropping right on the opening day of Google I/O. Grabbing headlines is standard practice in this industry, but releasing the announcement on the same day as a rival's big event carries more than a little subtext.

He isn't joining to slow down. As Anthropic's head of pre-training Nicholas Joseph revealed, Karpathy will build an entirely new team.

The team has just one core mission: accelerating pre-training research using the Claude model.

To translate what that means: Anthropic aims to achieve recursive self-improvement. Let AI train the next generation of AI, or let AI upgrade itself, removing humans from the loop entirely — or at minimum, cutting human intervention down to almost nothing.

There's no need to rehash Karpathy's resume. He is a founding member of OpenAI and the former lead of Tesla's Autopilot program.

In the two years since he left OpenAI, he has stayed far from idle. He launched Eureka Labs, an AI-native education initiative, and built autoresearch, an open-source tool for running automated AI experiments.

He has long identified as an educator and a believer in the open-source movement.

But Anthropic is a different beast. While the company has released seemingly open standards like the Model Context Protocol, Claude itself remains a closed-source commercial product at its core.

That's what makes this move so interesting: a person who has made a name for walking others through how to build neural networks is now setting out to research how neural networks can build themselves.

Karpathy left an opening in his tweet, noting he will return to education work down the line.

It's clear that cutting-edge R&D is what excites him most right now. As for balancing closed-source commercial work with his open-source ideals, that's a problem for another day.