Andrej Karpathy joined Anthropic. The OpenAI founding member's job: use Claude to train Claude.

Andrej Karpathy started at Anthropic this week. He posted the news to X on Tuesday, and the post cleared 3 million views inside an hour, per Fortune.

Karpathy is one of the people who built OpenAI in 2015. He left to run Tesla’s Autopilot vision work, came back to OpenAI for a year, then quit again in 2024 to start Eureka Labs, an AI-tutor company. The whiplash this week is that the person who has been arguing for years that AI education is undervalued is suddenly back inside a frontier lab. Specifically, the lab that beat his former employer’s coding model on most published benchmarks last quarter.

The role: Claude training Claude

Karpathy is joining the pre-training team under Nick Joseph, according to TechCrunch. An Anthropic spokesperson told the outlet he’ll “start a team focused on using Claude to accelerate pre-training research.” Pre-training is the heavy work that gives a model its raw capability before any reinforcement learning or fine-tuning gets layered on. It’s also the part of the stack where Anthropic has been quietly compute-bound.

In his X announcement Karpathy wrote: “I’ve joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D.”

The “Claude trains Claude” framing is not as recursive as it sounds. What Anthropic actually means is that current-generation Claude models will help researchers run experiments, generate and filter training data, and write the supporting tools for the next pre-training cycle. Several labs already do versions of this. Karpathy’s job is to make it good enough that it materially shortens the loop.

Why this matters

Pre-training is where the next generation of frontier models is won or lost, and the people who have actually shipped one are a small group. Karpathy is one of them. He led the early work on GPT-2 and was the original author of the nanoGPT reference implementation that a lot of mid-2020s researchers learned on. Hiring him into pre-training is not a brand grab. It’s a bet that he can compress the next training run from the inside.

It also lands on top of a busy compute month for Anthropic. Axios reported yesterday that the company is paying SpaceX about $1.25 billion a month for Colossus 1 and Colossus 2 capacity through May 2029, on top of the 220,000 GPUs it rented from xAI earlier this month and the Stainless acquisition on May 18. The shape of all three deals is the same: more compute, faster, with sharper tools wrapped around it. Karpathy is the human version of that pattern.

What’s still unclear

Eureka Labs is the open question. Karpathy founded it in July 2024 to build an AI-first version of the introductory deep-learning course he used to teach at Stanford. He told Fortune his work there is paused, not killed: “I remain deeply passionate about education and plan to resume my work on it in time.” The company has not posted a public update since late 2025. Whether it stays in stasis, gets sold, or runs in the background while he’s at Anthropic is the part nobody has answered yet.

The other open thread is OpenAI. Karpathy left it twice. The first departure made sense, he wanted to work on cars. The second one, in February 2024, was framed as a non-event, no drama, no specific reason. Joining a direct competitor 14 months later is the kind of move that retroactively rewrites the second departure. OpenAI has not commented.

What this means for you

If you’re following Claude releases, this is a leading indicator, not a launch. Pre-training cycles run for months. Anything Karpathy directly influences will land in the model series that comes after Opus 4.7 and the Mythos previews, not in the next dot release. Watch the next big version notes for a specific mention of “pre-training improvements” or a sharper compute-to-capability curve; that’s where the hire shows up.

For the broader hiring market, the signal is sharper. Frontier labs are now openly bidding for individual contributors who have shipped a real pre-training run. The talent pool is in the dozens, not the thousands, and most of them are already employed. Expect another high-profile move from OpenAI or Google DeepMind inside the next quarter. The Karpathy hire raises the floor on what an established pre-training lead can ask for, and competitors will quietly match.

For Karpathy himself, the most interesting line in his post wasn’t “I’ve joined Anthropic.” It was “get back to R&D.” After two years of building courses and writing YouTube explainers, he’s saying out loud that he wants to be in the loop again. That, more than the new title, is what’s worth tracking.