Andrej Karpathy, one of the co-founders of OpenAI, former head of AI at Tesla, and one of the most recognisable researchers in the field, this week moved to Anthropic - the competitor that for the last two years has regularly rattled both OpenAI and Google with the Claude model. The move is not symbolic. He will lead a team that will use Claude to train future Claude models.

"I've joined Anthropic," Karpathy wrote on X. "I think the next few years at the frontier of LLMs will be especially formative. I'm very excited to return to R&D."

More precisely, he joined the pre-training team - that most expensive and most intensive phase where the model acquires foundational knowledge and capabilities. Under Nick Joseph, his immediate boss, Karpathy will build a team to research how to use Claude to accelerate the pre-training process itself. Put simply: AI-assisted research, not just pouring in more chips.

This is the most significant signal from Anthropic in recent months. When OpenAI has a bigger market valuation, Google has more accessible infrastructure, and Microsoft is pumping ever larger capital into OpenAI - Anthropic has to find another way to stay at the top. Apparently that's through "smart" training, not just more training. The question is whether this is strategic work ethic or a final realisation that more chips won't fix the problem.

Karpathy's career reads like a chronicle of this era: OpenAI until 2017, Tesla where he led the autopilot team until 2022, OpenAI again until 2024, then his own startup Eureka Labs for AI in education. Whether Eureka Labs continues its work - he didn't answer clearly. "I remain deeply passionate about education and plan to come back to that work with time," he wrote. Which means: a pause.

At the same time, Anthropic added another notable hire. Chris Rolf, with 20 years of cybersecurity experience (Yahoo, Meta), joined their "frontier red team" - the team testing how the models can be abused for more serious threats. When one company invests at the research frontier and the defensive wall at the same time, that isn't coincidence - that's a plan.