The biggest names in AI keep switching jerseys, and Anthropic just made a hire that signals where the real research bet is heading.
Andrej Karpathy - one of the people who built the original OpenAI - just joined Anthropic, and he's not coming to make Claude better in the usual sense. He's coming to make Claude train Claude.
Karpathy posted on X on May 19 that he'd joined Anthropic. He started the same week, working on the pre-training team under team lead Nick Joseph.
Pre-training is where large language models pick up their core knowledge, and it's also where the biggest compute bills get racked up.
His specific role, according to Anthropic: launch a new team focused on using Claude itself to speed up pre-training research. Self-improving AI research isn't a marketing slogan at Anthropic - it's a hiring decision.
That's a meaningful shift in approach. The traditional pre-training playbook leaned on more chips and more data, and Anthropic is now betting that Claude itself can accelerate the work.
If you want a daily read on how AI is reshaping markets, Market Briefs breaks down the moves that matter in five minutes - and you get a free investing masterclass when you sign up.
Karpathy is one of a handful of researchers who can bridge LLM theory and large-scale training in practice. He co-founded OpenAI, then left in 2017 to lead AI at Tesla.
At Tesla, he ran Full Self-Driving and Autopilot until 2022, and he returned to OpenAI for a year before leaving again in 2024 to start the AI education startup Eureka Labs.
For Anthropic, this hire is bigger than headcount. The pre-training phase is the most expensive piece of building a frontier model, and using Claude to accelerate it could lower the cost curve or speed up the next model release.
The arms race used to be about who had more chips. Now it's about who has the smarter recipe for using them.
The broader AI talent war is intensifying alongside that shift, with major labs paying some of the highest comp packages in the industry to retain or attract senior researchers.
Anthropic also added Chris Rohlf this week to its frontier red team. Rohlf is a 20-year cybersecurity veteran who came from Meta and Yahoo, which gives him deep experience stress-testing systems against real-world threats.
That's two senior hires in one week aimed at very different problems - smarter models on one side, safer models on the other. Whether those bets pay off will show up in Anthropic's model release cadence and in its training costs.
Those costs are the bigger backdrop here. Frontier model training runs at an estimated nine figures per cycle, so anything that meaningfully shrinks that bill changes the economics of the whole industry.
If the time between releases tightens, or training costs drop, that's Karpathy's playbook showing up in the numbers.
For sharp daily reads on the AI race and the companies driving it, join Market Briefs - delivered every morning, with a 45-minute investing course as a bonus.
