A new AI lab called Core Automation posted its first message to X on Tuesday. No product demo. No funding announcement. Just a single claim: 'We are building the world's most automated AI lab. Our objective: systems that optimize and automate work, starting with research itself.'

Normally you would scroll past that. Startups claiming to build the future are as common as La Croix in San Francisco. But the roster behind Core Automation is the kind that makes hiring managers at Anthropic, Google DeepMind, and OpenAI check their Slack DMs with fresh paranoia.

The Roster

Jerry Tworek, listed as CEO and cofounder, was a vice president at OpenAI. Rohan Anil left Anthropic after Tworek, in his words, 'nerdsniped' him. Anil had previously worked at Google DeepMind before joining Anthropic, which he described as 'one of the best places to work for a researcher.' Anmol Gulati, a research scientist from Google DeepMind working on Gemini, said he is 'starting something new with some exceptional people.' Joanne Jang, who spent over four years as a general manager at OpenAI, has updated her X bio to 'trying to automate myself @coreautoai.'

Others who have confirmed joining include former Google DeepMind staffers Ehsan Amid and Avery Lamp, OpenAI's former head of people Julia Villagra, and Sai Surya Duvvuri, a former research intern at Google and Meta. The company's website says its team has 'helped build frontier models' and 'influential architectures.' That is underselling it.

The Thesis: Automate the Researchers

Gulati's post explains the thinking: 'I have increasingly felt that the current research paradigm, scaling models, data, and static deployment, will not get us all the way. We believe the next phase comes from something different: new learning algorithms, architectures beyond today's stack, and systems that automate the process of building itself.'

Translation: the scaling laws that got us from GPT-3 to GPT-5 are hitting diminishing returns. Instead of throwing more compute at bigger models, Core Automation wants to build systems that improve themselves. The AI lab that requires fewer humans to run. If that sounds familiar, it should. Recursive Superintelligence, the London lab we covered last week, raised $500 million at a $4 billion valuation with 20 employees chasing the same idea. Richard Socher, Jeff Clune, Tim Rocktaschel. Same thesis: automate the process of AI research.

The Talent War Escalates

Core Automation is the latest in a string of talent raids that are reshaping the AI industry's power structure. Yann LeCun left Meta to start AMI Labs. Mira Murati left OpenAI to start Thinking Machines Lab (which just signed a multibillion-dollar Google Cloud deal, despite Meta poaching seven of her founding members). OpenAI, Anthropic, and Google DeepMind have all lost senior researchers to startups offering cofounder titles, equity packages, and the promise of building something from scratch.

Business Insider reported last year that base salaries at AI startups rose rapidly as they competed for talent. But equity is the real weapon. Shawn Thorne at executive search firm True Search said startups offset the 'opportunity cost' for top researchers by offering cofounder titles, access to compute, and time for independent research. When your alternative is being a cog in a 1,000-person lab at Google, the pitch of building something new with 15 of the best people you have ever worked with is hard to refuse.

What to Watch

No funding has been disclosed yet, but with this team, the seed round will not be small. The pattern is predictable: stealth launch, elite team announcement, massive raise within weeks. Recursive Superintelligence did it. Thinking Machines Lab did it. Core Automation will almost certainly do it.

The bigger question is what happens when the best researchers at the best labs keep leaving to build smaller, more focused competitors. Anthropic retained 88% of its researchers according to Fortune. Google DeepMind retains 78%. OpenAI retains 67%. Those are decent numbers until you realize the 12-33% walking out the door are often the ones who mattered most.

DeepMind has reportedly started enforcing six-to-twelve-month non-compete clauses in response. Good luck with that in California.

The AI talent war just got a new front. And the weapons are not signing bonuses. They are ideas interesting enough to make world-class researchers quit the most prestigious labs on Earth to chase them.