Google’s four AI departures: “We wanted to build something differently”
The New Stack Amanda Caswell ● Covered by 6 sources
Google just reshuffled DeepMind's leadership and lost four star engineers to a new startup called Discovery Loop. Hassabis steps back, one exec now runs all of Gemini, and Google still gets a cut of whatever the ex-employees build.
Alphabet just wrapped its best stock market year since 2009, riding a wave of goodwill for Gemini and the AI machine behind it. So it's a strange moment for Google to be shaking up the very lab that engineered that comeback. On Wednesday, the company announced Demis Hassabis is stepping away from running DeepMind day-to-day, moving up to chair of Google DeepMind and chief scientist of Alphabet. Koray Kavukcuoglu, DeepMind's longtime CTO and the architect behind Google's chip and infrastructure strategy, now owns essentially everything downstream: Gemini model development, frontier research, the consumer app, developer tools, all of it under one person.
That consolidation happens at the same time four of DeepMind's most recognizable names are walking out the door. Jeff Dean, who joined Google back in 1999 and helped build Google Brain, MapReduce, and TensorFlow, is leaving alongside longtime collaborator Sanjay Ghemawat. Oriol Vinyals and Quoc Le, both co-authors on the seminal 2014 sequence-to-sequence paper with Ilya Sutskever, are going with them. Together they're founding Discovery Loop, a public-benefit corporation built around a fairly audacious bet: that AI can run the entire research cycle itself, proposing experiments, executing them, judging the results, and deciding what to try next, with minimal human steering.
Ghemawat put it plainly to Wired — Google's infrastructure was built for Search, ads, and consumer scale, not for this kind of open-ended scientific exploration, and the team wanted to build something different from how things get built inside Google. That's notable timing, because OpenAI's Astra recently ground through ten unsolved math and science theorems for roughly $2,000 in compute, proof that automated discovery isn't a thought experiment anymore. Discovery Loop is stepping into a race that's already producing results, not laying groundwork for a hypothetical.
Google isn't cutting these four loose, either. It's investing in Discovery Loop and signing on as its cloud provider, which means Google gets a financial stake in whatever the startup finds and another giant compute customer for Google Cloud, while Discovery Loop skips the years-long slog of building its own data centers. It's a tidy arrangement on paper, though it follows a rougher pattern: Google has already watched Noam Shazeer head to OpenAI and AlphaFold researcher John Jumper move to Anthropic. Investors weren't thrilled about the reshuffle regardless of the hedge — Alphabet shares dropped more than 5% on the news.
Hassabis, for his part, says he wants more room to think about AGI strategy and to spend time on Isomorphic Labs, DeepMind's drug-discovery spinout, telling employees that AGI now feels close at hand. Kavukcuoglu inherits a roadmap with real pressure attached: the flagship Gemini 4 has already slipped past its planned June launch, even as Gemini 3 in November put Google back in real contention with OpenAI and Anthropic. Whether one person running the whole stack accelerates that timeline or just concentrates the risk is the thing worth watching next.
My take
Watching Google fund the people who just quit to build a rival research shop tells you more about the labor market for frontier AI talent than any earnings call will. Nobody actually leaves anymore — they just renegotiate their leash, and the big labs would rather own a slice of the next thing than lose it outright. Call it AGI feudalism: everyone still reports to the same three or four landlords, they've just moved the furniture around.
Read more about this at: The New Stack
