Sundar Pichai told Google staff in a memo on August 5 that Demis Hassabis is stepping back from running Google DeepMind day to day, and that Jeff Dean is leaving after 27 years along with three other senior researchers to start a company Alphabet has invested in.
Hassabis becomes chair of Google DeepMind and chief scientist of Alphabet, and continues to lead Isomorphic Labs. Koray Kavukcuoglu, the lab's chief technology officer, becomes senior vice president of Google DeepMind reporting to Pichai, with responsibility for Gemini, frontier research and developer platforms.
Dean, Google's chief scientist, leaves with Sanjay Ghemawat, a senior fellow and longtime collaborator, Oriol Vinyals, a co-technical lead on Gemini, and Quoc Le, who co-founded Google Brain and spent years automating model design. Pichai thanked Dean for "an incredible 27-year run."
The four are founding Discovery Loop, a public benefit corporation intended to automate multi-step science and engineering experiments by running thousands of experimental loops in parallel. It starts with machine-learning research before extending to hardware design. The seed round was co-led by Radical Ventures and Khosla Ventures, with Alphabet taking a stake and serving as cloud partner. Neither the valuation nor the size of Alphabet's stake has been disclosed.
Alphabet shares fell as much as 5 percent during the session and closed down about 4 percent, then slipped a further 1.29 percent the following day to $357.75.
The departures follow a run of slipped releases. Gemini 3.5 Pro missed a June target and remains in partner testing after falling short of internal coding benchmarks, and a training data update in late June made results worse. Google confirmed on July 21 that pre-training had begun on Gemini 4. Current and former Google DeepMind employees told Axios that morale problems were contributing to delayed releases, citing burnout, competitive pressure, departures and internal disagreement over the company's military work.
Discovery Loop's founders say scientific progress is limited by the cost of running experiments rather than by a shortage of ideas. Several labs are working the same ground. OpenAI said on August 1 that an internal model solved ten open problems in mathematics and published machine-checkable proofs. Anthropic has said its models are accelerating its own development, and Andrej Karpathy has argued that humans are the bottleneck in AI research.
Kavukcuoglu has spent 13 years at the lab and takes over the Gemini roadmap with the next frontier model already in pre-training. Hassabis, who has said artificial general intelligence is close at hand, keeps a research remit at Alphabet alongside Isomorphic Labs, the drug discovery company he also runs.
Sources: CNBC, The Next Web, Bloomberg
–
By the Control Plane Editorial Team