Jeff Dean leaves Google for Discovery Loop: the DeepMind reshuffle and what it signals

On August 5, Google announced that Jeff Dean is leaving after roughly 27 years. He is not going alone. Sanjay Ghemawat, Oriol Vinyals, who was technical lead on Gemini, and Quoc Le are founding Discovery Loop with him, a company whose stated goal is automating the scientific experimental cycle. Google is a founding investor and the Cloud partner. Inside Alphabet, Demis Hassabis becomes chair of DeepMind and chief scientist of Alphabet, while Koray Kavukcuoglu takes over the day-to-day running of DeepMind as SVP, a role CNBC confirmed on August 13 reports to Sundar Pichai and carries the frontier and Gemini push.

Read it as a personnel story and it is big. Read it as a signal about where the centre of gravity is moving and it is bigger.

What actually left

Dean and Ghemawat built the infrastructure layer modern Google runs on. Vinyals ran the technical side of the model Google's entire AI strategy is named after. Le is behind a long list of the ideas that made scaling work. Losing any one of them would be a headline. Losing all four on one day, to one new company, is a statement that the most interesting problem is no longer the model.

The Hassabis move fits the pattern. Chair plus chief scientist is a role you take when you want to steer research direction rather than run a product organization. Kavukcuoglu now owns the operational frontier race. In org-chart language, Google is saying that building Gemini and doing science with it are two different jobs.

From building models to models doing science

Discovery Loop's pitch is the loop itself: hypothesis, experiment, analysis, next hypothesis, run by AI at a pace no lab matches by hand. That is not a fringe idea in 2026. On September 8, the same day DeepMind said it had mapped 9 billion DNA variants, OpenAI reported that an internal model coordinating on the order of 10,000 parallel agents had produced a proof about finite-time singularities in the Navier-Stokes equations and formalized it in Lean, after 88 hours, 2.7 million messages and around 130 billion output tokens. Whatever you think of the maths, the shape is the point: many agents, a machine-checkable verifier, and a loop that runs until it passes.

The labs spent the last five years teaching models to answer questions. The next five are about teaching them to run the experiment that produces the answer.

For the people who built the previous phase, the incentive is obvious. Inside Alphabet they were maintaining a product line. Outside, with Google's money and Cloud behind them, they get to define a category.

Why an engineer should care

This is not gossip if you plan infrastructure or hire researchers.

• Compute follows the loop. Experiment-running agents are token-hungry in a way chat is not; 130 billion output tokens for one result is real money and real hardware. Expect demand to shift toward long-running, verifiable batch workloads.
• Verification becomes the product. Lean, tests, simulators. If the output cannot be checked by a machine, the loop does not close. Anyone building agent pipelines should be investing in the checker, not the prompt.
• Talent moves to where the problem is. I wrote about the researcher-leaves-big-lab pattern in Mikolov after word2vec. The person who built the last thing rarely builds the next one inside the same org.
• Google Cloud is the tell. Google did not let this team walk to a competitor. Founding investor plus Cloud partner means Alphabet keeps the compute revenue and the option value while shedding the org overhead.

What to watch

Kavukcuoglu's first shipped Gemini generation will show whether the frontier push slows without Vinyals. Discovery Loop's first public result will show whether "automating the experiment cycle" is a product or a research programme. And the DNA-variant work is the internal comparison: AI-for-science stays at Google too, under a different structure.

The honest limitation

Everything above is inference from an org announcement, two press confirmations and a handful of research results. I have no inside information on why the four left, what Discovery Loop will actually build, or how much of Google's investment is cash versus compute credits. The Navier-Stokes result is OpenAI's own account of an unreleased model, and the DNA-variant figure is DeepMind's. Treat the direction as clear and the details as unverified.