Jeff Dean Leaves Google to Automate Science Itself

Jeff Dean is leaving Google after 27 years. He is not going alone. Sanjay Ghemawat, Quoc Le and Oriol Vinyals — three of the most consequential engineers and researchers the company has employed — are going with him, and the four have formed a startup called Discovery Loop. TechCrunch reported the departures on August 5.

The premise is narrower and stranger than "AI for science." Discovery Loop wants to automate the experimental loop itself — proposing a hypothesis, running the experiment, evaluating the result, and feeding that back into the next proposal — and then run that loop thousands of times in parallel rather than once at a time. The company describes itself as building systems that "leverage massive computational scale to fundamentally transform the speed and efficiency of innovation by automating complete experimental loops." Dean put the case to the New York Times in plainer terms: more experiments and better experiments, and the breakthroughs follow.

Its first customer is itself. Discovery Loop says it will begin by automating machine-learning research and engineering — using AI to design better AI — before moving outward into hardware design, drug discovery, materials and clean energy. That ordering is the whole thesis. Recursive self-improvement is the fastest domain to test because the lab and the subject are the same building, and a win there compounds into every domain that comes after.

The roster is what makes this more than another well-funded lab. Dean was employee #30 and Google's chief scientist. Ghemawat, a Google Senior Fellow, co-designed with Dean the distributed systems — MapReduce, Bigtable, the Google File System — that most of the modern cloud was subsequently modeled on. Le co-founded Google Brain and did foundational work on neural architecture search. Vinyals was a VP of research at Google DeepMind and a Gemini co-lead. Between them they have fingerprints on Search, TensorFlow, AlphaFold and Gemini.

The initial round is co-led by Radical Ventures and Khosla Ventures, with Kleiner Perkins, Lightspeed and Doerr Capital participating — and Alphabet itself putting money in. The round has not closed and the company declined to give a valuation. Discovery Loop is incorporated as a public benefit corporation, the same structure Anthropic and xAI chose, which binds the board to weigh a stated public mission alongside shareholder return.

Alphabet backing the vehicle that just took four of its most senior technical people is less contradictory than it looks. It buys Google exposure to whatever the loop produces without having to run the experiment internally, and it keeps the relationship warm. What it does not do is refill the seats. Google has spent 2026 losing senior researchers at a steady clip — a Gemini co-lead in June, Nobel laureate John Jumper to Anthropic the same week — and this is the largest single block yet.

The hard part is not the pitch, it is the physics. Automating a loop only pays if the evaluation step is cheap, fast and honest; ML research qualifies, because a training run grades itself. Wet-lab biology and materials science do not, and the bottleneck there has never been the shortage of hypotheses. Discovery Loop's real test is whether the compounding it expects to find in its first domain survives contact with the second.