Jeff Dean Left Google After 27 Years and Took Three More AI Pioneers With Him

Four people left Google. Between them they built MapReduce, Bigtable, TensorFlow, AlphaStar, seq2seq and AutoML.

Jeff Dean went after 27 years. Sanjay Ghemawat, his collaborator for most of those years, went with him. So did Oriol Vinyals from DeepMind and Quoc Le from Google Brain. The new company is called Discovery Loop.

What Discovery Loop says it is building

The stated goal is automating scientific and engineering research itself. Phase one is narrower and more pointed: automating machine learning research. After that the company has named hardware design, drug discovery and clean energy as targets.

It is structured as a public benefit corporation rather than a standard startup, which legally obliges the directors to weigh a stated public mission alongside shareholder returns. Radical Ventures and Khosla Ventures are co-leading the seed round.

The odd detail is that Google is a founding investor and cloud partner. That is an unusual way to lose four of your most senior researchers, and it suggests the split was negotiated rather than abrupt.

Why the names matter more than the titles

Job titles undersell this badly. It is easier to read the departure through the work.

• MapReduce and Bigtable (Dean and Ghemawat) are the reason large scale data processing looks the way it does. Hadoop existed because MapReduce was published.
• TensorFlow made deep learning something an ordinary engineering team could deploy rather than a research group's private toolchain.
• seq2seq (Le, with Ilya Sutskever and Vinyals) is a direct ancestor of the architecture behind every modern language model.
• AlphaStar (Vinyals) showed reinforcement learning handling a real time strategy game with imperfect information.
• AutoML (Le) was an early serious attempt at models designing models, which is now the thesis of the entire new company.

That last line is the point. Discovery Loop is not a swerve for these people. It is the AutoML idea taken to its logical end.

The bet underneath the company

Most frontier labs are still running a version of the same play: more compute, more data, larger models, better post-training. It works, and it is expensive, and the returns per dollar are getting harder to argue about.

Discovery Loop's bet is that the bottleneck is not model size but the speed of the research loop. If an AI system can propose experiments, run them, read the results and propose better ones, then the rate of progress stops being limited by how many researchers a lab can hire.

Get that right and the definition of an AI lab changes. The valuable asset stops being the model and becomes the loop that produces models. Get it wrong and it is an extremely well credentialed way to rediscover that research taste is hard to automate.

What this costs Google

Google is not short of researchers, and Gemini's scale is evidence the machine still works. But Dean and Ghemawat were institutional memory as much as headcount. A meaningful share of the intellectual foundation of Google's infrastructure just walked to a startup that Google is helping fund.

The wider pattern is worth watching. Senior AI researchers leaving large labs for smaller, mission-specific companies has been steady through 2026, and it is happening while regulators and safety institutes are still building the capacity to evaluate what those labs produce. Expertise is dispersing faster than oversight is consolidating.

What to watch

Discovery Loop has published a thesis and a founding team, not results. The things that would make it real:

• A concrete published result where the automated loop, not the humans, found something non-obvious.
• Whether the public benefit structure changes any actual decision, or stays decorative.
• Who else leaves Google in the next six months. Departures of this seniority tend to arrive in clusters, and the people who work on autonomous agent behaviour are the obvious next wave.

Four people is not a talent leak. It is a substantial piece of the foundation of modern Google infrastructure, relocated.