Who just walked out of Google DeepMind, and what are they actually building?
On 5 August 2026, four of the most senior names inside Google's AI stack left together. Not for OpenAI. Not for Anthropic. For a company that barely existed a few weeks earlier, called Discovery Loop.
Jeff Dean spent 27 years at Google. Employee number 30. Co-founder of Google Brain, chief scientist, co-technical lead on Gemini. He helped invent the systems that made search work at planet scale: MapReduce, Bigtable, Spanner, TensorFlow, the TPU story. Sanjay Ghemawat is the other half of that pair. Same vintage, same infrastructure papers, the quiet engineer whose name sits next to Dean's on the systems that still run the company. Oriol Vinyals was VP of research at DeepMind and a technical co-lead on Gemini, with AlphaStar and the old sequence-to-sequence work on his CV. Quoc Le co-founded Google Brain and built AutoML-Zero, the project that tried to let algorithms invent algorithms from scratch.
These four have worked together for between 14 and 30 years depending on the pairing. They vacation together. The idea for the company, they have said, came together only a few weeks before the announcement. Dean is CEO. The firm is a public benefit corporation, which is a legal way of saying they can choose a mission over a quarterly number and not get sued for it.
Why leave after 27 years
Sundar Pichai tried to keep them. Multiple meetings. It did not work. The public version is warm: Dean wanted to try something new, Google is a founding investor and the cloud partner, compute for year one is part of the deal. Alphabet even stayed on the cap table. Shares still dropped around 4 percent the day the news landed.
The private version is more interesting. In a giant lab, a radical change has to fight inertia first. Product calendars, Gemini launches, ads, search, the next demo. Vinyals put it simply: they wanted to build something different, and a large organisation is built to absorb difference until it looks like last year's plan. Dean has also said a small company can now rent the infrastructure it used to have to invent. Cloud compute means four people can raise serious money and skip the decade of plumbing.
Hold on. Google is investing in the company that just took its chief scientist and both Gemini co-leads. Why would you fund the people walking out the door?
Because the alternative was worse. Fight them, lose the relationship, watch them raise from someone else and build on someone else's chips. Keep a stake, keep the cloud contract, keep a research framework on ML systems. It is the polite version of not wanting to be locked out of whatever they invent next. It is also a tell: this was not a raid. It was an amicable split that still hurt.
The concept they walked out with is not "we will train a bigger Gemini." It is the scientific method as a loop. Propose an experiment, implement it, evaluate the result, go again. Run thousands of those loops in parallel. First customer: themselves. Point the loops at machine learning research until the system can invent better architectures than the transformer everyone is stuck on. Then point the same machinery at biology, drugs, chips, materials, clean energy.
That is why the names matter. Dean and Ghemawat know how to build the factory that runs a thousand experiments at once. Vinyals and Le know how models propose ideas and how automated machine learning actually fails. The bottleneck they keep naming is not code. It is taste. Current models are weak at coming up with a new idea worth trying. Early on, humans and machines will invent hypotheses together. Deeper automation is the later bet.
So yes, they left for startup freedom. That part is true and a bit boring. The sharper part is who they are. These were not four researchers who got a better offer. These were the people who built the stack the rest of the company still sits on, and they decided that stack was the wrong shape for the next problem.