News · 2026-08-05

Jeff Dean leaves Google after 27 years to build an AI that runs the experiments itself

Alphabet announced two things in a single memo on the morning of August 5, 2026: Demis Hassabis is handing off day-to-day control of Google DeepMind, and Jeff Dean, Google's chief scientist and a 27-year veteran, is leaving the company entirely. Dean is taking three more of Google's most senior researchers with him to found Discovery Loop, a startup whose premise is that AI should stop assisting researchers and start being the researcher. Alphabet shares fell immediately on the news.

Key facts

• Four senior researchers left at once: Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le.
• Announced August 5, 2026 in a memo from Alphabet CEO Sundar Pichai, with Discovery Loop launching the same day.
• Google itself invested as a founding investor and is providing compute for the startup's first year.
• Primary sources: Pichai's memo on blog.google, Discovery Loop's own site, and Steven Levy's same-day founder interview in Wired.

Start with the org chart, because it is the smaller half of the story. Hassabis becomes Chair of Google DeepMind and Chief Scientist of Alphabet, based in London, and in his own words hands over his day-to-day operational responsibilities. He keeps running Isomorphic Labs, the drug-discovery company DeepMind spun out, and his stated new job is to work with Pichai on strategic and global AGI matters. Koray Kavukcuoglu, previously DeepMind's CTO, becomes SVP of Google DeepMind reporting directly to Pichai, keeping his chief AI architect function and taking scope over Gemini model development, frontier research, and the Gemini app and developer teams. Pichai credited him by name for shipping WaveNet and DQN.

One correction worth making early: several secondary write-ups have described Kavukcuoglu as becoming CEO of Google DeepMind. The memo itself says SVP, reporting to Pichai. No CEO title appears in the primary source.

The four who left

Dean joined Google in 1999 as roughly its twentieth employee, cofounded Google Brain, and was technical co-lead on Gemini. Ghemawat, his collaborator of more than twenty years, co-architected the Google File System, Bigtable, and Spanner, which is to say much of the infrastructure the modern web runs on. Vinyals was VP of research at Google DeepMind and a technical lead on Gemini. Le cofounded Google Brain and did the AutoML work that first tried to automate architecture search.

Pichai's memo handles the departure warmly and briefly: "After an incredible 27-year run, Jeff Dean is at a moment where he wants to try something new, and we're excited to support him in that." Google is not just waving them off. It invested in the round, will be a Cloud partner, and agreed to supply compute for the first year. The round itself is co-led by Radical Ventures and Khosla Ventures with Lightspeed, Kleiner Perkins, and Doerr Capital participating. The amount and valuation were not disclosed, and Levy reports the founders declined to share them when asked directly. Discovery Loop is incorporated as a Delaware public benefit corporation, based in Palo Alto, with a lean in-person team and no office yet.

Market reaction was real and immediate, though the exact number is soft. Outlets reported an intraday decline ranging from about 4 percent to 5.4 percent. We flagged an overstated stock-drop figure in our June coverage of the earlier DeepMind departures, so treat the direction as well corroborated and the precise percentage as varying by source.

The bet

Discovery Loop's pitch is specific enough to argue with. Their site says scientific progress is bottlenecked by the sequential human loop of propose, implement, evaluate, refine, and proposes to replace it with parallel execution of thousands of experiments. Dean's framing on the site: "Imagine a future where a handful of people can conduct scientific research and engineering tasks much more rapidly, and with higher quality, than massive teams of scientists and engineers do today." Vinod Khosla, who met the founders in his office on a Saturday to keep the news quiet, put it more bluntly to Wired: "Humans have been using AI to do research, not using AI to be a researcher. The fundamental thing is that AI is the researcher."

The first customer is themselves. Le told Wired: "I'm very excited about automating machine learning. It might be that we will discover a different transformer architecture." The sequencing after that is chip design, drug discovery, materials, and clean energy.

The honest caveat, from an insider

The most useful line in the entire launch came from Vinyals, naming the unsolved part: "One of the things that we'll be obviously very focused on is how these models come up with new ideas to try. That's not something that currently they're super strong at."

That is the whole load-bearing assumption stated as an open problem by one of the founders. Running ten thousand experiments in parallel only compounds value if the experiments are worth running. If idea generation is the bottleneck rather than execution throughput, massive parallelism buys a faster treadmill, not a faster discovery rate. That tension is live in the research literature too, where results on agents improving themselves keep landing softer than the framing suggests, and where sampling more attempts beat clever self-reflection in all 36 comparisons of one recent study.

Dean told Wired the idea came together "only a few weeks ago," which reads as genuine rather than as a long-planned exit. Three weeks earlier, Hassabis had publicly proposed a FINRA-style industry reviewer for frontier AI. He now occupies a strategy seat rather than running the lab that ships the models. That is not a contradiction, but it is a coincidence of timing worth noticing rather than resolving.

Primary source, verified: read the paper →

Ground Truth. (2026, August 5). Jeff Dean leaves Google after 27 years to build an AI that runs the experiments itself. Ground Truth. https://groundtruth.day/news/jeff-dean-leaves-google-to-build-an-ai-that-is-the-researcher.html

@misc{groundtruth:jeff-dean-leaves-google-to-build-an-ai-that-is-the-researcher,
  title  = {Jeff Dean leaves Google after 27 years to build an AI that runs the experiments itself},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/jeff-dean-leaves-google-to-build-an-ai-that-is-the-researcher.html}
}

Topics: industry · google · deepmind · startups · ai-research · automation