Jeff Dean is leaving Google after 27 years. The chief scientist is co-founding Discovery Loop, an AI startup that wants to automate scientific research, with three of Google’s most decorated researchers: Sanjay Ghemawat, Quoc Le, and Oriol Vinyals.
Google confirmed the news in a memo from Sundar Pichai posted Wednesday. Alphabet stays on as a founding investor and cloud partner. Dean will be CEO.
Why Jeff Dean left
Dean is Google employee 30. He helped build the search infrastructure, co-founded Google Brain in 2011, and led early Gemini work.
The New York Times reports Jeff Dean’s profile inside Google waned in recent years as other executives, including co-founder Sergey Brin, asserted control over Google’s all-in AI bet.
The four founders have worked together for stretches of 14 to 30 years, per The Decoder. Ghemawat has been Dean’s collaborator for more than two decades.
Le helped build Google Brain’s early models. Vinyals oversaw Gemini’s chatbot design with Jeff Dean.
This isn’t four strangers pooling résumés. It’s a team that already shipped the defining systems of the modern AI era.
The departure is part of a bigger reshuffle. Demis Hassabis steps down as DeepMind CEO to become Alphabet chief scientist and DeepMind chair. Koray Kavukcuoglu takes over as DeepMind SVP, overseeing Gemini development.
Alphabet shares fell about 4% after the news, CNBC reported. The reshuffle lands as Gemini 4, the flagship model Google targeted for a June launch, remains unreleased.
What Discovery Loop wants to build
Discovery Loop is a public benefit corporation with a simple pitch: compress the scientific method. The company’s site describes systems that propose experiments, run them, evaluate results, and iterate at machine speed, thousands of loops in parallel.
The plan starts with machine learning research and engineering, then expands to bigger domains. The founders frame it as taking on the National Academy of Engineering Grand Challenges, from better medicines to clean water to cheap solar.
These four bring real receipts. MapReduce, BigTable, Spanner, TensorFlow, TPUs, AlphaFold, Gemini. Between them they wrote chunks of the internet’s plumbing and most of the modern AI stack.
Google is backing the startup as a founding investor and cloud partner, with a shared research framework for ML systems. Radical Ventures and Khosla Ventures led the seed round, with Kleiner Perkins, Lightspeed, and Doerr Capital participating.
The recursive self-improvement angle
Discovery Loop’s real bet is that AI can build better AI. Dean told the New York Times the goal is to “more fully automate what has traditionally been a very human-intensive experimental loop,” cutting human iteration out entirely at the limit.
That’s the same direction OpenAI’s Astra math breakthroughs point: models that don’t just assist research but run it.
The difference is scale. Jeff Dean’s team built the systems that made internet-scale computation work, and they’re betting that experience transfers to automated experimentation.
I’ve watched this talent pattern before. Noam Shazeer left for OpenAI, Peter Norvig went to Recursive Superintelligence, and now four of the biggest names in one group walk out together. Alphabet’s reaction is telling: instead of fighting the exit, it’s funding the company.
My take
Jeff Dean leaving Google is the end of an era, and Alphabet knows it. The company that built modern AI infrastructure is watching its founders leave with a blessing and a checkbook.
The interesting part isn’t that Dean left. It’s that Alphabet is funding the company that will chase the same frontier. Google is betting that Discovery Loop’s breakthroughs flow back through the research framework and cloud contract.
That’s the same open-weight wave playbook from a different direction: feed the infrastructure, own the returns. Except here the infrastructure is a research loop, and the returns are scientific breakthroughs instead of inference fees.
Discovery Loop’s ambition is the part to watch. Automating the experimental loop is the kind of moonshot that either changes how science gets done or stays a funding story. With this team, I’d bet on the first outcome, but the timeline is the question.
What happens at Google next
The reshuffle gives Koray Kavukcuoglu the DeepMind operating seat at a bad moment. Gemini 4 missed its June window, and coding benchmarks lag rivals.
The talent drain is now visible at the top. Losing Jeff Dean, the engineer most associated with Google’s technical dominance, is the kind of departure that rattles the rank and file even when it’s friendly.
Pichai’s memo leans on momentum: the Gemini app passed 950 million monthly users and Gemma models passed 900 million downloads.
Demis Hassabis stays in the arena as Alphabet chief scientist, which matters. He framed the change around AGI being “close at hand” and doubled down on Isomorphic Labs, his drug discovery company. The science-automation bet isn’t leaving Google, it’s just splitting into two companies now.
For builders, the practical takeaway is simple. The people who shipped the infrastructure you run on are now betting that AI can run science the way it runs software. If they’re right, the next few years of research tools come from a six-person lab in Palo Alto, not a 200,000-person company.
