Jeff Dean Left Google After 27 Years to Automate His Own Job

TL;DR

Google just rewired the top of its AI org chart in a single announcement. Sundar Pichai's memo on Tuesday says Demis Hassabis is stepping back from running Google DeepMind day-to-day to become Chair of DeepMind and Chief Scientist of Alphabet, with longtime CTO Koray Kavukcuoglu promoted to Senior Vice President and reporting straight to Pichai. In the same breath: Jeff Dean, the engineer whose fingerprints are on Google Search, MapReduce, TensorFlow, and the TPU, is leaving after 27 years to co-found Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le. The new company's stated mission is to automate the experimental loop of machine learning research itself. Alphabet is a founding investor and cloud partner anyway, and the market still shaved roughly 4 percent off the stock.

One afternoon, three job changes

Pichai framed the memo as the next chapter of Google's AI momentum. What it actually describes is the simultaneous exit of the two most senior scientific figures in the company's AI effort from their operational roles.

Hassabis, who has run DeepMind since founding it in 2010 and picked up a Nobel Prize in Chemistry along the way, relinquishes day-to-day control. His new remit as Alphabet Chief Scientist is deliberately vague: AGI strategy, the big picture, advising DeepMind leadership on models and research. He keeps running Isomorphic Labs, Alphabet's AI drug-discovery spinout, and stays in London. In a statement quoted by Fortune, he says he believes AGI is "close at hand" and that getting the next steps right is critical.

Dean's move is the sharper break. He was Google's Chief Scientist, employee-badge royalty since 1999, and the person a generation of infrastructure engineers grew up imitating. He is not retiring; he is starting a competitor-adjacent lab with three of the most decorated researchers Google had left.

Who actually runs Gemini now

Kavukcuoglu is the operational winner. He has been at DeepMind for over 13 years, led the deep learning team behind WaveNet and DQN, and already held the Google-wide Chief AI Architect title, which he keeps. As SVP he now oversees Gemini model development, frontier AI research, and the Gemini app and developer teams, without a standalone CEO title and reporting directly to Pichai.

What he inherits is not small. Pichai's memo cites more than 950 million monthly users on the Gemini app and over 900 million downloads of the open-weight Gemma family. The tell is in the structure, though: the research lab that once negotiated its independence from Mountain View now reports to Alphabet's CEO like any other product division.

The founding four

Discovery Loop's founder list reads like the citations section of the last 20 years of computer science. Dean and Ghemawat co-wrote the MapReduce and Bigtable papers and were, for decades, Google's most famous pair-programming duo. Vinyals co-led Gemini and built AlphaStar. Le co-founded Google Brain and did foundational work on sequence-to-sequence learning. The company's own site puts it bluntly: three of the most-cited AI researchers and two of the most-cited distributed systems researchers, on one cap table.

Jeff Dean is the engineer about whom other engineers wrote Chuck Norris-style facts ("Jeff Dean's PIN is the last four digits of pi"). He is now building a company whose explicit goal is to make that kind of engineer less necessary.

What Discovery Loop actually does

The pitch is not another frontier chatbot. Discovery Loop wants to automate the experimental loop of research: hypothesis, experiment, analysis, next hypothesis, with as little human in the middle as possible. Initial focus is automating machine learning research and engineering, using its own tooling on itself first, then expanding toward the National Academy of Engineering's Grand Challenges: medicine, health informatics, solar energy, clean water, cybersecurity.

The mechanism is scale-out, not genius-in-a-box. Frontier research today still mostly moves at the speed of a human staring at loss curves and deciding what to try next. Discovery Loop's bet is to run thousands of experiments in parallel and have the system queue its own follow-ups. Think of it as swapping one lab bench with a brilliant scientist for a warehouse of self-driving lab benches that file their own results and schedule their own next runs. The scientist does not disappear, but the bottleneck does.

Structurally it is an independent public benefit corporation, hiring for what it calls a lean, in-person team.

The money, and the market's verdict

Per TechCrunch, the initial round is co-led by Radical Ventures and Khosla Ventures, with participation from Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet itself. Radical's Jordan Jacobs joins the board. The founders are not disclosing the round's size or valuation, which for a team like this usually means the number would embarrass everyone involved.

Google stays close: founding investor, cloud provider, and a research collaboration framework on ML systems and infrastructure. That makes this less a defection than a spin-out with a handshake. Investors read the org chart anyway and sent Alphabet down roughly 4 to 5 percent on the day, per Fortune and CNBC.

Why builders should care

First, the automation-of-research race just got its most credible entrant. OpenAI has been signaling the same direction with its unreleased Astra results, and METR spent July trying to price what autonomous AI R&D is actually worth. When the person who built the infrastructure layer of modern ML decides the next infrastructure layer is the research loop itself, that is a strong prior about where the leverage is.

Second, Gemini's roadmap now runs through Kavukcuoglu and reports into Pichai, which is the most product-shaped configuration DeepMind has ever had. Whether that speeds Gemini up or grinds the research edge off it is the question every Gemini API customer should be watching over the next few releases.

Third, talent gravity is real. Four principal-level researchers leaving in one day, with Alphabet's blessing and Alphabet's money, tells you where senior researchers think the marginal unit of progress lives: not inside a 180,000-person company, but in a lean team pointing automation at the research process itself.

Key Takeaways

• Demis Hassabis stepped back from running Google DeepMind day-to-day; he is now Chair of DeepMind and Chief Scientist of Alphabet, and still runs Isomorphic Labs.
• Koray Kavukcuoglu, DeepMind's longtime CTO, becomes SVP over Gemini models, frontier research, and the Gemini app and developer teams, reporting directly to Sundar Pichai.
• Jeff Dean left Google after 27 years to co-found Discovery Loop with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le.
• Discovery Loop is a public benefit corporation aiming to automate the ML research loop first, then broader scientific and engineering discovery.
• The round is co-led by Radical Ventures and Khosla Ventures with Alphabet participating as founding investor and cloud partner; size undisclosed.
• Alphabet stock fell roughly 4 to 5 percent on the announcement.

Sources: Google (Sundar Pichai memo), Discovery Loop, TechCrunch, Fortune, The Decoder, 9to5Google