Jeff Dean Leaves Google; Hassabis Becomes DeepMind Chair

Google reshaped the top of its AI organization in a single day. On August 5, 2026, the company confirmed that Jeff Dean, its chief scientist and a 27-year veteran, is leaving to co-found a startup called Discovery Loop, and that Demis Hassabis is stepping back from running Google DeepMind day to day to become its Chair and Chief Scientist of Alphabet. Koray Kavukcuoglu, previously DeepMind’s CTO and Alphabet’s chief AI architect, is being promoted to Senior Vice President of Google DeepMind, reporting directly to CEO Sundar Pichai.

The changes touch two of the most consequential figures in modern computing and land at a moment when Google is fighting to keep its Gemini models at the frontier against OpenAI and Anthropic. Alphabet shares fell roughly 5% on the news.

Who is leaving, and who is joining

Jeff Dean joined Google in 1999 and, with his longtime collaborator Sanjay Ghemawat, built much of the infrastructure that made large-scale computing possible: MapReduce, Bigtable, Spanner, and the Google File System. Dean later co-founded Google Brain and, after the 2023 merger of Brain and DeepMind, became Google’s chief scientist. Ghemawat is leaving with him.

They are not going alone. Two other senior researchers are joining Discovery Loop as co-founders:

• Oriol Vinyals, a Gemini co-lead whose 2014 sequence-to-sequence paper helped establish the encoder-decoder paradigm behind modern machine translation.
• Quoc Le, whose research helped popularize pre-training large models on massive datasets — a foundational idea behind today’s large language models.

Google framed the departures as amicable. The company is participating as an investor in Discovery Loop and will serve as a cloud partner, meaning Dean’s new venture will run at least partly on Google infrastructure even as it operates independently.

What Discovery Loop is trying to build

Discovery Loop is incorporated as a public benefit corporation, and its stated mission is to automate the scientific method itself. Rather than having humans run experiments one after another, the founders want to build AI systems that propose an experiment, run it, learn from the result, and iterate — thousands of times in parallel, recursively.

The team plans to start with machine learning research — using AI to design better AI — before pointing the same systems at the physical world. The named targets are ambitious: engineering, medicine, materials design, and clean energy. It is the “recursive self-improvement” thesis that has circulated in AI circles for years, now packaged as a company with a marquee founding team.

The initial funding round is being co-led by Radical Ventures and Khosla Ventures, with participation from Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet. The founders’ pitch leans on their track record: the people who built the systems that let Google index the web and train frontier models now want to build the system that automates discovery.

Hassabis moves up and out of daily operations

Hassabis co-founded DeepMind in 2010, sold it to Google in 2014, and has run it — and then the merged Google DeepMind — through the Gemini era. He shared the 2024 Nobel Prize in Chemistry for AlphaFold’s work on protein structure prediction, research now commercialized through Isomorphic Labs, which he will continue to lead.

In his new role as Chair of Google DeepMind and Chief Scientist of Alphabet, Hassabis is stepping away from operational management to focus on long-horizon research and Alphabet’s overall AI-for-science strategy. Google is positioning the move as a promotion that frees him to concentrate on artificial general intelligence and scientific breakthroughs rather than a departure — but the practical effect is that the person who has run DeepMind for over a decade is handing the controls to someone else.

Kavukcuoglu takes the controls

That someone is Koray Kavukcuoglu, who joined DeepMind in 2012 and rose to CTO before adding the chief AI architect title across Alphabet. He has been closely involved in shipping Gemini, and his promotion to SVP reporting to Pichai signals that Google wants the person running the model roadmap to also run the organization that builds it.

The consolidation is notable. Where DeepMind previously balanced Hassabis on research direction and Dean as chief scientist across Google, the operational weight now sits with a single executive who owns both the architecture and the org chart. For a company that has struggled at times to move as fast as its smaller rivals, tightening that line of authority is the clear intent.

The context: Gemini’s uneasy frontier

The shakeup does not happen in a vacuum. Google has spent 2026 racing to keep Gemini competitive, shipping rapid updates and pushing AI into Search even as that transition threatens its core advertising business. The company’s capital spending has surged to fund the data centers and accelerators those models require, and investors have grown sensitive to any sign that the returns are not keeping pace.

Losing a chief scientist of Dean’s stature — plus two Gemini-caliber researchers — to a startup building recursive AI research systems is exactly the kind of headline that unsettles that trade. The 5% drop in Alphabet’s stock reflects the market pricing in both the talent loss and the uncertainty of a leadership handoff at the worst possible time to look distracted.

What it means

The talent war has entered its endgame. For most of the past three years, the story was labs poaching each other’s researchers with nine-figure packages. Dean’s exit is a different signal: the most senior people at the most established labs now believe the frontier is moving fast enough that a startup with the right team and enough compute can compete from a standing start. When a 27-year Google veteran leaves to build recursive AI-for-science, it says the ceiling on what a new entrant can attempt has risen sharply.

Google is betting on focus over star power. Consolidating operational control under Kavukcuoglu while elevating Hassabis to a research chairmanship is a bet that a cleaner command structure ships better models than a constellation of famous names. That can work — decisive execution has been Google’s weak point relative to OpenAI and Anthropic — but it also removes the two people most associated with DeepMind’s scientific credibility from the day-to-day. The Gemini roadmap now has to carry that reputation on its own.

Discovery Loop is the story to watch. If the recursive-research thesis works even partially — AI systems that meaningfully accelerate ML research or materials design — it validates the most aggressive claims about self-improving AI and reshapes how every lab allocates researchers. If it stalls, it becomes another cautionary tale about star founders and grand visions. Either way, Google has hedged: it’s an investor and the cloud provider, so it captures upside and keeps a window into what its former scientists build next.

The near-term risk is continuity. Frontier model development runs on institutional knowledge and tightly coupled teams. Losing Dean, Ghemawat, Vinyals, and Le in one stroke removes deep expertise from the Gemini effort precisely as the next model generation is being trained. Google’s challenge over the coming quarters is proving that the bench is deep enough that the loss of its most decorated researchers doesn’t show up in the product.