Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le are leaving Google DeepMind to found Discovery Loop, a Public Benefit Corporation aimed at automating scientific and engineering research. The seed round is led by Radical Ventures and Khosla Ventures, with participation from Alphabet. On the same day, Demis Hassabis moves from CEO to Chair of DeepMind and Chief Scientist of Alphabet, handing operational control to Koray Kavukcuoglu, who becomes SVP overseeing Gemini, frontier research and product teams.

The move tells two things that matter for those building with AI. First: frontier labs stop being monoliths. Deep research and product teams now live under different roofs, because their rhythms and metrics are incompatible. Dean and colleagues didn’t found another generalist model startup: they’re targeting a vertical domain, automated scientific discovery. Second: specialization beats scale. When you have the strongest name in AI infrastructure and choose a narrow target, the message is clear. Generalist models leave thin margins and value shifts toward those solving a specific problem completely.

For those bringing AI into their organizations, the signal is practical. Your vendors are reorganizing around specializations. Google is positioning Gemini as an operational product, while frontier research fragments into smaller, targeted initiatives. When choosing a supplier, look at who owns operational responsibility for the product you actually use, not just the lab name on paper.

As we reported in July when Hassabis proposed a global US-led AI watchdog, DeepMind’s governance was already shifting (/en/radar/hassabis-us-led-ai-watchdog). Today that shift becomes explicit reorganization.

In detail

The change has two sides that need separate reading.

DeepMind’s governance. Hassabis moves to Chair and Chief Scientist of Alphabet. The official description speaks of focus on long-term strategy, AGI and science, with strengthened commitment to Isomorphic Labs (the drug discovery startup already inside Alphabet). Koray Kavukcuoglu, previously CTO, takes operational control as SVP: Gemini, frontier research, product and development teams. This is the first time since DeepMind and Google Brain merged that operational leadership has changed hands. The ecosystem narrative reads two ways: a governance reset to sharpen Gemini execution, which hasn’t updated the Pro version in six months, and a clean separation between long-term thinking and product delivery.

Discovery Loop’s launch. Dean, Ghemawat, Vinyals and Le form one of the strongest founding teams in AI infrastructure history. Dean and Ghemawat built MapReduce, BigTable and the infrastructure on which Google scaled machine learning for twenty years. Vinyals led work on reinforcement learning (AlphaStar) and multi-agent systems. Quoc Le was the principal researcher behind Sequence-to-Sequence and many foundational transformer advances.

The stated mission is to automate the scientific discovery cycle: formulate hypotheses, design experiments, run them, interpret results, iterate. Not a generalist model, but a system channeling ML and automation toward research and engineering problems. The Public Benefit Corporation structure suggests a governance model balancing investor returns and scientific mission.

Google is investing in the seed round. Sources stress the departures are friendly, but one question remains: why wasn’t Discovery Loop born inside Google? Sources don’t give an explicit answer. Context helps read it: approval cycles at a giant like Alphabet are barely compatible with the iteration speed a startup for automated research needs. Dean and colleagues want execution freedom, not a bigger budget.

What changes for you. Two concrete implications. On product: Gemini now has a single operational owner, which could speed releases or concentrate risk on one person. If you use Gemini in your pipelines, the Pro version’s six-month silence is the signal to watch. On research: AI for science stops being a side project in a general lab and becomes a core mission of an independent company. For those working in applied research, it means automated discovery tools could arrive as product sooner than as papers.

What we don’t know. We don’t know the financial terms of the round or a timeline for a first product. Discovery Loop has a team, a mission and investors, but nothing public on architecture or timeline. The narrative of “monolith fragmenting” fits the trend (see also Sprocket and vertical agents), but remains an interpretation, not proven fact. Available sources are all from the newsletter ecosystem and tweets from those directly involved: no independent investigation into the internal reasons for reorganization.