What exactly changed at Google DeepMind?

Demis Hassabis stepped back from day-to-day leadership of Google DeepMind to become its chair, and simultaneously took a newly created role as Alphabet's chief scientist. Koray Kavukcuoglu, previously DeepMind's chief technology officer, now runs daily operations as senior vice president of Google DeepMind, reporting directly to Sundar Pichai while retaining his separate role as Google's chief AI architect. Hassabis continues to lead Isomorphic Labs.

Who left Google, and where did they go?

Jeff Dean departed after 27 years, alongside senior fellow Sanjay Ghemawat, DeepMind vice president Oriol Vinyals, and Google Brain co-founder Quoc Le. The four are founding Discovery Loop, a public benefit corporation whose stated aim is to automate the process of discovery in machine learning, science and engineering.

Who is funding Discovery Loop?

The initial round is co-led by Radical Ventures and Khosla Ventures, with participation from Lightspeed, Kleiner Perkins, Doerr Capital and Alphabet. Reporting indicates Google will also serve as the company's cloud provider. Wilson Sonsini advised on the launch and initial funding.

What will Discovery Loop actually build?

At launch the stated focus is automating machine-learning research and engineering itself — the loop of proposing, running and evaluating experiments. The company has signalled later expansion into hardware design, drug discovery and clean energy. No product, model or technical paper has been published.

Is this a sign that Google is in trouble?

That reading is not supported by what was actually announced. Hassabis remains chair and gains an Alphabet-wide title, Kavukcuoglu was already the CTO and is a continuity appointment rather than an outside hire, and Alphabet is an investor in the departing team's company. It is a restructuring in which Google retains a stake in both outcomes. Whether it works is a separate question from whether it is a crisis.

What does this mean for Gemini and DeepMind's research output?

Nothing announced changes model roadmaps or research direction. The operational risk is concentrated in the departures, not the reshuffle: Dean and Ghemawat built much of Google's distributed systems and ML infrastructure lineage, and Vinyals and Le are core research leaders. Institutional knowledge of that depth does not transfer in a handover document. Expect no visible effect this quarter and an unmeasurable one after that.

Why does the public benefit corporation structure matter?

A PBC is a for-profit company that is legally permitted — not required — to weigh a stated public mission against shareholder returns. It raises conventional venture capital and can exit conventionally. In practice it is a governance signal about intent rather than a binding constraint, and it has become the default structure for AI labs that want research framing alongside venture funding.

Google DeepMind Loses Its CEO and Chief Scientist in One Day

On 5 August, Google DeepMind announced that Demis Hassabis was moving from chief executive to chair. On the same day, Jeff Dean announced he was leaving Google after 27 years — with Sanjay Ghemawat, Oriol Vinyals and Quoc Le — to found a new company.

Two announcements, one day. Most of the coverage read them as one event: a lab losing its leadership. That reading misses the more interesting detail, which is in the cap table.

Alphabet is an investor in the company its people left to build.

Executive summary

Demis Hassabis moves from CEO to chair of Google DeepMind, and takes a newly created role as Alphabet chief scientist. He continues to lead Isomorphic Labs.
Koray Kavukcuoglu, previously DeepMind CTO, takes daily operations as SVP of Google DeepMind, reporting to Sundar Pichai, while remaining Google’s chief AI architect.
Jeff Dean departs after 27 years, with Sanjay Ghemawat, Oriol Vinyals and Quoc Le.
The four founded Discovery Loop, a public benefit corporation aiming to automate discovery in ML, science and engineering — later expanding to hardware design, drug discovery and clean energy.
Funding is co-led by Radical Ventures and Khosla Ventures, with Lightspeed, Kleiner Perkins, Doerr Capital and Alphabet participating. Google is reported to be the cloud provider.
No product, model or paper exists yet. Everything technical about Discovery Loop is currently a stated intention.

The reshuffle is not the story

Read the Google DeepMind change on its own and it is unremarkable, close to routine.

A founder-CEO who has run a research lab through its transition into a product organisation hands operations to the CTO who has been running the technical side anyway, and takes a title that lets him think about direction instead of headcount. Hassabis said as much: he wanted “the time and space to focus on the big picture.” Kavukcuoglu is an internal promotion, not a search-firm hire, which is what continuity looks like. Hassabis keeps the chair seat and gains scope across all of Alphabet.

If that had been the only announcement, it would have been a personnel note.

The departure is the story, and its shape is unusual

Four people left. Not four engineers — four of the most consequential technical people Google has employed.

Dean and Ghemawat’s joint work is a substantial fraction of the reason large-scale distributed computing looks the way it does. Vinyals was a DeepMind vice president. Le co-founded Google Brain. This is not a team leaving; it is a lineage leaving.

And Alphabet wrote a cheque.

That is the detail worth sitting with, because it changes what kind of event this is. A competitive departure looks like litigation, non-solicits and silence. This looks like a spin-out that was negotiated: Alphabet invests, Google provides the compute, the founders get an independent company, and the mission is framed as a public benefit corporation.

Whatever the internal reasoning, the outcome is that Alphabet has optionality on both sides of a bet it could not otherwise place. If automating ML research works, it holds equity and sells the cloud capacity underneath it. If it does not, the cost was a minority investment and four salaries it was paying anyway.

What “automating discovery” is actually claiming

Discovery Loop’s stated aim is to automate the process of discovery and progress in machine learning, science and engineering — starting with ML research and engineering itself, then branching into hardware design, drug discovery and clean energy.

Starting with ML research is the right first target, and not for the recursive-self-improvement reasons that phrase usually attracts. It is the right target because ML research is the rare scientific domain where the full experimental loop is already software. Proposing a change, running it, and measuring the result requires no laboratory, no reagents, no fabrication step and no waiting. The feedback loop is bounded by compute, and compute is precisely what a Google cloud partnership supplies.

Drug discovery and hardware design are much harder for the same reason in reverse: the loop leaves the computer, and everything outside the computer is slow.

There is a useful caution here from the research literature the same week. A position paper at ICML 2026 argues that language models handle induction — pattern matching — and increasingly handle deduction, but lack the mechanism for abduction: the leap that generates the new premises in the first place. Its author has been careful to note it is a personal position rather than an institutional one, and it is an argument rather than an experiment. But it names the correct dividing line for a company like this.

Automating the running and evaluating of experiments is a large, tractable, valuable problem, and current systems are already good at it. Automating the choosing of which experiment is worth running is the harder claim, and it is the one on which “automating discovery” either means something or does not.

Nothing in the launch tells us which one Discovery Loop is building. That is not a criticism — it is day one, and there is no product, no model, no paper. It is a note about what to look for when there is.

What this actually predicts

Three things follow with reasonable confidence, and one does not.

Talent gravity has inverted. The interesting fact is not that senior researchers left a large lab. It is that four people with unlimited internal resources concluded a startup was the better vehicle, and their employer agreed enough to fund it. That is a statement about the speed of decision-making inside large organisations, not about their compute budgets.

The spin-out-with-investment pattern will repeat. It resolves a genuine problem for both sides: incumbents cannot move fast enough internally to keep their best researchers, and researchers cannot raise frontier-scale compute alone. Expect more of these, structured similarly.

The PBC framing means less than it reads. A public benefit corporation is permitted to weigh a stated mission against returns; it is not compelled to. It raises normal venture money and exits normally. It has become the default legal wrapper for AI labs that want research framing alongside venture funding, and it should be read as a signal of intent rather than a constraint on behaviour.

What does not follow is any conclusion about Google’s model roadmap. Nothing announced touches Gemini, research direction or publication strategy, and Kavukcuoglu has been running the technical organisation for years. Anyone reading this as a capability story is reading a governance document as a benchmark table.

The verdict

The reshuffle is continuity dressed as change. The departure is change dressed as continuity.

Google gave up four of its most senior technical people and bought a position in what they do next. That is a reasonable trade if you believe the constraint on frontier research is organisational speed rather than resources — and the fact that Alphabet made it suggests somebody there does believe that.

Discovery Loop has no product, no published research and an enormously ambitious mission statement. Judge it on its first paper, not its founding team. The founding team only tells you the ceiling.