Key Points
・On August 5, 2026, Alphabet announced that Demis Hassabis would step away from the day to day running of Google DeepMind, and on the same day Jeff Dean and three colleagues left to found Discovery Loop, a company built to automate scientific research.
・DeepMind made its name with AlphaGo and AlphaFold, then merged with Google Brain after ChatGPT and turned into the division that ships Gemini.
・The most useful reading is not a talent exodus but a functional split: frontier AI has grown expensive enough that research for revenue and research for science are being placed in separate buildings.
▼ Watch this article as a video
Hassabis Changes Seats and Four Researchers Leave
Alphabet announced a restructuring of its AI organisation on August 5, 2026. Demis Hassabis, who had served as chief executive of Google DeepMind, becomes chairman of that unit and chief scientist of Alphabet. He steps away from daily operations to concentrate on long term research toward AGI.
In a memo to staff, Hassabis cited the approach of AGI as his reason for changing roles. He remains chief executive of Isomorphic Labs, the drug discovery company spun out of DeepMind and held under Alphabet.
Day to day operations and Gemini model development pass to current chief technology officer Koray Kavukcuoglu. His title is senior vice president of Google DeepMind rather than chief executive. He reports directly to Sundar Pichai and continues to hold the Chief AI Architect role at Google that he took up in June 2025. His remit covers Gemini model development, frontier AI research, and the Gemini app and developer teams.
On the same day, Google chief scientist Jeff Dean and three colleagues announced they were leaving to found a company called Discovery Loop. The other three are Sanjay Ghemawat, Oriol Vinyals and Quoc Le. Based in Palo Alto and incorporated as a public benefit corporation, it will work on automating scientific research with AI.
Google stated on its official blog that Alphabet is a founding investor in and cloud partner to the new company. Alphabet has not explained the background to these moves or why they were announced together. Alphabet shares were reported to have fallen 4 percent after the announcement.
Related Articles
From a London Lab to Google’s Core Business
Who is leaving, and what they built
The four departures land heavily because of what these particular people built rather than how many of them there are.
Jeff Dean joined Google in 1999, when the company had a few dozen employees, and spent 27 years designing the successive generations of infrastructure that spread web scale workloads across hundreds of thousands of machines. He co-founded the internal AI research team Google Brain in 2011, worked on TensorFlow, and has been chief scientist for Google as a whole since 2023. Sanjay Ghemawat co-designed that same infrastructure with him. Oriol Vinyals and Quoc Le were leading Gemini model research until recently.
Discovery Loop proposes to make AI run the loop of science itself: form a hypothesis, run an experiment, analyse the result, form the next hypothesis. It is a for profit company with a public benefit charter. Alphabet and several venture funds have invested, with amounts undisclosed, and Google will supply compute for at least the first year.
How a research lab became the core of the business
The short version is that a lab founded far from Google’s business became its lifeline in roughly a decade.
| Date | Event |
|---|---|
| 2010 | DeepMind founded in London |
| January 2014 | Google acquires DeepMind (terms undisclosed) |
| March 2016 | AlphaGo beats Lee Sedol 4-1 |
| November 2020 | AlphaFold2 achieves a breakthrough in protein structure prediction |
| November 2022 | ChatGPT released publicly |
| April 2023 | Google Brain and DeepMind merge into Google DeepMind |
| October 2024 | Hassabis shares the Nobel Prize in Chemistry |
| November 2025 | Gemini 3 released |
DeepMind was founded in London in 2010 by a group whose stated goal was AGI, and it kept a lab-like freedom for years after the 2014 acquisition. AlphaGo and AlphaFold were the results, and AlphaFold brought Hassabis a share of the 2024 Nobel Prize in Chemistry. Both were demonstrations of AI as an instrument of science rather than a product line.
ChatGPT changed that in November 2022. The company most associated with AI had been beaten to the product moment by OpenAI. Google merged its internal AI research team with DeepMind the following year and put Hassabis in charge. Co-founder Sergey Brin, long retired from operations, is reported to have returned to the floor and become deeply involved in Gemini around the same period.
The lab for AGI and science thereby became the division that builds Gemini. Gemini 3, released in November 2025, was judged to have closed much of the gap with rivals. The next model, Gemini 4, had been expected in June and had still not shipped when the August restructuring was announced.
The 2020 precedent at OpenAI
English language readers will recognise the shape of this immediately, because it happened five years ago at OpenAI.
Dario Amodei, then head of research at OpenAI and one of the central figures behind GPT-2 and GPT-3, left the company at the end of 2020. In 2021 he founded Anthropic with his sister Daniela and other OpenAI alumni. At the time Anthropic read as a research house focused on reliable, steerable AI. Today it fields Claude and competes head on with OpenAI and Google.
That precedent is why departures from a frontier lab now carry so much weight. Independence has already produced a rival to the parent once. Whether the shape repeating means the substance repeats is a separate question, and it is the question this article turns on.
One more data point sits just before the restructuring. In June 2026, John Jumper, who shared the Nobel Prize in Chemistry with Hassabis, announced he was leaving Google DeepMind for Anthropic.
What the Talent Exodus Story Misses
Half right, and half misleading
Reading August 5 as a story of talent flight gets you halfway and then walks you into a wall.
The flight side is real. Reuters carried a market strategist’s assessment that the absence of someone of Jeff Dean’s standing will be felt for some time. Nor was it isolated. In June, Noam Shazeer, a Gemini co-lead, went to OpenAI, and in the same week John Jumper announced his move to Anthropic. The four departures in August sit at the end of that line.
The AlphaFold team, DeepMind’s most celebrated group, has been reported disbanded, with key members reassigned to Gemini work or to Isomorphic Labs. Around a quarter of the authors of the original paper are reported to have left the company. There is room to read that as scientific research losing its footing inside Google.
Set against that are facts a pure exodus story cannot absorb. Hassabis is staying at Google. Alphabet is a founding investor in Discovery Loop and is supplying its compute. Gemini development passed to an operational lead reporting to Pichai in what looks like a planned handover. Companies do not fund defections to the enemy.
Reading Hassabis’s move as either a demotion or a promotion is the other trap. The title went up, to chief scientist of all of Alphabet, while the company says he will have almost no direct reports. A larger title and a smaller daily grip on the organisation are perfectly compatible. The question worth asking is not up or down but what the authority is now attached to.
The details point the same way. Kavukcuoglu’s centre of gravity is Mountain View rather than London, and communications and legal functions are reported to be folding into Google proper. Executives at Google Cloud, the commercial engine of the AI business, are reported to have welcomed the change.
The Financial Times is reported to have concluded that influence over AI strategy has shifted toward Sergey Brin, and other reporting describes Kavukcuoglu as an operator close to him. The reading that the centre of gravity moved from the research capital in London to the commercial capital in California has a fair amount of material behind it.
It looks like Anthropic, but the capital runs the other way
The 2020 case and the 2026 case are alike in outline and different in three specifics.
The first is the axis of the split. The people who founded Anthropic put their research weight on how a powerful AI should be built and controlled. The four founding Discovery Loop are betting on what AI should be used for. One division is about development philosophy, the other about application domain.
The second is the direction of the capital. OpenAI and Anthropic now compete directly for models and customers. Google has invested in the company its people left for and contracted to sell it compute. The stake and its terms are undisclosed, so the thickness of that tie cannot yet be measured.
The third is how much rupture there was. Anthropic’s founding is described as following differences over development approach. For Google, there is little material at present confirming that the four rejected any company policy. The investment weakens the rupture reading, and a reorganisation of roles, a functional split, currently explains more than an ideological break.
Pull the frame back one step, however, and the two converge again. Microsoft stands behind OpenAI, Amazon and Google are shareholders in Anthropic, and Alphabet is behind Discovery Loop. However independent the lab, frontier research cannot leave the compute of large capital. That has not changed in five years.
Commercialisation is the enemy of research and also its wallet
The clean opposition of research against commerce runs into an inconvenient fact: AlphaFold was itself built with Google’s money.
AlphaGo and AlphaFold2 existed because of the profits of an advertising business and one of the largest compute estates on earth. Commercial pressure narrows the freedom of a research group and pays for it at the same time.
The Financial Times is reported to have described internal unhappiness at the decision to release AlphaFold to the scientific community for free. Free release also buys recruiting, reputation and external validation, which makes this less a clash of idealism against pragmatism than it first appears.
Corporate labs have lived in this tension since long before AI. Bell Labs, which produced the transistor and information theory, could sustain basic research because a telephone monopoly generated an enormous surplus. Xerox PARC, which built the prototype of the personal computer, watched Apple commercialise what it had invented. Surplus from the main business feeds the lab; a threat to the main business restructures the lab. The pattern is half a century old.
What is striking is that the same force ran in opposite directions at OpenAI and Google. OpenAI started as a lab, became a business as development costs ballooned, and shed part of its research-minded staff. Google’s acquired lab became the core business, and its science-minded researchers are leaving.
Opposite entrances, identical structure at the far end. Frontier research requires enormous money, money requires commercialisation, and commercialisation prioritises product velocity. That is not the personality of any one company. It is the tension this industry is built on.
Can Discovery Loop stay a science company?
Anthropic is the reminder that the sign over the door at founding is not a forecast.
In 2021 it was widely read as a small safety focused lab. Within a few years it was a frontier company with Claude and one of OpenAI’s largest rivals. In fairness the company had described general, high performing AI from the start, so outside perception catching up with reality is part of what happened. Either way, the founding label did not predict the destination.
For Discovery Loop today there is nothing to evaluate except a founding statement. What it becomes in five years will be decided by behaviour.
The loop it proposes to automate, hypothesis to experiment to analysis to next hypothesis, is continuous with the capabilities of a general purpose AI agent: set a problem, gather information, iterate, evaluate your own results. TechCrunch reports that the company will also explore recursive self improvement. Capability sharpened as a scientific instrument could turn out to be general capability.
Running autonomous experiments and working across every domain are still different things, though, and a technology being transferable is different again from a company choosing to transfer it.
Rather than guess, here are four observation points for a change of direction. Does the target domain widen beyond drug discovery and materials? Does the company start building its own foundation models? Does it ship an external API or enterprise products? Does its compute sourcing extend beyond Google Cloud? When those move, the sign over the door is being repainted.
Alphabet’s investment can be read as a hedge across all four branches. Success in science returns investment gains and cloud demand; a turn toward general capability stays within reach through the equity. The stake and terms are undisclosed, so how far that reading holds depends on numbers nobody outside has seen.
Can the Lab Survive as a Form?
A reorganisation that sorted research by where it belongs
If the August 5 restructuring did one thing, it sorted research by where it should sit. Research that makes the next Gemini stronger goes to an operational structure reporting to Pichai. Long horizon work on AGI and science goes to Hassabis in his new role. Some of the research that matters without obviously paying goes outside the company entirely, to a Discovery Loop that Alphabet has funded. Roles that shared one building were placed in three.
Whether that was a designed division of labour or an orderly presentation of departures nobody could stop is not something an outsider can settle now. One reading has London’s research culture thinning; another has each part free to concentrate. The answer will come from what follows Gemini 4 and from what Discovery Loop looks like in a few years.
The longer lasting question is the one underneath. Can basic research, product development and revenue responsibility live in a single organisation? Nobody has answered that since Bell Labs, and the frontier AI industry is now facing it with more money than has ever been aimed at the problem. The ten years of a lab called DeepMind will read as the first record of that experiment.
Reference Links
- The next chapter of our AI momentum｜Google
- Google shakes up AI leadership as DeepMind chief shifts role｜Reuters
- Discovery Loop｜Discovery Loop
- Jeff Dean and other top AI researchers are leaving Google to launch their own startup｜TechCrunch
- Google DeepMind Reshuffles After CEO Demis Hassabis Steps Aside｜TIME
- Google DeepMind disbands its Nobel-prize winning AlphaFold team｜Engadget
- Anthropic raises $124 million to build more reliable, general AI systems｜Anthropic
