Demis Hassabis steps down, Jeff Dean leaves Google
Demis Hassabis becomes Google DeepMind chair and Alphabet chief scientist; Jeff Dean exits after 27 years to found Discovery Loop. What changed, and why.
Amara Singh is a seasoned technology journalist with a background in computer science from the Indian Institute of Technology. She has covered AI and machine learning trends across Asia and Silicon Valley for over a decade.
On August 5, 2026, Sundar Pichai published a memo titled "The next chapter of our AI momentum." Two things happened inside it. Demis Hassabis stopped being chief executive of Google DeepMind. Jeff Dean stopped being an employee of Google. Everything else in the memo is scaffolding around those two facts.
Both moves have been read as a single "AI shakeup," which is the wrong frame. They point in opposite directions. One pulls a founder deeper into Alphabet. The other pushes four of the company's most senior researchers out of it, with Alphabet's money following them out the door.
What actually changed on the org chart
Hassabis is now Chair of Google DeepMind and Chief Scientist of Alphabet. He keeps running Isomorphic Labs, the drug-discovery spinout. What he gives up is day-to-day operational control of the unit he co-founded in 2010 and sold to Google in 2014. In his framing, the point is to concentrate on the long-range AGI question rather than shipping cycles: "It's critical that we collectively get the next steps right to ensure this all goes well for humanity."
The operating job goes to Koray Kavukcuoglu, until now DeepMind's chief technology officer and a researcher there since the years before the Google acquisition. His new title is SVP of Google DeepMind, and the reporting line is the detail worth reading twice: he reports to Pichai, not to Hassabis. Gemini model development, frontier research, and the Gemini app and developer teams all sit under him.
So the chair role is not a demotion dressed as a promotion, and it is not a promotion either. It is a separation. Research direction and product execution, previously fused in one person, are now two boxes with two independent reporting lines into the same CEO. Google has run this experiment before, in the other direction: the 2023 merger of Google Brain and DeepMind existed precisely to collapse two research organizations into one chain of command under Hassabis. Three years later, the company is splitting the function again, along a different seam.
What Jeff Dean is actually doing
Dean leaves after 27 years to co-found Discovery Loop, a public benefit corporation. He is not going alone. Sanjay Ghemawat, Oriol Vinyals, and Quoc Le are co-founders. That roster is the story.
It is worth being concrete about what these four built, because the abstraction "senior researchers" undersells it. Dean and Ghemawat authored the infrastructure layer that made Google's scale possible and then made everyone else's possible too: the Google File System, MapReduce, Bigtable, and Spanner, a lineage that the entire open-source big-data ecosystem was reverse-engineered from. Dean went on to co-found Google Brain, drive DistBelief and then TensorFlow, and sit close to the TPU program that gave Google its one durable hardware advantage over competitors renting Nvidia silicon.
Vinyals co-led Gemini and, before that, led AlphaStar; he is a co-author of the sequence-to-sequence paper that reframed translation and much else as a learned mapping problem. Le co-authored that same paper and then pushed the neural architecture search and AutoML line of work, which is the closest existing antecedent to what Discovery Loop says it wants to do. These are not four people who happened to resign in the same week. They are a research group with a shared thirty-year working history choosing to keep working together somewhere else.
Discovery Loop's stated goal is to "automatically solve important problems in machine learning, science, and engineering." Concretely, it wants systems that run the whole experimental loop without a human in each step: propose an experiment, implement it, run it, evaluate the result, iterate. The first target is recursive. The team plans to automate machine learning research and engineering first, then apply that capability to its own stack before extending into hardware design, drug discovery, and clean energy.
Alphabet is a founding investor and the cloud provider. Radical Ventures and Khosla Ventures co-led the seed.
The public benefit corporation structure is worth a note, because it keeps being described as if it were a charity. It is not. A PBC is an ordinary for-profit company with ordinary shareholders; the only difference is that its charter names a public purpose alongside profit, which gives directors legal cover to weigh that purpose in decisions without breaching their duty to investors. OpenAI's restructured for-profit arm and Anthropic both use the form. It signals intent and it shapes what a board can defend in court. It does not cap returns, and it does not stop anyone from selling.
Read the two moves together
Alphabet shed roughly 5% of its market value on the day it announced it had converted four of its most senior researchers from employees into a cap-table position, and that is the honest summary of the trade.
Google keeps optionality. If Discovery Loop works, Alphabet owns equity in it and sells it the compute it burns, which means a meaningful share of the invested capital returns as Cloud revenue regardless of the scientific outcome. If it does not work, the cost was a seed check rather than four executive compensation packages and the internal friction of housing an open-ended research bet next to a quarterly product roadmap.
This is the structure Microsoft used with OpenAI and Amazon used with Anthropic, applied to talent Google already had. The novelty is that the counterparty was internal until the day it was not.
There is also a precedent inside Google for how this ends. The eight authors of the 2017 Transformer paper nearly all left, and the startups they founded (Character.AI, Cohere, Adept, Sakana) became the second generation of the field. Google then paid roughly $2.7 billion in an August 2024 licensing arrangement with Character.AI whose main effect was to bring Noam Shazeer back and install him as a Gemini co-lead. The Justice Department has since examined whether that structure was designed to avoid merger review.
Read against that history, the Discovery Loop investment looks less like a farewell and more like a call option written early. Google learned that letting a research group walk out unfunded costs billions to reverse. Funding them on the way out is cheaper than reacquiring them later, and it buys information Google would otherwise have to pay a premium for.
What Google is actually giving up
The cost is harder to price than the stock move suggests.
Automating machine learning research is the capability that compounds. Whoever gets it first improves everything downstream of it faster than competitors can respond, because the improvement applies to the improving system. That is the entire thesis behind the phrase "recursive self-improvement," stripped of its more speculative connotations. Google is now a shareholder in that outcome rather than the owner of it, and a shareholder does not set the research agenda, the safety posture, or the publication policy.
There is a second-order cost too. Dean and Ghemawat were not only researchers; they were the technical arbiters of last resort, the people whose design review settled arguments that no process could. Organizations rarely notice the loss of that function until a decision that used to take a week starts taking a quarter.
The reshuffle is best understood not as a succession but as Google deciding which parts of its research agenda belong on the balance sheet and which belong on the org chart. Gemini and frontier research stay inside, under an operator reporting to the CEO. AGI strategy stays inside, under a founder with a chair title and no operating load. Open-ended automation of the scientific method goes outside, funded but not managed.
What to watch next
Three things will tell you whether this held.
Whether Kavukcuoglu's Gemini cadence changes. He inherits model development, the app, and developers under a single line. If shipping speed rises through the next two Gemini releases, the separation worked. If the unit slows while it re-forms, the fused structure was doing more work than it appeared to.
Whether Hassabis's chair role has teeth. Chief Scientist of Alphabet is a title with no headcount attached. Its power comes entirely from whether Pichai routes decisions through it. Watch who signs off on the next frontier-model safety or release decision, and whether Isomorphic gets more of Hassabis's calendar than DeepMind does.
Whether more researchers follow Dean out. Four co-founders left together, and Discovery Loop's own founding note describes a group that has worked together for between 14 and 30 years. Groups with that much shared history recruit from the same bench, and a startup with Alphabet's compute behind it can pay in equity that a mature company cannot match. The second wave, if there is one, will say more about Google's retention than the first did.
Neither move was forced by a bad quarter. Google's AI position in mid-2026 is strong, Gemini is competitive at the frontier, and the TPU advantage is real. That is what makes the reshuffle legible. This is a company with room to choose, choosing to hold its product research close and its most speculative research at arm's length, with a check attached and a seat on the cap table instead of a seat on the org chart.