Demis Hassabis calls himself a scientist first. 12 years after buying DeepMind, Google has let him be one

A year ago, in his office at DeepMind's London headquarters, Demis Hassabis told The Guardian something that sounded like a throwaway regret. If he'd had his way, he said, they would have "left it in the lab for longer" and done more things like AlphaFold. Cured cancer, maybe. Instead ChatGPT arrived, the race started, and the man who wanted to be left alone with the proteins became the public face of Google's answer to it.It was not the first time he had said it, or acted on it. Twenty years earlier he had wound up a games company to go and study amnesia.That one he decided for himself. On Wednesday, nine days after his fiftieth birthday, Google decided the second one.The move was dressed as a promotion, and by title it is one. Sir Demis Hassabis becomes Chair of Google DeepMind and Chief Scientist of Alphabet, and keeps Isomorphic Labs, the drug design company he spun out of DeepMind in 2021. What he does not keep is the organisation. Koray Kavukcuoglu, DeepMind's chief technology officer and one of the first researchers Hassabis ever hired, takes Gemini model development, frontier research and the Gemini app, and reports to Sundar Pichai.Everything about it looks like a man being moved aside. Almost nothing about the past twenty years suggests he would mind. Hassabis has never hidden how little he enjoyed the corporate half of his career, and once before, he gave up a company to escape it.By 2005 he had a decade of the games industry behind him: Theme Park at seventeen, a double first at Cambridge, then Elixir Studios, the London developer he founded at 22 and ran for seven years. That year he sold off the rights to the games he had built, wound the company up, and went back to university.What he went back for was not computing. It was cognitive neuroscience, at UCL, under Eleanor Maguire. His first paper, in PNAS in 2007, showed that patients with hippocampal amnesia could not imagine new experiences either: the brain remembers its past and pictures its future with the same machinery. Science named it one of the ten breakthroughs of the year. He founded DeepMind the year after he left.He sold it in 2014. Google paid £400 million for a lab with fifty-odd people, no product and no revenue, which was roughly what Hassabis had told investors to expect when they asked what DeepMind would sell: nothing yet, because it was building the most important thing of all time. What Google had that he wanted was compute, on a scale no independent lab could buy. What Google wanted was less clear to him, and he negotiated for an ethics board and a set of limits on how the work could be used, then went back to the research.It did not hold. Sebastian Mallaby's biography of Hassabis has him and his co-founder Mustafa Suleyman spending close to three years from 2015 trying to prise DeepMind out of Google altogether, under an independent board, with Reid Hoffman prepared to put a billion dollars behind the split. Pichai refused. What Hassabis wanted from that fight was not power but distance, and he lost it.He stayed, and the lab changed around him. DeepMind, he told CNBC in 2024, was "like the engine room of the company"—a line that sounds like a boast until you notice that engine rooms sit below deck. He made his researchers sit on papers for six months before publishing in case what they had found was commercially sensitive, a rule that went down badly in a lab whose scientists had spent a decade measuring themselves by what they got into Nature. The scoreboard changed with it: user growth in, journal citations out.None of this was hidden. Six current and former employees told Business Insider that Hassabis had drifted from the job for about a year, one longtime Googler reducing it to a sentence: he does not care about chatbots, he wants to use AI to cure cancer. The parts he had been shedding were the ones he never wanted—refereeing team politics, sitting through product decisions that did not hold his attention.Google's side of it is less flattering. The Financial Times, citing a dozen people familiar with the company, reported that senior executives had grown frustrated by what they saw as his lesser focus on the commercial demands of the AI business. A person close to the company disputed there had been any tension.Which makes the obvious question one of what he is actually left with. The lazy reading of Wednesday is that Isomorphic is a science project handed to a scientist to keep him busy, and to keep him at all, in a year when Google has been losing people it could not afford to lose. The money says otherwise. In May it closed a $2.1 billion Series B led by Thrive Capital, with Alphabet, Abu Dhabi's MGX and Singapore's Temasek joining, taking outside capital past $2.6 billion. Its partnerships with Novartis and Eli Lilly are worth close to $3 billion between them.This is not a hobby. It is one of the largest private financings in AI drug discovery, and he has been running it alongside a job that already had him working, by his own account, until four in the morning. It is also unproven. Five years in, Isomorphic has yet to put an AI-designed drug into clinical trials.What he intends to do with it he set out in Stockholm in December 2024, accepting a prize for work that started with a question about brains. His argument since UCL has run roughly like this. A brain is not a store of facts; it is a device that builds models of the world and uses them to guess what happens next, which is what imagining and remembering and hypothesising all turn out to be.Build a machine that works the same way and you do not get a better encyclopaedia. You get something that can go and find out what nobody knows yet—and then you point it at the questions where finding out has been the bottleneck for fifty years.The lecture had two halves. The first went to AlphaFold and the fifty-year problem of predicting a protein's three-dimensional shape from its amino acid sequence: the roughly 170,000 experimentally determined structures that trained the system, and the accuracy down to the width of a single atom that led the contest grading the field since 1994 to declare the problem solved in 2020. Then he gave all of it away. Over 200 million structures, released free, used since by more than two million researchers across 190 countries.Not everyone at Google admired the gesture. The FT reported this week that the giveaway became a source of internal tension, an enormous investment returning almost nothing commercially, and that the AlphaFold team has since been broken up. John Jumper, who shared the Nobel with him for it, left for Anthropic earlier this year.Then he generalised it. AlphaFold, he argued, was not a one-off but a template: find a problem with a vast combinatorial search space, a clear objective to optimise against, and enough data or a good enough simulator, and the same method applies. He called the result digital biology, and proposed a conjecture from the podium: that any pattern found in nature can be efficiently discovered and modelled by a classical learning algorithm.Isomorphic is that thesis with patients attached. Its stated goal is to solve all disease, a phrase Hassabis uses without visible embarrassment, and its next target beyond molecules is a virtual cell able to predict how an intervention plays out before anyone touches a pipette. In his memo to staff on Wednesday he wrote that AI's highest application has always been human health, and asked what better demonstration there could be than helping to "cure diseases like cancer."Why now, when Isomorphic has existed since 2021 and the wish to be back in the science for longer than that? Because until now Google could not afford to let him go.ChatGPT arrived in November 2022, Pichai declared a code red, and Larry Page and Sergey Brin were called back in to help find an answer. Brin stayed. He has been in Mountain View most days since, sitting with the researchers training Gemini, arguing about loss curves, weighing in on who gets hired—the technical detail he has said is where his scientific interest actually lies. He is not on the org chart and does not need to be. Kavukcuoglu relocated from London last year and now runs the models from a desk beside his. He is also, by some readings inside the company, on a path towards eventually succeeding Pichai.The symmetry is hard to miss. One founder came back into the trenches because the race is the interesting part; the other spent three years trying to get out of the building. Hassabis built DeepMind in London to prove a frontier lab did not have to sit in Silicon Valley. Gemini is run from Silicon Valley now.If Google was calling a founder back in, it was never going to let Hassabis walk out. Instead it did the opposite, merging DeepMind with Google Brain, its own in-house AI lab, in 2023 and putting him over both, then folding in the safety and research teams and the Gemini apps business until the lab was the company's AI engine in fact as well as in his phrasing.Nothing about Hassabis changed after the merge. What changed was what Google needed from the person holding his job. Gemini 3 had put Google back in front after a bruising year, which in this business means very little for very long. Every lab has led it at some point and been overtaken within months of saying so, and Google is behind again already—the follow-up has slipped since June, and Pichai conceded on the last earnings call that the company needs to improve at agentic coding.Train, ship, lead for a quarter, get overtaken, train again. That is the job now. Somebody has to want it, and by every account of the past year, he no longer did.Hassabis has a phrase he keeps returning to for where we are: the foothills of the singularity. Pichai quoted it back at him on Wednesday, framing the new role as one that lets Hassabis give his full attention to shaping what AGI becomes.What he means by it is narrower than what his peers mean. At Davos in January he held to five to ten years, and said his bar was higher than theirs because he counts AGI as a system that can produce new hypotheses, not one that performs economically valuable tasks. What is still missing, he said, is continual learning, better memory, robotics that works, and world models—systems that understand the physical world well enough to predict it, which he works on personally.A machine that can have an idea nobody has had is a scientist's target. He set it before Google owned any of this, and it is not reached by shipping quarterly. Some of it has started arriving anyway: DeepMind systems took gold at the International Mathematical Olympiad last summer and are now pointed at Millennium Prize problems.Five to ten years is a short time to spend in the wrong job. He has said for twenty of them that the first thing to do with AGI is cure disease, and whoever does it first will be whoever has already built the apparatus to use it. Isomorphic is his.DeepMind is not, and that is the price. The lab he founded in 2010 and spent sixteen years shielding is now run out of California, by a man he hired, reporting to a chief executive who once refused to let him take it away.London took it hard. Current and former staff told the FT the news went through the building like a shock, some of them convinced it ends the research culture Hassabis spent a decade defending, and a former executive at a rival firm said the calls from people ready to leave started that day. A friend told the paper Hassabis himself was relieved. Both of those can be true at once, and probably are.The DeepMind he is handing over is not the one he built. Everyone who worked there can see that research is no longer what it is for, which is why the other half of Wednesday's news was Jeff Dean leaving after twenty-seven years to start a company that will do the thing DeepMind stopped making room for. Hassabis says the models are in good hands with Kavukcuoglu, and so, presumably, is the frontier that has been late since June.What he takes with him is what he came for in the first place: the science. It is the narrowest job he has held in a decade and the only one he ever wanted, and it is not the retreat it looks like. He still thinks a machine will one day have an idea nobody has had. He would rather be standing in a lab when it does.