Part IThe Story
On Thursday, October 10, 2024, the day after the Royal Swedish Academy of Sciences announced that half of that year's Nobel Prize in Chemistry would go to Demis Hassabis and John Jumper, the newest chemistry laureate in Britain was at a card table. A childhood chess friend was hosting a poker-and-chess evening at his house in London. The timing was luck: some of the world's best players were in the city for a large tournament. Magnus Carlsen came. So did several other former world champions of chess and poker. They played, as Hassabis later put it, "long into the night." I don't know who won. He hasn't said, and the omission may be the point.
"For me, that was the perfect way to celebrate," he told an interviewer from the Nobel Foundation that December in Stockholm. "It couldn't have been more apt."
It was apt. A man who had spent his adult life persuading machines to play games had just been honored by the oldest prize in science for teaching a machine to predict how proteins fold, and he marked the occasion by sitting down at a board with people who could still beat him. In the same interview he admitted something that sounds like a confession: "I'm not that good at celebrating successes... Because I'm always looking straight away as soon as I do something onto the next thing." So he celebrated the only way that didn't require stopping. He played the next game.
What follows tries to explain how a boy who learned chess at four by watching his father play his uncle ended up there: a CBE, then a knighthood, the Lasker, the Breakthrough Prize, the Gairdner, then the Nobel, and in 2025 one of the "Architects of AI" on Time's Person of the Year cover. It also tries to explain why, through all of it, the board never left the table. The short version is that Hassabis treated games as a curriculum and not a hobby. Chess taught him to think, game design taught him to build, and games became the place where his machines learned. The longer version runs through a butcher's surname, a failed simulation of a fictional country, amnesiac patients who could not picture a beach, a hospital data scandal, a negotiation over autonomy with a parent company that went on for years and that DeepMind lost, and a document in Stockholm. In none of it does he ever quite stop moving.
By the Numbers
The Hassabis Ledger
2300Elo rating at age 13, his chess peak (January 1990)
£400MReported price Google paid for DeepMind in January 2014
4–1AlphaGo's score against Lee Sedol in Seoul, March 2016
200M+Protein structures released through the AlphaFold database
2M+Researchers in 190 countries who have used AlphaFold
£1.33BDeepMind Technologies revenue, 2024
$600MIsomorphic Labs' first external round, led by Thrive Capital (2025)
One Letter Changed
The family name was Hassapis, Χασάπης in Greek. It means butcher.
Somewhere along the way the son changed it. Ingo Althöfer, the German mathematician and computer-games researcher who wrote a long appreciation of Hassabis for the ICGA Journal in 2024, described the change in the vocabulary of the field that would make Hassabis famous:
— Ingo Althöfer, ICGA Journal (2024)He executed a point mutation by changing 'p' to 'b'.
A point mutation is a single substituted base in a strand of DNA. It is the smallest change a genome can undergo, and sometimes the change matters a great deal. One of Hassabis's younger brothers still carries the original name. Between the two brothers, then, sits a family record of what one letter can do: one sibling kept the butcher's name, and the other took a variant that now appears on a list of Nobel laureates in chemistry next to Frederick Sanger and Dorothy Hodgkin.
The parents were not built for conventional careers either. Costas Hassapis was a Greek Cypriot who grew up an only child and was the first in his family to attend university. According to Sebastian Mallaby, whose biography The Infinity Machine appeared in 2026, he "was too much of a bohemian free spirit to abide office work." He wanted to be a singer-songwriter and earned his living selling toys out of the back of a van. Angela, Demis's mother, was Chinese Singaporean. She grew up poor and was orphaned, then adopted by a relative. She trained as a nurse but worked as a retail clerk and, part-time, as a cleaner.
Put those two sketches next to each other and you can see the household that produced him. The father had an education he chose not to cash in, and the mother had cashed in everything she had and gotten less than she deserved for it. One supplied the restlessness, a man who found offices unbearable and sold toys from a van, and the other supplied the knowledge, earned the hard way, that nobody hands you anything. The child they raised in North London had a Greek name, a Singaporean mother, and a Cypriot father, grew up in an English city, and would later choose to stay there when everyone with money told him to leave. Decades on he told a Singaporean newspaper that his parents had told him to "chase your passions, go deep with them." It is an unusual thing for a bohemian toy-seller and an orphaned nurse to say to a son, and the son did exactly what they said.
He still describes himself as proud of the Greek side. He understands some of the language, visits Greece regularly, and has, in his words, "a lot of contact with the Greek side of things." In September 2025 he appeared onstage with Prime Minister Kyriakos Mitsotakis at an Endeavor Greece event. The Greek Cypriot press had claimed him well before the Nobel; Parikiaki, the London Greek-Cypriot paper, ran "Demis Hassabis the child prodigy" in April 2014, three months after the Google sale. When he won in Stockholm, the Cyprus Mail headline read "Nobel prize win for Greek Cypriot scientist." The Straits Times ran "Nobel Chemistry laureate Demis Hassabis has Singapore in his blood." Each place that could claim him did. He is a Londoner whom other places have quietly adopted, which is roughly what happened to his mother, in reverse.
The Inanimate Lump of Plastic
He learned chess at four. He took it seriously from the start. At thirteen he reached master standard with an Elo rating of 2300, a peak recorded on the FIDE list in January 1990, and he captained England's junior teams. A rating of 2300 at thirteen marks a strong player, though not a prodigy of the very first rank. In practice it meant a childhood spent at tournaments, in training camps, and across boards from adults.
The camps are where the rest of his life starts. In Stockholm he described them this way:
"On some of the chess training camps you would use chess computers, physical boards, to try and train and get better at. As we were playing against these chess computers, I remember being fascinated by the fact that someone had managed to program this inanimate lump of plastic to actually play chess. So although I was meant to be training my chess openings and things, I actually got more fascinated by the program and the AI underneath it."
The image of a chess prodigy who stopped looking at the board and started looking at the box is the closest thing this story has to an origin. The other boys wanted to know which move was best. He wanted to know how the plastic could know.
In 1984 he bought his first computer, a ZX Spectrum 48K, using his chess winnings. That detail recurs in every profile because it makes the logic plain. The game paid for the machine, and the machine later paid for everything else. He taught himself to program from books. On a Commodore Amiga he wrote his first AI program, which played Othello (the game also called reversi). A boy who could beat most adults at one board game built a machine to play another.
School was secondary. He was home-schooled by his parents for a year and then attended Christ's College, a comprehensive school in East Finchley, where he finished his A-levels two years early, at sixteen. Asked in Stockholm which teacher had mattered most, he chose his A-level maths teacher, a Mr. Lovegrove, who "was teaching us how to think and how to break down problems in a really disciplined, efficient way." In the same breath he attached that to his chess upbringing. Being trained to decompose problems by a maths teacher and being trained to evaluate positions by chess come to the same skill once you remove the subject matter, and he seems to have noticed this early.
He kept playing long after he stopped needing to. He represented Cambridge in the Oxford–Cambridge varsity chess matches of 1995, 1996, and 1997 and earned a half blue. He won the Pentamind, the all-round championship at the London Mind Sports Olympiad, five times: in 1998, 1999, 2000, 2001, and 2003. He won the Decamentathlon in 2003 and 2004. He was part of the team that won the Diplomacy world team championship in 2004 and finished fourth individually in 2006. He has cashed six times at the World Series of Poker, once in the Main Event. His FIDE rating in March 2019 was 2220, which means that at forty-two, while running one of the most important research labs in the world, he was still playing rated chess. These days he plays online late at night, "a little bit like going to the gym for the mind." He used to play League of Legends on a team with his children, his brother, and a family friend.
Most people who become serious about something drop their childhood games. Hassabis has kept all of his.
Twenty Years Too Early
Cambridge accepted him and then asked him to wait a year because he was too young. He spent the gap year on a contest in Amiga Power, the August 1992 issue of which ran a "Win-a-job-at-Bullfrog" competition. He won. Bullfrog Productions in Guildford was then one of the most inventive studios in British games, and he started out as a playtester on Syndicate. At seventeen he was co-designer and lead programmer on Theme Park, released in 1994 and built with Peter Molyneux.
Molyneux needs a sketch of his own. He was Bullfrog's designer, the man who would later found Lionhead Studios and make Black & White, a "god game" in which the player oversees a world of autonomous villagers. His whole career was organized around one fantasy: giving the player a world full of little minds and watching what they do. Over the years he became known as much for what he promised as for what he shipped. He was the first adult professional to treat the teenage Hassabis as a collaborator. Their sensibilities matched well. Molyneux wanted to play God, and Hassabis wanted to understand what God would need to build.
Theme Park sold several million copies and helped create a genre of simulation sandbox games. Hassabis has said the commercial success mattered less to him than what it showed:
"That of course has AI as the heart of the game. That's why everyone liked Theme Park, because anytime you played, anytime you designed your own theme park, all the little people would come in and play on your rides and everything. And every time, the game would be different because the AI would react to how you as the player designed the theme park. That was one of the things that after I wrote that and I saw how many people enjoyed playing it and interacting with the AI, that convinced me that AI was what I wanted to spend my whole career on."
He was then offered a seven-figure sum to stay in games, and he said no. A seventeen-year-old turned down a fortune to go read computer science. He had earned enough in the gap year to pay his own way through university, and he took the Computer Science Tripos at Queens' College, graduating in 1997 with a double first. It was at Cambridge, in the nineties, that he first came across the protein folding problem. "Kind of like a puzzle really," he called it. He put it away for about twenty years.
Lionhead came next. Molyneux had just founded it, and Hassabis became lead AI programmer on Black & White. He stayed only a short time. In 1998 he left to found his own company, Elixir Studios, in London, and signed publishing deals with Eidos Interactive, Vivendi Universal, and Microsoft. He was twenty-two.
Elixir's first game, Republic: The Revolution, is the most revealing failure in his career, because the ambition and the error in it are the same thing. It was a political simulation, and the premise was an AI model of the workings of an entire fictional country, in which the player led a revolution against the regime. The scope was so large that the release slipped. When the game finally shipped in 2003, it was a reduced version of the original design. Reviews were lukewarm and it scored 62 on Metacritic. Evil Genius, Elixir's second game, was a tongue-in-cheek Austin Powers parody about running a lair. It did better, scoring 75. Both games earned BAFTA nominations for James Hannigan's scores. Neither saved the company. In April 2005 Elixir's intellectual property and technology rights were sold to various publishers and the studio closed.
It is tempting to read Elixir as the necessary failure that gets narrated in every founder story. Hassabis tells it with more precision than that, and with less consolation:
— Demis Hassabis, Nobel Prize interview, December 2024Some of the ambitions that we had for the game design and the technologies were a little bit too ahead of where the technology was. I've described it as being 20 years ahead of your time. Actually, what I realised was that timing is very important.
Twenty years ahead of its time is an exact figure, and the arithmetic supports it. In 2003, a simulated country full of autonomous agents was beyond what a PC could run. By the mid-2020s, DeepMind was building SIMA, an agent that follows natural-language instructions across nine commercial video games, and Genie, a model that produces playable 3D worlds from a text prompt or a sketch. A lab Hassabis ran had built the simulation engine that Republic needed about twenty years earlier. The idea was right, but the hardware arrived two decades late.
The lesson he drew is the most practically useful sentence he has ever said: "You need to pick hard problems to solve and be ambitious with that, but you've also got to pick the right time." His later career can be read as an extended application of that lesson. Every hard problem he picked afterward had a test he could run and an approaching moment when the computing power would exist to run it.
The Patients Who Could Not Imagine a Beach
After Elixir, Hassabis did something founders rarely do after a failure. He went back to school and studied how minds work.
He enrolled in a PhD in cognitive neuroscience at the UCL Queen Square Institute of Neurology under Eleanor Maguire. His reasoning was strategic. If the aim was to build intelligence, the only working example anyone had was the human brain, and it seemed sensible to understand how it was put together before trying to make another one. The thesis, finished in 2009, was titled Neural Processes Underpinning Episodic Memory.
His first academic paper, published in PNAS in 2007 with Dharshan Kumaran, Seralynne Vann, and Maguire, carried the blunt title "Patients with hippocampal amnesia cannot imagine new experiences." The hippocampus was long known to be necessary for remembering. The paper showed systematically, for the first time, that patients with hippocampal damage also could not imagine. Asked to picture themselves in a new scene, they produced fragments, pieces without a coherent spatial setting. The New York Times headline was "Amnesiacs May Be Cut Off From Past and Future Alike." Science named the work one of its top ten breakthroughs of 2007.
The finding linked two things that look different: remembering the past by reconstructing it, and imagining the future by constructing it. Hassabis and Maguire proposed that a single process underlay both, which they called scene construction, meaning the generation and online maintenance of a complex, coherent scene. Later he generalized the idea into what he called a "simulation engine of the mind," a system that imagines events and scenarios in order to plan.
The connection between this and his games career should be stated plainly. In Theme Park he had written the AI that let small simulated people react to a world. In Republic he had tried to simulate a whole country. In his first scientific paper he found that the human brain runs a simulation engine, and that if you damage the hardware it runs on, a person loses both the past and the future together. The man who would build AlphaGo, whose search algorithm plays out imagined futures before choosing a move, began his scientific career by showing that imagining futures is a core function of a mind. Looked at from 2026, the progression seems deliberate, as if each stage of his career prototyped the next one.
After the PhD he worked as a visiting scientist at MIT in Tomaso Poggio's lab and at Harvard. He then won a Henry Wellcome postdoctoral fellowship to the Gatsby Computational Neuroscience Unit at UCL, working with Peter Dayan. In 2009, at Gatsby, he met a New Zealander who had been thinking about the same thing for even longer.
Shane Legg was born in 1973 or 1974 and attended Rotorua Lakes High School on New Zealand's North Island. He finished his undergraduate degree at Waikato in 1996, and that same year completed a master's thesis at Auckland titled Solomonoff Induction, a study of the mathematically ideal (and uncomputable) way to predict anything from data. In the early 2000s he and Ben Goertzel revived and popularized the term "artificial general intelligence," at a time when, by MIT Technology Review's later account, talking about AGI "would put you on the lunatic fringe." He worked at Goertzel's startup Webmind, then did a PhD at the Dalle Molle Institute in Lugano under Marcus Hutter on theoretical models of superintelligent machines. His 2008 dissertation was called, unhedged, Machine Super Intelligence, and it won a $10,000 prize from the Singularity Institute. Then he did a postdoc in finance, which is a strange detour for a man whose thesis title sounded like the name of a bomb.
Hassabis came to the problem of general intelligence through games and the brain. Legg came to it through mathematics and a serious worry. By 2011 he was telling an interviewer on LessWrong that he took existential risk from AI seriously. The two of them made an odd but complementary pair: a builder who wanted to know how imagination works, and a theorist who had already calculated where the work might end up.
Solve Intelligence, Then Use It to Solve Everything Else
DeepMind Technologies Limited was incorporated at Companies House on September 23, 2010, and launched on November 15. It had three founders. Hassabis supplied the vision and the ability to recruit. Legg supplied the theory and, later, much of the safety program. The third founder was Mustafa Suleyman, a friend of Hassabis through family, who handled what neither scientist wanted to handle: applications, policy, and how the technology would meet the institutions of the world. Suleyman was the founder who looked outward, and that would make him the most exposed of the three.
The mission statement was famous almost immediately: solve intelligence, then use intelligence "to solve everything else." It sounds like hubris, and it is. It is also a precise claim about leverage. If intelligence is the general-purpose tool behind every other advance, then building it counts as working on all problems at once. Hassabis would later make the same point in Stockholm through an image of himself on a plane: "Sometimes when I'm flying over to the US for a business trip or something on a 747, I sometimes look out the window and think, 'how have we as humanity manage this with our sort of primate brains?'" To him intelligence is the force that built the 747, and building more of it would build more 747s. The logic has no ceiling, and he has never presented it as having one.
He recruited David Silver, a friend from Cambridge and a former partner at Elixir. Silver's own career had followed the same path as Hassabis's. He went from games, to a doctorate in reinforcement learning, to the conviction that systems which learn from experience would eventually beat systems taught by rules. He would become the main architect of AlphaGo and first author on the paper describing AlphaGo Zero. In other words, the man who would build the Go machine had also helped ship a game whose AI promised more than the hardware could deliver.
Money came from people with large ideas about the future. Horizons Ventures and Founders Fund invested, as did Scott Banister, Peter Thiel, and Elon Musk. Jaan Tallinn, the Estonian engineer behind Skype who had become one of the first serious private funders of AI-risk research, invested early and advised the company. The cap table reads like a list of people who believed AGI was coming and were afraid of it in varying degrees.
There was one condition Hassabis did not concede. A later Business Insider headline summed it up: "How DeepMind convinced billionaire Peter Thiel to invest without moving the company to Silicon Valley." DeepMind stayed in London. That was not sentiment, or not only sentiment. London had UCL, Gatsby, Cambridge an hour north, Oxford an hour west, and a pool of academic talent that no other AI lab was competing for. It was also where he had grown up and where the Elixir studio had been. The son of a man who couldn't stand offices insisted on keeping his own company in his own city.
The first public result was modest and technically stunning. In December 2013, Volodymyr Mnih, Koray Kavukcuoglu, Silver, and colleagues published "Playing Atari with Deep Reinforcement Learning." A single system, the Deep Q-Network, combined a convolutional neural network with Q-learning and learned to play Atari 2600 games from the 1970s and 1980s, including Pong, Breakout, and Enduro. It surpassed human performance on some of them using only the raw pixels on the screen and the score. Nobody gave it the rules. The same code, unchanged, played different games. A system that learns from experience flails first. Early on it barely plays, then it finds what works, and eventually it plays some games better than any person. For a former games designer, the setup was elegant: a game is a world with a score, the score supplies the reward, and the reward is enough to learn from.
The research came out in December 2013. Within weeks several large companies were bidding. Facebook courted DeepMind, and so did Musk and others. On January 26, 2014, Google confirmed it would buy the company. The price was reported variously as between $400 million and $650 million, and the British press settled on £400 million. The Telegraph called him "the secretive computer boffin with the £400 million brain." The Evening Standard ran an exclusive interview headlined "London's megamind." That October the same paper ranked him the third most influential Londoner.
He sold because he had to. DeepMind needed a backer with deep pockets to pay researchers, buy compute, and expand. The company's later history shows the price of that bargain was much higher than £400 million and was paid in autonomy, a little at a time. The Atari paper reached the cover of Nature on February 26, 2015, with Kavukcuoglu, a Turkish-born researcher who co-authored that first paper, among its names. Kavukcuoglu stayed for the next eleven years. In August 2026, when Hassabis stepped back from the CEO role, Kavukcuoglu was one of the senior vice presidents running the lab.
Seoul, and the Machine That No Longer Needed Us
Go was the obvious next target. It has many more board positions than chess, and the brute-force search that let IBM's Deep Blue win at chess in 1997 breaks down against it. Before DeepMind, programs played Go at amateur strength.
In October 2015, behind closed doors, AlphaGo beat Fan Hui, the European champion and a 2-dan professional, five games to none. It was the first time an AI had beaten a professional Go player. The result was published in Nature in January 2016 and made the cover. The system used two deep neural networks: a policy network that proposed promising moves and a value network that judged positions. These were tied together with a Monte Carlo tree search that used the networks to imagine possible continuations before choosing a move. In other words it was a simulation engine. It had first been trained on about thirty million moves from human games and then improved by playing against itself.
In March 2016, in Seoul, it played Lee Sedol, a 9-dan professional and former world champion, in a five-game match for a $1 million prize. AlphaGo won four games to one. Greg Kohs filmed the match for the documentary AlphaGo, which won a Grand Prix at Cannes Lions that June. The Korea Baduk Association made the program an honorary 9-dan professional. In May 2017, at the Future of Go Summit in China, AlphaGo beat Ke Jie, then the world's top-ranked player, three games to none, and the Chinese Weiqi Association also gave it 9-dan rank. Nature named Hassabis one of its ten people of 2016. In April 2017 Time listed him in its 100 most influential people, and the short tribute was written by Ray Kurzweil, a futurist describing the man most likely to build what Kurzweil had predicted.
The more consequential result came that October and got less attention. AlphaGo Zero was trained with no human games at all. It began with random play and learned only by playing itself. It used one neural network instead of two, and four of Google's TPU chips where the original had needed forty-eight. A team of about fifteen people built it. After three days of training it beat the version that had defeated Lee Sedol, and it eventually won a hundred games out of a hundred. A BBC headline described AI as becoming "more alien." In December, AlphaZero, a generalization of the approach, reached superhuman play in chess and shogi as well. In 2019 MuZero learned Go, chess, shogi, and Atari without even being given the rules.
The sequence follows the pattern Hassabis had set at the chess camps. The first version learned from people, the second learned from itself, and the third learned without being told what the game was. Human knowledge went from foundation, to support, to something that could be discarded. Hassabis has never said it this bluntly, but the direction was obvious: the more completely the machine dropped what humans could teach it, the stronger it played.
Games stayed the test bed. In 2016 Hassabis named StarCraft, a real-time strategy game with hidden information, as the next challenge. In January 2019 AlphaStar beat professionals ten matches in a row. The win came with an asterisk the company acknowledged: the agent could see the entire map at once, which no human can, and a version without that advantage lost the next match. By October 2019, with that advantage removed and playing all three races on the public European ladder, AlphaStar reached Grandmaster. In 2020 Agent57 exceeded the human benchmark on all fifty-seven Atari games. The work that started with Pong took seven years to finish the full set.
The games also produced real-world results. In 2016 a Google data-center engineer named Jim Gao, who had already been applying machine learning to energy efficiency, saw AlphaGo and asked DeepMind to help build a reinforcement-learning system that could recommend actions to the cooling plants. Operators found the recommendations counterintuitive, but they met every safety constraint, and Google reported cutting the energy used for cooling by 40 percent. A later, more autonomous version came up with strategies that surprised long-time operators, such as using winter conditions to produce colder-than-normal water. It was the first strong evidence that the game-playing methods could save real money.
Then the attention faded. A LexisNexis count found 1,842 news stories mentioning DeepMind in 2016 and 1,363 in 2019. In June 2020 CNBC ran "Why the buzz around DeepMind is dissipating as it transitions from games to science." The headline was written six months before the result that would earn Hassabis his Nobel.
The Cost of Being Close
The Google acquisition came with a safeguard that was supposed to reassure people and instead became a puzzle. As part of the deal, DeepMind set up an AI ethics board. Neither Google nor DeepMind would say who sat on it. Forbes ran "Inside Google's Mysterious Ethics Board" eight days after the sale. A body meant to build trust in powerful technology had members nobody could name.
Hospitals were where that kind of secrecy did the most damage. In September 2015, seeking a socially useful project, DeepMind signed an information-sharing agreement with the Royal Free London NHS Foundation Trust to co-develop Streams, a clinical app that alerted staff to patients at risk of acute kidney injury. The NHS still ran on paper and fax machines, and the work was needed. Staff at the Royal Free later said the app saved a "huge amount of time" and made a "phenomenal" difference.
But in April 2016 New Scientist obtained the agreement itself. It covered the three hospitals run by the trust, which treated an estimated 1.6 million patients a year, and it gave DeepMind access to admissions, discharges, transfers, accident and emergency, pathology, radiology, and critical-care records. That included whether patients had been diagnosed with HIV, had suffered from depression, or had ever had an abortion. The Guardian ran an op-ed headlined "Google's DeepMind shouldn't suck up our NHS records in secret." New Scientist followed up by reporting that the project had not obtained approval from a key confidentiality advisory group.
The verdicts came in 2017. In May, Sky News published a leaked letter from Dame Fiona Caldicott, the National Data Guardian, which said that in her "considered opinion" the data had been transferred on an "inappropriate legal basis." In July, after a year-long investigation, the Information Commissioner's Office found that the Royal Free had failed to comply with the Data Protection Act. Suleyman and Dominic King, DeepMind's clinical lead, wrote the company's response in four words that are now quoted as often as anything else DeepMind published: "we need to do better."
That October DeepMind set up DeepMind Ethics & Society, a separate unit to fund outside research on privacy, fairness, economic impact, and governance. It was distinct from the anonymous board created in 2014. A year later, on November 13, 2018, DeepMind announced that its health division and Streams would be absorbed into Google Health. Privacy advocates said this broke an earlier assurance that patient data would not be connected to Google. A DeepMind spokesman said the data would stay separate. The Guardian headline was "Google 'betrays patient trust' with DeepMind Health move."
Here the logic of the 2014 sale becomes visible. DeepMind had sold itself to buy computing power and keep its research independence. The independence turned out to be partial, and the first part of the lab to be absorbed was the part that dealt with the public. In December 2019 Suleyman left for a policy job at Google. Of the three founders, the one whose job had been to explain DeepMind to the world was the first to leave it.
None of this was entirely Hassabis's doing, and none of it was entirely separate from him. A chess player is used to an opponent who sits across the board, makes moves within known rules, and accepts the result. Patients in a public health system are not opponents, and the rules governing their records are written by the institutions involved, not by the players. The Royal Free affair taught DeepMind that the trust of the people whose data it uses is part of its working infrastructure and can be lost, and that one leaked contract could do more damage than any rival lab.
There was a longer contest too, quieter and slower, between DeepMind and its owner. In 2019 The Economist titled a report "DeepMind and Google: the battle to control artificial intelligence." DeepMind's executives spent years trying to secure more autonomy, including structures that would have placed the lab's most powerful work under independent oversight. In May 2023, Parmy Olson of the Wall Street Journal published the ending in her headline: "Google Unit DeepMind Tried—and Failed—to Win AI Autonomy From Parent." The game had been lost a month earlier, by a side DeepMind had not been watching.
The Lighthouse
He had carried the protein problem since his undergraduate years. In Stockholm he said why it mattered to him: "I've always believed it's quite hard to describe that system just with a few mathematical equations, so it's not like physics. Actually, I think AI systems and AI algorithms are quite a good match, a good description language, for biology." For more than twenty-five years he had talked about biology as an information-processing system with Sir Paul Nurse, the Nobel-winning cell biologist who heads the Francis Crick Institute (Hassabis joined its scientific advisory board in 2016). For a man trained to see problems as puzzles, protein folding was nearly ideal. It had a clear objective, a long scientific history, an external benchmark, and a payoff that included drug discovery and a basic understanding of disease.
DeepMind took up the problem in 2016, the year of Seoul. The external benchmark was CASP, the Critical Assessment of Techniques for Protein Structure Prediction, a competition held every two years in which teams predict structures that have been determined experimentally but not yet published. In December 2018, at CASP13, the first AlphaFold produced the most accurate prediction for 25 of the 43 proteins in its category. Hassabis's description to The Guardian was more revealing than he probably intended:
— Demis Hassabis to The Guardian, December 2018This is a lighthouse project, our first major investment in terms of people and resources into a fundamental, very important, real-world scientific problem.
A lighthouse is placed so ships know where the coast is. The phrase admits that the earlier projects, Go included, were preparation for a point at which the methods would be turned onto the real world. Games had been practice, and from here on the problems were real.
AlphaFold 2 was the result. In November 2020, at CASP14, it achieved a median global distance test score of 87.0 on the hardest category of targets, up from a median below 60 two years earlier. Its typical error was under one angstrom, smaller than the width of an atom, and its accuracy was competitive with laboratory methods. The organizers said the problem was essentially solved. Andriy Kryshtafovych, one of the adjudicators, called the result "truly remarkable" and said protein structure prediction had been "largely solved." Nature's headline quoted a researcher: "'It will change everything.'" In 2021 Science named it Breakthrough of the Year. That made four appearances for Hassabis's work on the journal's annual list, starting in 2007 with the amnesia paper about imagining the future.
What he did next with the result is what distinguishes him. In July 2021 DeepMind released AlphaFold 2's code. A week later it published predicted structures for nearly every human protein and for the full proteomes of twenty other organisms studied closely by scientists. In July 2022, through a database built with EMBL's European Bioinformatics Institute, it released predictions for more than 200 million proteins, which is essentially every protein known to science, free to anyone. By Nobel week the database had been used by more than two million researchers in 190 countries. Hassabis said the reason was arithmetic: "We knew that we could only do a tiny, tiny fraction of the downstream work that AlphaFold could enable." He was proudest of the collaborations nobody would have paid for, such as work with the Drugs for Neglected Diseases initiative on leishmaniasis and dengue fever, diseases of poor countries that large pharmaceutical companies tend to ignore. "It's been really pleasing for us to just be able to give those nonprofit institutes the protein structures, so they can get on straight away with designing drugs."
The 2023 prizes followed quickly: the Lasker, the Breakthrough Prize in Life Sciences, the Canada Gairdner, and, oddly for a lapsed Greek Orthodox Londoner, membership in the Pontifical Academy of Sciences. Then came October 9, 2024. Half the Nobel Prize in Chemistry went to David Baker of the University of Washington for computational protein design, and the other half to Hassabis and Jumper for structure prediction. Geoffrey Hinton shared the physics prize the same week. AI had won two Nobels in one week.
John Jumper deserves a sketch, though the record on him in these sources is thin. He led the AlphaFold team and became, alongside Hassabis, the scientist the Nobel committee named. The pairing matters as much as the person. A lab founded to solve intelligence was honored for solving a narrower, older problem in chemistry, and the honor named the man who ran the lab together with the man who ran the project. DeepMind had always worked that way, with a long-term vision at the top and a team assembled to deliver one specific result.
The AlphaFold record has caveats, and the company has not always welcomed them. AlphaFold 3, released in May 2024 with Isomorphic Labs, extended prediction to how proteins interact with DNA, RNA, and drug-like small molecules. But preliminary results from CASP16 that November showed AlphaFold 3-based models did not clearly beat older methods on protein-ligand prediction. The best entries in the pharmaceutical-target category combined AlphaFold 2 predictions with human visual inspection and manual adjustment. Elsewhere in DeepMind's growing catalog of scientific claims, outside reviewers kept pushing back. Virginia Vassilevska Williams called AlphaTensor's matrix-multiplication result "a little overhyped." Anthony Cheetham and Ram Seshadri reviewed the materials-discovery tool GNoME and could not find any "strikingly novel" materials among its predictions. Communications of the ACM pointed out that DeepMind had never published the comparative benchmarks experts had asked for to support AlphaChip's claims. A lighthouse is visible from a long way off, and so is anything placed near it that doesn't hold up.
In February 2021, months before the AlphaFold code was released, Hassabis incorporated a second company to capture the commercial value the free database would never earn. Isomorphic Labs, a separate Alphabet subsidiary with Hassabis as CEO, was announced on November 4, 2021, with a mission to turn structural prediction into drug design. It opened a second office in Lausanne in December 2022. In January 2024 it signed partnerships with Novartis and Eli Lilly. In April 2025 it raised $600 million in its first external round, led by Thrive Capital. In February 2026 it announced a Drug Design Engine that it said doubled AlphaFold 3's performance on a generalization benchmark. The arrangement is neat: the open database serves the public good and sets the standard, and the company sells what comes next. He gave away the map of every protein and kept the business of designing drugs against them.
The Year the Board Changed Shape
ChatGPT launched in late November 2022, and the competitive landscape changed overnight. DeepMind had spent twelve years preparing for one kind of contest, deliberate and scientific and built around benchmarks, and found itself in a race decided by consumer chat products. The transformer architecture behind the new systems had been invented in 2017 at Google itself, by Google Brain, the parent company's other AI lab. DeepMind had built its own language models, including Chinchilla, Sparrow, Flamingo, and the 604-task generalist Gato. None of them had been released to the public in the way ChatGPT had.
In April 2023 Alphabet merged Brain into DeepMind to form Google DeepMind and put Hassabis in charge of the combined organization. On paper he won: he got the bigger organization, the larger budget, the merged headcount, and responsibility for Google's flagship model. But the merger also ended the long effort for autonomy, because the lab that had wanted independent oversight now sat at the center of Google's product strategy. Gemini, the multimodal model meant to compete with GPT-4, launched on December 6, 2023. It was followed by Gemini 2.0 in December 2024, Gemini 2.5 with reasoning in March 2025, and Gemini 3 Pro on November 18, 2025, built into Google Search the day it shipped. DeepMind's accounts for 2024 show revenue of £1.33 billion and net income of £174 million. The company once mocked as a research boutique for running Atari experiments was now responsible for the products Google most needed to work.
Through all of this Hassabis kept a position that was easy to caricature and difficult to hold. In May 2023 he signed the one-sentence statement that "mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war." Legg signed it too. Hassabis also said a pause would be very hard to enforce worldwide and that the potential benefits for health and climate justified continuing. That October he told The Guardian that AI risk had to be treated as seriously as the climate crisis. In August 2025 he told the same paper that AI would be "10 times bigger than the Industrial Revolution – and maybe 10 times faster." He was warning about the danger and accelerating it in the same breath, and he does not see a contradiction, because he was trained to calculate many moves ahead and assumes the people across the board will do the same.
His stated concerns are specific and limited to two. "One is, these are general purpose systems... it can also be repurposed by bad actors for harmful ends." The other: "As the systems become more and more powerful themselves, more autonomous, more agent like, can we understand and keep control of those systems?" Being able to understand and keep control of a system he built is the same thing he was denied as a subsidiary of Google, which may explain why the second concern is so clearly put.
Rivals came after his staff. In August 2025 the Wall Street Journal reported that Microsoft was recruiting from DeepMind with the promise of "less bureaucracy." The Microsoft AI division making that pitch was led by Mustafa Suleyman, the co-founder who had left DeepMind for Google in 2019 and then left Google. It is notable when your former partner's sales pitch is that his place is less bureaucratic than yours. DeepMind then employed about six thousand people. The small London startup that had once persuaded Peter Thiel to fund it without moving was now one of the bureaucracies.
On August 5, 2026, Alphabet announced a reorganization. Hassabis would step down as CEO of Google DeepMind, become its chairman, and take the title of chief scientist of Alphabet. Jeff Dean, the Google engineer who had co-led the old Brain organization and who for twenty years was the closest thing Google had to a resident genius, was leaving. Kavukcuoglu, from the 2013 Atari paper, would help run the lab. That summer DeepMind moved into Platform 37, Google's building in King's Cross Central, a few miles from the North London streets where Hassabis grew up and where he still lives with his family. The company had never left London, though it had stopped being entirely his.
The Business Insider report on the change led with the phrase "steps down." It reads as accurate, though the step was also a step up, toward the thing he had said in Stockholm he was still trying to do: "My original ambition in AI is to build what we sometimes call artificial general intelligence... I guess that was my original goal when I got into AI, 30 plus years ago and that's still my goal now." He kept the scientific question and handed the management of the lab to someone else. His father never liked offices either.
Restoring the Missing Letters
One DeepMind project has nothing to do with Gemini, medicine, or money, and it fits the founder's life story better than anything else the company has built. In March 2022 the lab published Ithaca, named after the island Odysseus spends the Odyssey trying to reach. It is a neural network that restores missing text in damaged ancient Greek inscriptions and estimates where and when they were made. On its own it restores the missing characters with 62 percent accuracy, attributes the location correctly 71 percent of the time, and dates inscriptions to within thirty years. The authors reported that historians working with Ithaca reached 72 percent accuracy, compared with 25 percent alone. The classicist Eleanor Dickey pointed out that the test subjects were students, and that it remained unclear how much the tool would help "genuinely qualified editors." The pattern held as usual: a remarkable result with a reasonable caveat attached.
Ithaca's job is to examine broken Greek words and supply the letters that are missing or wrong. It works one character at a time, much like the point mutation that turned Hassapis into Hassabis.
He is fifty now. He is a Knight Bachelor, a Fellow of the Royal Society, and a member of the US National Academy of Engineering. In 2026 he received the Royal Society of Arts' Albert Medal and used the occasion to call for "a new Renaissance." He is the subject of The Thinking Game, which premiered at Tribeca in 2024 and was made by Greg Kohs, the filmmaker who shot AlphaGo in Seoul. He is also the subject of Mallaby's The Infinity Machine, a title chosen by someone who has watched him closely and concluded that he does not stop. In 2025 he entered the Ukie Hall of Fame, the British games industry's honor roll, which is a fitting recognition for the seventeen-year-old who programmed Theme Park's little people and decided that AI would be his life's work.
He still plays chess late at night, online, against strangers.
In 2016, after Seoul, the Korea Baduk Association made AlphaGo an honorary nine-dan professional, the highest rank in the game. In 2026, ten years after the Lee Sedol match, the same association gave Demis Hassabis an amateur rank of seventh dan.
The program holds the higher rank, two levels above the man who built it.
Part IIThe Playbook
The principles below are drawn from three decades of decisions, not from Hassabis's speeches, though he is quoted where his own words make the point most precisely. The common thread is his insistence on treating hard problems as games: environments with clear rules, measurable outcomes, and a moment of readiness that matters as much as the moves. Some principles below come from things he did well. Others come from the parts of the record he would probably like to replay.
The principles
- 1.Pick the hard problem, then pick the hour.
- 2.Find the scoreboard before you find the solution.
- 3.Study the only working example.
- 4.Write a mission that compounds.
- 5.Stay where your unfair advantage lives.
- 6.Sell to buy compute, and price the autonomy you lose.
- 7.Name one lighthouse.
- 8.Remove the human crutch on schedule.
- 9.Give away what you cannot fully exploit, and sell the layer above it.
- 10.Treat trust as load-bearing infrastructure.
- 11.Run failure as a binary search.
- 12.Recruit at the seams between fields.
Principle 1
Pick the hard problem, then pick the hour
Elixir Studios was ambitious and had the right idea. Republic: The Revolution tried to simulate an entire country with agents twenty years before the hardware could support it, and the company closed in 2005. Hassabis's own diagnosis is the most important strategic insight in his career: ambition without timing gets you a delay, a scaled-back product, and a Metacritic score of 62. The same instinct that made Republic fail made AlphaGo succeed, because in 2015 deep learning, reinforcement learning, and TPUs had all matured at roughly the same moment.
Most founders treat timing as luck and hindsight. Hassabis treats it as a variable to estimate. His question is not only whether a problem matters but whether the conditions now exist (compute, data, methods, a benchmark) for real progress on it. He learned about protein folding as a Cambridge undergraduate in the nineties and did nothing with it until 2016, after deep learning had shown it could handle complex structure. He carried the puzzle for twenty years and waited.
Tactic: For every ambitious project, write down the two or three external conditions it depends on, such as hardware, data, regulation, or market behavior, and don't commit fully until at least one has visibly changed in the past eighteen months.
Principle 2
Find the scoreboard before you find the solution
Every DeepMind breakthrough happened in a setting with an unambiguous score. Atari games have points. Go has a winner. StarCraft has a public ladder. Protein folding had CASP, an independent biennial competition that judged predictions against structures determined experimentally. Even the data-center work had power usage effectiveness, a single number that either improves or doesn't.
This is a game designer's habit applied to research. A game is a world with rules and a reward signal, and a reward signal is what lets a learning system, or a learning organization, improve without arguing about whether it has. AlphaFold's credibility came less from DeepMind's own claims than from CASP's adjudicators saying the problem was "largely solved." The projects that drew the strongest outside skepticism, such as GNoME and AlphaChip, were the ones where the scoreboard was disputed or comparative benchmarks never appeared.
Tactic: Before staffing a project, identify the external, adversarial benchmark that would convince a skeptic it worked; if none exists, make building that benchmark the first milestone.
Principle 3
Study the only working example
After his company failed, Hassabis spent years earning a PhD in cognitive neuroscience. Most people would call that a detour. It was due diligence on the only general intelligence available to study. His hippocampal amnesia research showed that remembering and imagining depend on a shared process of scene construction, and he extended that into the idea of a "simulation engine of the mind" used for planning. AlphaGo's design, with learned intuition proposing moves and tree search imagining where they lead, has the same structure.
The principle is broader than AI. When you want to build something that has never existed, the closest working analogue, even in a different field, usually contains clues the specialists in your field have missed. DeepMind made this a structural commitment, mixing neuroscientists with engineers from the start, and its 2018 Nature paper on grid-like cells emerging in navigating agents brought the insight back into neuroscience.
Tactic: Identify the one existing system in nature, history, or another industry that already does what you're trying to build, and spend real time studying it, not a reading list but a stretch of genuine apprenticeship, before you finalize the architecture.
Principle 4
Write a mission that compounds
"Solve intelligence, then use it to solve everything else" sounds grandiose because it is. It is also an argument about leverage. If the tool is general, each improvement makes every downstream application stronger. That is how a lab founded to play Atari ended up working on proteins, weather forecasting (where its models beat NOAA's on hurricane track and intensity in 2025), sorting algorithms now in the C++ standard library, and ancient Greek inscriptions.
The two-stage structure of the mission also handled a recruiting problem. Researchers who wanted to work on fundamental questions joined for the first half, and researchers who wanted real-world impact joined for the second. A mission naming a single product would have attracted only one of those groups.
Tactic: Phrase your mission as a capability followed by the domains it unlocks, so that each improvement in the core capability visibly expands what the organization can do next.
Principle 5
Stay where your unfair advantage lives
DeepMind took Peter Thiel's money without moving to Silicon Valley. At the time the conventional view was that serious technology companies had to be in the Bay Area. Hassabis saw that London gave him what Palo Alto couldn't: UCL, the Gatsby Unit, Oxford and Cambridge nearby, a pool of academic talent in neuroscience and machine learning with no local competitor for it, and lower salary expectations than California.
Being outside the main market became an advantage. DeepMind could hire researchers who would never have joined a Valley startup, could build a scientific culture separate from a product culture, and later sponsored machine-learning chairs at Cambridge, Oxford, and UCL that kept the talent supply going. The company is still in London in 2026, now in a new building at King's Cross.
Tactic: Map where the talent you most need is underpriced or unclaimed, and locate there even if investors insist on the obvious hub; then build institutional ties, such as chairs, fellowships, or joint appointments, that keep the supply flowing.
Principle 6
Sell to buy compute, and price the autonomy you lose
The £400 million sale to Google in January 2014 is usually described as an exit. Hassabis treated it as financing. DeepMind needed a backer with enough money to cover salaries and, most of all, compute at a scale venture capital couldn't provide. Google paid for AlphaGo, AlphaFold, and the TPUs both ran on. On the research outcomes alone, the sale is one of the most productive acquisitions in the history of technology.
But the price wasn't only the purchase price. DeepMind Health was absorbed into Google Health in 2018. The years-long campaign for independent governance ended in failure, reported by the Wall Street Journal in May 2023, a month after DeepMind was merged with Google Brain and placed at the center of Google's commercial response to ChatGPT. The autonomy DeepMind thought it had kept turned out to be on loan, and it was taken back when the parent company needed it.
⚖
The Bargain, Itemized
What DeepMind received from Google, and what it gave up, over twelve years.
| Year | Event | Ledger |
|---|---|---|
| 2014 | Acquisition for a reported £400M; anonymous ethics board | Gained Compute, runway |
| 2016 | AlphaGo beats Lee Sedol on Google TPUs | Gained Global credibility |
| 2018 | DeepMind Health absorbed into Google Health | Lost Control of applied health work |
| 2020–22 | AlphaFold 2 solves structure prediction; 200M structures released | Gained Scientific legitimacy |
| 2023 | Merger with Google Brain; autonomy bid fails | Lost Independent governance |
| 2026 | Hassabis becomes Alphabet chief scientist; leaves CEO role | Traded Operations for research focus |
Tactic: When you sell to a strategic buyer, list in writing the specific decisions you expect to keep control of, and assume each one will be renegotiated the first time the buyer's core business is threatened.
Principle 7
Name one lighthouse
In 2018 Hassabis called AlphaFold "a lighthouse project," DeepMind's "first major investment in terms of people and resources into a fundamental, very important, real-world scientific problem." The phrase did internal work. It told researchers, Google, and the public where the coast was, so that everything else could be read relative to that one project.
Organizations that chase many problems at once tend to spread their best people too thin. DeepMind ran dozens of research lines, but at any given point one flagship carried the lab's reputation: Atari, then Go, then proteins. Each took years and each was chosen to show something the previous one couldn't. Go showed the methods could handle extreme complexity, and proteins showed they could produce real science.
Tactic: Declare one flagship project per era, staff it with your best people regardless of their other commitments, and make its success criteria public so the whole organization can see where things stand.
Principle 8
Remove the human crutch on schedule
The original AlphaGo learned from about thirty million moves of human play. AlphaGo Zero learned from none, used four TPUs instead of forty-eight, needed a team of about fifteen people, beat its predecessor after three days, and won a hundred games out of a hundred. AlphaZero generalized the method to chess and shogi, and MuZero learned without being given the rules.
The lesson is about sequence. Human knowledge was useful for getting started and became a constraint once the system was strong. Hassabis's team planned to remove it. This applies beyond machine learning. Many organizations build processes around expert judgment, and the expert scaffolding that made the first version possible can become the ceiling on the next one.
♟
The Shedding Sequence
Each version of the Go program relied less on human knowledge, and each was stronger.
2015
AlphaGo trained on human games beats Fan Hui 5–0.
2016
AlphaGo beats Lee Sedol 4–1 in Seoul, on 48 TPUs.
2017
AlphaGo Zero learns from self-play alone on 4 TPUs; beats AlphaGo 100–0.
2017
AlphaZero extends the method to chess and shogi.
2019
MuZero masters Go, chess, shogi, and Atari without being told the rules.
Tactic: For each process that depends on expert judgment, set a date to test a version that relies on it less, and compare the results instead of assuming the experts are what makes it work.
Principle 9
Give away what you cannot fully exploit, and sell the layer above it
DeepMind released AlphaFold 2's code and more than 200 million predicted structures free, through a public database built with EMBL-EBI. Hassabis's reasoning was blunt: the company "could only do a tiny, tiny fraction of the downstream work that AlphaFold could enable." Two million researchers in 190 countries did the rest, including nonprofits working on neglected tropical diseases who would never have been paying customers.
That was not charity at the expense of strategy. In February 2021, months before the release, Hassabis incorporated Isomorphic Labs to capture value at the next layer up, designing drugs against the structures. The free database made AlphaFold the standard, and the standard made Isomorphic credible to Novartis, Eli Lilly, and Thrive Capital's $600 million round. Open access at the base layer created demand for the proprietary product above it.
Tactic: Find the asset in your portfolio whose value to the world far exceeds what you could ever capture from it, release it, and set up a separate entity to sell the next layer of capability built on top.
Principle 10
Treat trust as load-bearing infrastructure
The Royal Free agreement covered records from three hospitals serving 1.6 million patients a year, including diagnoses of HIV, depression, and abortion. The clinical value was real; staff called Streams "phenomenal." But the data transfer was found to rest on an "inappropriate legal basis," the trust was found to have breached data protection law, and DeepMind had to say publicly "we need to do better." The anonymous ethics board set up in 2014 made matters worse by signaling secrecy where openness was needed.
The general point is that in high-stakes areas, legitimacy is a dependency, not a public-relations task you deal with after launch. A technically excellent product built on a governance shortcut can lose more than its development cost, and the absorption of DeepMind Health into Google in 2018 shows that lost trust also reduces your negotiating position with your own owners.
!
Two Approaches to Legitimacy
The same organization handled trust very differently in two domains.
| Dimension | NHS / Streams (2015–18) | AlphaFold (2018–22) |
|---|---|---|
| Validation | Risk Contract leaked to press | Scale Independent CASP adjudication |
| Disclosure | Agreement surfaced via New Scientist | Code and structures published openly |
| Governance | Anonymous ethics board | Public partnership with EMBL-EBI |
| Outcome | ICO finding; unit absorbed by Google | Nobel Prize in Chemistry, 2024 |
Tactic: Before any project that touches sensitive data or public institutions, have an independent outsider review your legal basis and disclosure plan as if they were a hostile journalist holding the contract.
Principle 11
Run failure as a binary search
Hassabis has a distinctive way of thinking about failure: "I sort of think of it as splitting the hypothesis space into two, and you keep splitting it into two, and it doesn't really matter if in any particular one experiment it didn't work or it did work, it still helps you advance overall to the next question." That is a binary search. A well-designed experiment produces useful information whichever way it comes out.
The important condition is well-designed. An experiment with an ambiguous outcome teaches nothing either way. That connects back to Principle 2: you need a clear scoreboard to make failure informative. AlphaStar's early full-map-vision win, followed by the loss once the advantage was removed, is the method in practice. The loss told the team exactly what had to be fixed before the Grandmaster result could count.
Tactic: Before running any experiment, write down what you would conclude if it succeeded and what you would conclude if it failed; if both conclusions would be "unclear," redesign the experiment.
Principle 12
Recruit at the seams between fields
Hassabis's own progress happened where fields met: games and AI, AI and neuroscience, AI and biology. His advice to young researchers is to get "quite deep understanding and knowledge of at least two areas," because the biggest advances happen "in the in-between points where they connect." DeepMind started with machine learning, neuroscience, engineering, and mathematics and later added philosophers, ethicists, chemists, and physicists.
The founding team followed this rule. Hassabis brought games and neuroscience, Legg brought the mathematics of induction and superintelligence, and Suleyman brought applications and policy. Silver brought reinforcement learning and the Elixir experience. Each new domain, from proteins to weather to ancient Greek, required hiring people who could translate between AI and that field's specialists.
Tactic: When you hire for a hard problem, look for people with real depth in two fields that rarely overlap, and give them problems that sit exactly where those fields meet.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Demis Hassabis, Nobel Prize interview, Stockholm, December 6, 2024As we were playing against these chess computers, I remember being fascinated by the fact that someone had managed to program this inanimate lump of plastic to actually play chess.
— Demis Hassabis, Nobel Prize interview, December 2024Even when we started DeepMind, back in 2010, nobody was really working on AI, certainly not in industry... At the time, everyone thought we were pretty crazy. Everyone was like, "well, we know AI doesn't work." But I was going to do it no matter what.
— Demis Hassabis, Nobel Prize interview, December 2024I've always been in a bit of a hurry. I think I've always had this notion of that life's pretty short if you want to explore and experience all the amazing things in the world and maybe add a little bit of impact and value to that.
— Demis Hassabis, Nobel Prize interview, December 2024The science of AI is about trying to explore and understand what intelligence is. And the best expression of understanding something is actually trying to build it.
— Demis Hassabis, Nobel Prize interview, December 2024I don't really see any distinction in my life between work and life. It's all just one big adventure.
Maxims
- Timing is a variable, not luck. A problem twenty years ahead of its hardware is a mistake in timing, however good the idea.
- Find the scoreboard first. Choose problems where an independent judge can say whether you won.
- Study the only working example. Before building something new, learn how the closest existing thing actually works.
- Make the mission compound. Build the general capability and let the specific applications follow from it.
- Stay where your edge is. Locate where your talent is underpriced, not where investors expect you to be.
- Every sale has a hidden price. The autonomy you keep after an acquisition is on loan and can be reclaimed.
- Plan to drop the scaffolding. Human expertise that gets the first version working can cap the next one.
- Give away the base, sell the layer above. Open access sets the standard and creates demand for the proprietary product built on it.
- Trust is infrastructure. A governance shortcut can cost more than the whole product earned.
- Keep playing. The habits that formed you are worth practicing long after they stop being your job.
