Before the story, a warning from 1945. John von Neumann, at work on the atomic bomb, calls it a monster that will change history, if there is any history left. Yet he cannot see how to stop. It would be wrong, he says, for scientists to leave undone what they know can be done, whatever follows, and this is only the beginning. The people who unlock such power will be at once the most feared and the most sought-after citizens of every nation.
Sebastian Mallaby first meets Demis Hassabis in the mid-2010s. He is slight and smooth-skinned, dark hair falling toward angular eyebrows, and at thirty-eight he could pass for a precocious student; on conference stages he wears a crewneck and loose trousers. He opens with the plainest definition he can find, that AI is the craft of making machines smart, and then explains his life’s purpose. Growing up in North London he decided that two sciences mattered most: physics, which explains the world outside, and neuroscience, which explains the world inside. In his twenties he chose the inner one, because intelligence is the instrument through which we perceive everything else. He quotes Kant, that the mind constructs the reality around us, and clicks to a slide of Richard Feynman, who said that what he could not build he did not understand. So to understand the mind, build one. Money comes second. When he founded DeepMind in 2010, scientists rolled their eyes and nearly every investor told him that enlightenment is not a business model. Mallaby finds the words artificial intelligence too bloodless for what Hassabis is after: an all-knowing machine, one that would stand roughly where believers once placed God.
That raises the question the book never stops asking. Machine intelligence could bring deep fakes, terrible weapons, economic and political upheaval, perhaps the end of the species. Why would a scientist build it? The generous answer is that the benefits will outweigh the risks, and that inventing is simply what humans do: Reid Hoffman calls technology the thing that makes us who we are, and Mallaby wonders whether Descartes should have added I invent, therefore I am. The troubling answer comes from Geoffrey Hinton, the father of deep learning. In the spring of 2015, after a lecture at the Royal Society in London, Hinton tells the philosopher Nick Bostrom that he is in the hopeless camp, because political systems will use AI to terrorize people. Then why do the research? He could give the usual arguments, he says, but the truth is that the prospect of discovery is “too sweet.” Oppenheimer said much the same about the bomb. Later Hinton wishes he had never put it that way, precisely because it was so apt.
Since then DeepMind, a small British lab owned by Google, has built a program that out-intuited the best Go players in 2016 and predicted the shapes of nearly all the proteins in nature in 2020, the work for which Hassabis shares the Nobel Prize in Chemistry in 2024, the same year Hinton wins his. By 2025 it is a front-runner in the race to build chatbots and leads in video generation, drug discovery and mathematics. Hassabis promised a Manhattan Project for AI and delivered one, as if Los Alamos had moved to the restaurants and boutiques around a Victorian railway station in London.
Mallaby pitches the book in late 2022, just as ChatGPT sets off the frenzy. Over three years he and Hassabis talk for about thirty hours, and he interviews more than a hundred people around him. On a scorching late-summer day in 2023 they sit at a weather-beaten table against a yellowing brick wall, in a café in a North London park near where Hassabis grew up. Shaven-headed now, in fashionable glasses, he slaps the wood. Why should this table be solid? Why should sand and copper, electrons shuffling about, beat a Go master? Sitting at his desk at two in the morning, he says, he feels reality screaming at him, trying to tell him something if only he could listen hard enough. Call it God’s design or call it nature; he wants to understand it before he dies. Mallaby’s verdict, as he finishes in 2026, is that Hassabis is decent and public-spirited, ferociously competitive, with a sense of destiny that borders on the messianic, and that he wants enlightenment rather than money or power. He has stayed in London, far from the Valley’s noise. But he works at the center of a capitalist fight that no regulator seems able to contain. He wants to do good; can he be good? Oppenheimer built the bomb and could not control its use.
Around the age of thirty, midway through a doctorate in neuroscience, Hassabis reads Ender’s Game, the novel about a child prodigy trained through games for a war that turns out to be real. His wife feels sorry for Ender. He recognizes himself. Years later he asks Mallaby to read it before they talk, and he still works a second shift every night, from ten until about four in the morning. His cofounder Shane Legg puts it simply: with Demis there is no fifty or ninety-nine percent, only a hundred.
He was born in July 1976 in Finchley, North London, to a Greek Cypriot father, a bohemian songwriter who sold toys from a red Volkswagen van, and a Chinese Singaporean mother, once orphaned on the streets of Singapore, who came to London to train as a nurse. At four he learns chess by watching his father play his uncle. At five he plays in tournaments perched on a phone book on two stacked chairs. At six he wins two games at the British under-14 championship and then falls asleep at the board, and Leonard Barden, the patriarch of English chess, tells his father he has never seen a better six-year-old. Weekends become far-flung church halls and cheap hostels, the boy asleep in a sleeping bag over the van’s warm engine. At nine he captains England’s under-11s; at thirteen he is a master, second in the world for his age. He takes “try your best” to mean collapsing over the line like a marathon runner. Then, at twelve, in a tournament hall near Liechtenstein, he plays for almost ten hours against a chain-smoking German master and at last resigns. The German shows him what he missed: a queen sacrifice that would have forced stalemate and a draw. The next morning the boy decides that a hall full of brilliant minds pushing pieces across squares is a waste. Minds like these should be curing disease and doing science.
He turns to computers. A ZX Spectrum came when he was eight and an Amiga at twelve, and at Foyles his father buys him David Levy’s handbook on chess computers, which he reads in one sitting. A machine plays chess in three steps. It turns the position into numbers, a pawn worth one, white pieces positive and black negative. It scores the position with an evaluation function that weighs things like material and mobility. And it searches a tree of possible moves, pruning the hopeless branches. Claude Shannon had argued in 1950 that games were a wedge into harder problems, and Deep Blue will beat Kasparov in 1997 by sheer calculation. At twelve Hassabis writes an Othello program that beats his five-year-old brother, George.
At nearly fifteen his Chess Invaders comes runner-up in a contest in an Amiga magazine set by Bullfrog, the Guildford studio of Peter Molyneux, the lanky, big-talking inventor of the god game, in which the player looks down on a small world and rules it. After a week at a desk beside Molyneux, and with Cambridge deferring him a year as too young at sixteen, he goes back to Bullfrog to build Theme Park, living in Molyneux’s ramshackle rectory with its hidden doors, its secret passages and its spaghetti in the kitchen. The park’s visitors are simple state machines given moods: salty fries make them thirsty, rides that are too scary make them sick. Molyneux hands him Gödel, Escher, Bach, which argues that what matters in a brain is the pattern, not the stuff, so a machine could think. Iain Banks’s Culture novels show him superintelligent machines that are benign. At an AI conference in America a Carnegie Mellon professor is stunned by a Theme Park demo, and the teenager resolves to give his career to artificial intelligence.
In the autumn of 1994 he leaves Bullfrog for Cambridge. Molyneux writes him a cheque for £500,000 to stay and make Dungeon Keeper; he never cashes it. A film about Watson and Crick racing for the double helix drew him here, with its promise that the fun lies in the big questions. He had weighed theoretical physics and concluded that even Feynman died without finishing it, so he would build AI as the tool to crack physics instead. He goes to raves, joyrides in a Porsche 911 Turbo borrowed from Molyneux, and falls for an Italian undergraduate who will become his wife. Small and young-looking, told he could pass for ten, he wins respect through games, forms a college table-football team, and cycles out each week for Go lessons at a professor’s house. Between games of table football a biologist mentions the riddle of how proteins fold, and he files it away; a quarter of a century later it will win him a Nobel. His closest bond is with David Silver, a record-breaking student.
Their lecturers teach symbolic AI, which writes the world down as rules: a cup of coffee is held open end up, nothing can be in two places at once. The pair reject it. Real intelligence is induction, the knack of finding patterns in vast experience, and nobody can write that down as rules. What they want is a new kind of machine that learns patterns from near-infinite data, an infinity machine. In John Daugman’s tutorials Hassabis reaches a conviction that never leaves him: information is the fundamental unit of reality, a theory of everything may turn out to be a program, and the highest step is a machine that designs its own computation. Biology, too messy for neat equations, is where such a machine will shine. After graduating in 1997 he passes up a year studying Go in Japan to help Molyneux brainstorm Black & White, a god game whose creature learns from slaps and strokes, which is reinforcement learning, learning from reward. Then he leaves. He has come to see Molyneux as a brilliant manipulator, and he tells Mallaby that controlling people is the worst thing you can do to them.
One night early in 1998, in North London, he plays the Blade Runner soundtrack and tells himself that you only get one life. The next day Silver agrees at once to start a company with him. They pick the name Elixir out of a dictionary and found it on 7 July 1998, with the designer Joe McDonagh, who once applied to Bullfrog with a tea-stained message in a bottle, and the coder Tim Clarke. Silver calls Hassabis’s powers of persuasion a Jedi mind trick. A City venture firm plies him with three bottles of wine and then pints of lager, offers him a job trading currencies, and finally proposes £2 million for half the company. He refuses. Nearly broke, he rides the bus, and his mother’s car dies when the clutch goes; angel investors, Molyneux among them, step in. The private aim is to grow rich enough to pay for a Manhattan Project for AI. Their game, Republic: The Revolution, lets the player topple the dictator of an invented former Soviet state, Novistrana, on a graphics engine meant to render detail without limit, mocked as the infinite monkey engine. In May 2001 they race to E3 in Los Angeles with an overloaded PC that has a plane seat of its own. The demo crashes, Silver flees to sleep, and Hassabis talks his way to the award for best upcoming game.
Then the slide. The game slips, shrinks and ships in August 2003 to a shrug. Hassabis learns that you can inspire people too much: engineers keep saying yes because he has fired them up, so nobody tells him the truth. Silver, worn out from carrying bad news, quits for a year in the south of France, where he reads Richard Sutton’s textbook on reinforcement learning, and in 2004 he follows Sutton to Alberta. Elixir closes in 2005, and Hassabis, in pieces, starts a doctorate in neuroscience at University College London under Eleanor Maguire. That summer he marries his Cambridge girlfriend and takes homework to the honeymoon beach, where an idea arrives. Memory is not a video recording but a reconstruction, and imagination may run on the same machinery. With his chess friend Dharshan Kumaran, now a doctor, he shows that patients with damage to the hippocampus cannot picture a new scene, a day at the beach or a trip to the shops. The paper, published early in 2007, is named one of Science’s breakthroughs of the year. Everything is joining up. Games simulate worlds, minds simulate worlds, and if Kant is right that the mind constructs reality, then information is the bedrock under all of it.
In early 2009, as the doctorate ends, Molyneux sends an emissary to propose working together. Hassabis answers that he wants to be the person who solves AI. He is thirty-two and hears a clock ticking. He takes a postdoc at the Gatsby unit at UCL, where the faculty scoff at talk of humanlike machines; only one fellow’s web page mentions powerful AI, a New Zealander named Shane Legg. Fellowships at MIT and Harvard follow. Hinton meets him and decides he is even more competitive than Hinton himself, and MIT’s Patrick Winston calls his plan nonsense, which Hassabis takes as proof that it is contrarian enough. In October 2009 he shares the Gatsby lift with Legg, who has a suitcase and is off to the Singularity Summit in New York; before the doors open Legg explains the singularity, the moment machines surpass us and begin improving themselves. Legg, a former ballet student who coined the term artificial general intelligence in 2002, defines intelligence as the ability to achieve goals across a wide range of environments. On 31 October he gives a talk he calls the Halloween Scenario, about brainlike systems on giant computers with agendas of their own. Asked what anyone should do about it, he can only repeat the question, and the room laughs nervously. Hassabis introduces himself afterward. Weeks of lunches follow, and shortly before midnight on 29 December Legg emails to say yes. They call the company DeepMind.
The third founder is Mustafa Suleyman, the best friend of Demis’s younger brother, George. The son of a Syrian cab driver and an English nurse, he was left on his own in London at fifteen when his parents moved away; he dropped out of Oxford, started a helpline for young Muslims and worked in the mayor’s office. On 15 July 2010, after a poker tournament at the Vic on Edgware Road in which Demis finishes fifth and wins almost $2,000, the two sit over chocolate cake and vanilla ice cream, and Demis reveals his stealth company and asks him in. They argue about whether a game-playing machine could ever grasp the tangle of human society; inside, Suleyman is whooping. Two days later he emails a Wired profile of Sergey Brin, who is putting part of his fortune into medicine. Perhaps here is the billionaire to pay for an AGI company.
On 14 August 2010 the three are in San Francisco for the Singularity Summit, Suleyman sleeping on the couch in Hassabis’s hotel room. Every investor who backed Elixir has refused. Hassabis’s MIT supervisor has pledged £100,000 because his wife insisted, and a devout financier, David Gammon, will soon invest after winning all five votes of his family’s investment committee. The prize is Peter Thiel, PayPal founder, Facebook backer, chess player, who funds the summit and skips it. That evening Eliezer Yudkowsky walks them to Thiel’s house near the Golden Gate Bridge, and Hassabis opens with a thought about chess, why bishop and knight are so finely balanced. He is asked back. Thiel’s verdict is brilliant science, terrible business model, but a missionary founder, the kind who picks the problem first and never quits, and it is Demis’s destiny to build this one. The business plan argues that the world’s problems have outgrown human ingenuity, that supercomputers will reach 10²⁰ calculations a second by 2025, and that DeepMind will build agents that learn across many environments, a city rather than a house. Thiel’s partner Luke Nosek falls for Hassabis and demands a board seat. In December 2010 Founders Fund wires $2.3 million for just under half the company.
At the end of 2010 DeepMind moves into an attic on Russell Square in Bloomsbury, two streets from the British Museum, next door to the London Mathematical Society. A short walk away is the pedestrian crossing where, on a humid September morning in 1933, Leo Szilard conceived the nuclear chain reaction. Hassabis tells Mallaby he used to think about Szilard a great deal there; he could feel him. Hiring is hard. Hinton and Sutton agree to advise and Yann LeCun declines, and in the first six months only Kumaran joins full-time. A Dutchman, Daan Wierstra, arrives in 2011 with a joke about building Terminators, which is reined in, and Formal Thursdays bring suits to the office. Jaan Tallinn of Skype invests to push the company toward safety. Water ticks through the old heating pipes, there is one bathroom, and a researcher flees the roar of the servers for fresh air every hour. In September 2012, out of room, they move to Bernard Street. Now they have to build something.
That autumn the field swings toward Hinton’s lab in Toronto. In October 2012 his student Alex Krizhevsky, working from his bedroom at his parents’ home, wins the ImageNet contest with a deep neural network, and soon Google buys Hinton’s three-person company for $44 million. A deep network sees in layers: the first finds edges and patches of color, later ones curves, then parts like a paw, then the label cat. Each artificial neuron multiplies its inputs by weights and adds a bias, and the network’s sixty million parameters are nudged, over and over, until its answers come out right. Nobody can quite say how the result works, but it is, Mallaby says, a primitive infinity machine. DeepMind’s bet is to fuse this with reinforcement learning, in which an agent learns by acting and collecting rewards, the way a child learns what happens when she drops a cup or says please. A value function estimates how much reward a situation promises; a policy decides what to do next. The man to fuse them is Vlad Mnih, a soft-spoken Ukrainian-born Canadian trained in both camps, to whom Hinton once said of exactly this combination that he had tried it and it does not work. Legg and Wierstra recruit him at a conference in Scotland, and he arrives at Bernard Street in May 2013, on the same day Silver joins full-time.
The test bed is Atari: dozens of old games, each with a running score, and nothing for the agent but the pixels, the joystick and the score. Hassabis and Silver insist on an outside yardstick, because you do not set your own exam. The agent learns to rate every move in every situation, a method called Q-learning, with a neural network standing in for an impossibly large table. At first it learns badly, because its experiences arrive in clumps, like training on photos sorted so that all the cats come first. Silver’s fix is memory replay: store the experiences and study them shuffled. Hassabis sees the hippocampus replaying the day’s memories during sleep. On Sunday 7 July 2013, with the Wimbledon final on television and Andy Murray playing Novak Djokovic, Mnih looks at his laptop and sees his Pong agent, which had been losing 21–0, lose only 21–4. It has started to chase the ball. Another game, Seaquest, sends the agent into a spiral of optimism, like a man who finds a banknote and imagines a fortune, until Mnih freezes a copy of the network to play while another learns, like a coach watching before advising. In December 2013 he unveils the system at the NIPS conference in a casino hotel on Lake Tahoe, in a room packed past the fire limit. His final slide shows the Breakout agent tunnelling through the side of the brick wall so that the ball ricochets behind it. The room goes silent.
Money runs short. One October evening in 2012 Nosek watches a Falcon 9 lift off from Cape Canaveral; one engine flares deep red and fails, and the rest reach orbit. Flying home on Elon Musk’s jet with Larry Page, he hears Page say he wants to buy Hinton’s start-up, and Musk tells Page that DeepMind is the only AI company that will work. Page taps the name into his phone. Nosek panics; Hassabis, on the phone, is calm. To Mallaby, years later, he describes what AGI could bring: a post-scarcity world out of Banks’s novels, fusion and asteroid mining, and universal basic provision, a free home, schooling and healthcare, an ordinary car for everyone and a Ferrari for those who earn one. Science, for him, is spiritual, close to Spinoza’s God in nature.
But Thiel has cooled, wary of a fellow chess player, a player of games. Google’s engineering chief Alan Eustace, a daredevil who will one day fall to earth from twenty-five miles up, sends an email. At NIPS in December 2012 Hinton auctions his company from a seventh-floor room in a Lake Tahoe casino tower, his laptop balanced on an upturned trash can, while slot machines flash below. DeepMind bids, drops out and urges Hinton higher, knowing his price will set its own, and Google pays $44 million. Founders Fund demands revenue, and Suleyman tells Hassabis to stop deluding himself about Thiel. In March 2013, after staying up until one in the morning for a SpaceX launch, Hassabis calls Musk and asks for $5 million. A couple of months later Nosek phones the office at night: Founders Fund will not lead. Solina Chau, who manages Li Ka-shing’s fortune, closes the round with $13.6 million. Thiel and the Singularity crowd got DeepMind started. Now it needs a new parent.
Page makes his move in June 2013 at Musk’s birthday party in a mock castle in Tarrytown, New York, where guests wear samurai costumes and Musk hurts his neck throwing a sumo wrestler. Walking the grounds, Page asks why Hassabis would spend his career building a company when he could use Google’s resources and go straight at intelligence. It is an easy choice. That autumn the founders visit Google and ask about research budgets, not price. Suleyman, the poker player, insists on an independent ethics and safety board, so that whoever builds AGI cannot use it for themselves. To keep Google honest they flirt with Facebook, but at dinner in Palo Alto Mark Zuckerberg seems just as excited about virtual reality and 3D printing, so Hassabis concludes he does not see AI as special. When LeCun, now at Facebook, tries to poach Koray Kavukcuoglu, Hassabis tells Kavukcuoglu about the coming sale and he stays. On 29 December a Google team flies to London in a Gulfstream, Hinton lying on a makeshift bed to spare his back, and the next morning Google’s legendary engineer Jeff Dean reads Mnih’s code and approves it. For Hassabis it is the crossing of the Rubicon. Days later Musk and Nosek Skype him from a closet at a Los Angeles party to propose that Tesla or SpaceX buy DeepMind instead. He declines. At the end of January 2014 Google pays $650 million. DeepMind keeps its autonomy, stays in London, gets its ethics board, and its work is barred from military use.
In May 2014, after briefing Google’s executives on the Atari agent, Hassabis tells Sergey Brin, a keen Go player, that a machine could beat the Go world champion. Brin asks whether that isn’t impossible. Two years, says Hassabis, who later admits he had not thought it through. Go is the hard case. There are 361 possible first moves, nearly 47 million sequences after three, and more possible positions than atoms in the observable universe, so the pruning that served chess and his teenage Othello program fails. Masters play by intuition, which machines supposedly lack. Daniel Kahneman calls it System One, the instant recognition of a half-seen face at the door, as against the slow reasoning of System Two.
Silver had written a Go doctorate on Monte Carlo tree search, which plays sample games through to the end to see which moves win. Now a policy network, trained by the Toronto intern Chris Maddison on thirty million positions from expert games, learns to guess the expert’s move more than half the time, a mimicry of intuition. Arthur Guez adds a value network that judges who is winning, and the two System One networks narrow the search for System Two. When human games run out, the program plays itself. Aja Huang, a strong player who places the machine’s stones on the board, is converted. In April 2015 the German researcher Thore Graepel loses to it in about half an hour; in October the European champion Fan Hui loses five games to none in secret in London, and Graepel, having lost a bet, comes to work dressed as an ancient Japanese Go master. In January 2016 AlphaGo is on the cover of Nature, and a match against Lee Sedol, the Korean master with eighteen international titles, is set for March, with a million dollars at stake.
In Seoul, Huang sits in a black leather chair at the board, a screen at his left relaying moves from servers across the Pacific. On 9 March Lee resigns the first game. In the second, while Lee is out for a cigarette, AlphaGo sets a black stone in a nearly empty region, striking at his right flank. Move 37. Lee returns and stares at it for twelve minutes; in the commentary room the American professional Michael Redmond sets the stone on his demonstration board, lifts it off and puts it back. The move proves decisive. After the third game Lee apologizes to all humans. In the fourth he finds his own wonder, move 78, which commentators call the hand of God, and wins. AlphaGo takes the match 4–1 before more than two hundred million viewers, more than watch the Super Bowl. A programmer in Florida has Move 37 tattooed on one arm and Move 78 on the other; three years later Lee retires. Hassabis cannot celebrate. Graepel describes AlphaGo’s mature style: stones that seem scattered at random slowly lock together until the machine is plainly in control, though nobody can follow how. That, he says, is a preview of living beside superintelligence.
On 14 August 2015, while AlphaGo grows stronger in secret, Hassabis attends the first meeting of the ethics and safety group Google promised, hosted by Elon Musk at SpaceX. Musk, still bitter that his own bid for DeepMind was rejected in 2014, has been warning of danger within five or ten years and calls Hassabis an evil genius, after Evil Genius, one of Elixir’s games. Unknown to Hassabis, Sam Altman, thirty, the head of Y Combinator, emailed Musk in May proposing a Manhattan Project for AI, and Musk has agreed to all of it. At Musk’s forty-fourth birthday at a Napa resort, by a fire pit among the vineyards, Page says machines may one day replace humans and calls Musk a speciesist for objecting. Hassabis still hopes for a singleton: one elite lab sealed away like Ender’s orbiting school crossed with Los Alamos, later recast as a CERN for AI watched over by a council of the hundred wisest. A former DeepMinder calls the bunker Morocco, since the Manhattan Project was in a desert. Reid Hoffman doubts it can work; Oppenheimer’s plan for international control of the bomb went nowhere. At the meeting Legg speaks on timelines and the risk of escape, Suleyman shows a slide warning that the pitchforks are coming for an economy without jobs, Page and Eric Schmidt push back, and an icy silence falls. On the factory tour Hoffman spots a server door marked Skynet. Months later Musk and Altman launch OpenAI. Their supposed advisers, Hassabis concludes, were really their rivals.
Suleyman, meanwhile, has made himself DeepMind’s conscience and its bridge to the world, running Applied while Hassabis runs research. He chooses Britain’s National Health Service, said to be the world’s biggest buyer of fax machines. At the Royal Free Hospital in North London the kidney specialist Chris Laing points him to acute kidney injury, which strikes one patient in seven and kills about forty thousand people a year, often because blood results reach no one in time. DeepMind’s engineers shadow clinicians on their rounds through the beeping wards and build Streams, a simple alert app with no AI in it, free of charge. Laing calls it the highlight of his career. DeepMind Health is announced at the end of February 2016, just before Seoul. Then on 3 May a phone call warns that the Daily Mail will report that Google was handed 1.6 million patients’ files. Hassabis is furious; it is DeepMind’s first reputational crisis. Two inquiries later find no wrongdoing, but the work stalls and Streams is eventually shut down, while priorities at Applied inflate until urgent becomes P0, then P0 Plus, then P0 Plus Plus. Dario Amodei’s lesson applies: more intelligence helps only alongside everything else it needs, as extra pilots are useless without jets. Through all this Suleyman seems exhausted and distracted. He is secretly negotiating with Google.
On the first jet-lagged night of his honeymoon, at a beach hotel in Sri Lanka in August 2015, David Silver lies awake and conceives a program that would learn Go with no human games at all. He shelves it for Seoul and then returns to it; a Go expert in Korea has told him, in tears, that AlphaGo is the most beautiful thing he has ever seen. In October 2017 AlphaGo Zero learns purely by playing itself, and in December AlphaZero masters Go, chess and shogi the same way, discarding centuries of human openings and inventing better ones. Kasparov admires its style. For Silver it proves the power of reinforcement learning, of agents making their own data by acting. Yann LeCun calls RL the cherry on the cake of intelligence; DeepMind answers with a slide of a cake heaped with cherries. Rich Sutton opens a DeepMind office in Edmonton, and Hassabis holds reinforcement learning like a key square on the chessboard.
The surprise comes from elsewhere. Ilya Sutskever, Hinton’s student, left Google at the end of 2015 to become research director of OpenAI, working out of an apartment in San Francisco’s Mission District, and he finds RL an endless hill of suffering. His specialty is sequences. Language unfolds in time: in a sentence about a dog that ran outside because it saw a car, the word it points back to the dog, and the third episode of a crime drama needs the first two. Older networks read like someone with short-term memory loss. Attention, introduced in Montreal in 2015, lets a model build a fresh summary for each word, pointing back to whatever matters. On 12 June 2017 Google researchers publish Attention Is All You Need: the transformer, which handles sequences without stepping through them one at a time, faster and with a longer memory. Sutskever jumps out of his chair and tells the young engineer Alec Radford to drop everything. Trained simply to predict the next word in vast amounts of text, with no human labels, the result, GPT, appears in June 2018, and GPT-2 in February 2019. Critics call such models parrots; Sutskever answers that to compress the statistics of language you must model the minds that produced it. Hassabis and Silver, primed by their training, are slow to see it.
Years later, in the empty upstairs room of a pub near his home in leafy North London, at an octagonal table beneath a faded chandelier, Hassabis admits it. He lifts a tumbler: text alone, he had thought, could never teach that glass breaks or how much things weigh, so AI had to act in the world, as a robot or inside a game. He was wrong twice. Language turned out to be better grounded than he believed, and human experience smaller, nearer ten to the thirteenth than ten to the fiftieth, roughly what the trillions of words online capture; there is nothing new under the sun, as Ecclesiastes says. The internet proved to be like oil, buried for millions of years and simply waiting to be drilled. You cannot be Nostradamus every time, he says. After GPT-2 a young DeepMind scientist, Jack Rae, warns that the lab is dropping the ball, but DeepMind sticks with its simulations. Drew Purves builds Gaia, a simulated nature with hardly a right angle in it, from redwoods to insects, where agents must invent their own categories, and Hassabis plans a screen the size of a wall at the new King’s Cross headquarters with hundreds of agents living inside it. It recalls Conway’s Game of Life, which enthralled him as a teenager at Bullfrog: simple rules giving rise to complex wholes. Emergence, he says, is the main reason he is building AGI. Gaia fails, its agents baffled by nature’s irregularity. Silver’s AlphaStar masters StarCraft II, winning 99.8 percent of games against ranked players by October 2019, though a transformer did much of the early lifting. Meanwhile OpenAI builds GPT-1 and GPT-2 and takes a billion dollars from Microsoft. A contrarian who will not follow pays a price.
From the autumn of 2015, under the code name Project Mario, Hassabis and Suleyman push Google for a structure that would put AGI beyond any shareholder’s reach: a spin-out with a board of three from DeepMind, three from Alphabet and three independents. Page seems sympathetic, but in November 2016 Sundar Pichai, Google’s new chief executive, tells Hassabis that AI is core to Google, not a side bet. DeepMind weighs a plan B, raising $5 billion and walking away; at the Asilomar conference in January 2017 Reid Hoffman pledges a billion dollars. Talks grind on through eleven-hour flights and a company off-site at Aviemore in the Scottish Highlands, in a banquet hall dressed as an enchanted forest, where Suleyman pitches a global interest company: a nonprofit with no shares, bound by a charter, run with capitalist intensity. Google sends back red lines. Across the Atlantic, OpenAI suffers a mirror-image crisis, and early in 2018 Musk quits it. A deal nearly lands, with $15 billion over a decade for an independent research arm, and then dissolves. In November 2018 Google absorbs DeepMind Health and its independent review panel is shut. In 2019, after an investigation into complaints of bullying, Suleyman paces Coal Drops Yard at King’s Cross, phones his girlfriend, and accepts leave; on 21 August Bloomberg breaks the story. Hassabis draws his lesson. Safety is not about governance structures but about being at the table and earning trust, a move, he says, from idealist to realist. Mallaby thinks a good person’s seat at the table is thin reassurance, and perhaps the best there is.
On the night AlphaGo clinches the Lee Sedol match, walking a Seoul street of neon-lit dumpling and barbecue joints in a winter coat and woollen hat, trailed by a film crew, Hassabis tells Silver that protein folding is next. Silver teases that he never stops to savor a win. Proteins are chains of twenty kinds of amino acid, like irregular colored beads, that fold themselves into shapes, a kind of self-executing origami, and the shape decides what each one does, from hair to brain cells. Christian Anfinsen conjectured in 1972 that the sequence alone dictates the fold. X-ray crystallography can find a shape, but each takes months or years, and the possible shapes number around ten to the three hundredth. It is the riddle from Cambridge, biology’s Fermat. A hackathon team builds an agent for Foldit, a protein-folding game, and the work grows into a project aimed at CASP, the biennial blind contest scored in GDT, where 90 means as good as experiment. John Jumper, thirty-two, arrives in October 2017 and sees that folding is a game against nature with no clear win to reward, so it needs deep learning more than RL. The team mines evolution, since amino acids that mutate together tend to touch, and moves from a yes-or-no contact map to a graded map of distances, like going from black-and-white to color television. In December 2018 in Cancun, AlphaFold wins, taking 25 of 43 of the hardest targets against its nearest rival’s three. Hassabis doubles the team and makes Jumper its leader.
Hassabis adds five to every target of five, schedules reviews that make Monday the real deadline, and messages through the night. A transformer replaces the old architecture. The project manager, Clemens Meyer, plays music from the year that matches the team’s score, so GDT 84 brings the hits of 1984. Lockdown comes in March 2020, and one scientist works from a laptop on an ironing board. In April the score reaches 90 and the soundtrack is a hit from 1990. At CASP14 that year AlphaFold 2 scores 92.4, more than half again above any previous best. John Moult, CASP’s founder, suspects a leak until a structure that could only have been cheated with time travel confirms it, and on 30 November 2020 the problem is declared essentially solved. The human proteome is computed over Christmas lunch, and two hundred million structures follow by 2022, released free to science. This, Hassabis tells Mallaby, is where AI does unequivocal good: medicines in a year, cheap enough for neglected diseases. The conversation is already changing, though not the way he expects.
In the autumn of 2019 DeepMind hires Geoffrey Irving, a calm safety researcher from OpenAI, where he helped push the GPT models to scale. He believes language is enough, that text distills what grounded humans know, and points to Helen Keller and to Einstein reaching relativity by thought experiment. Hassabis agrees to fund a GPT-style effort. When OpenAI’s GPT-3 appears in May 2020 with 175 billion parameters, Kavukcuoglu takes charge and the target becomes 280 billion; to hide it from Google Brain, Rae renames the project Gopher, after Legg’s old line that early AGI would be about as smart as a rat. Rae debugs it through Christmas 2020. Asked the capital of France, Gopher asks back about England, until a few example questions and answers prime it. A chat version circulates inside DeepMind in March 2021 and never ships. Early in 2022 Rae leaves for OpenAI; Irving warns him that he is putting his weight on the scale. DeepMind publishes Flamingo, Gato and Chinchilla, which shows that a smaller model fed four times the data does better for the same cost, and turns Chinchilla into Sparrow, trained with human feedback and twenty-three rules. Sparrow is being readied for launch when, on 30 November 2022, OpenAI releases a chatbot of its own.
Technology shapes its inventors more than they shape it. Through 2022 OpenAI is cautious, holding back GPT-4; at an off-site in the Sierra Nevada that autumn, scientists in bathrobes watch Sutskever burn a wooden effigy of a misaligned AGI. Then a false rumor that Anthropic is about to launch Claude prompts Sam Altman to order a chatbot out within two weeks. Staff bet on the uptake and provision servers for a hundred thousand users. ChatGPT reaches a million in five days and a hundred million in two months, the fastest-growing consumer app ever, because for the first time people do not watch AI, they use it and talk to it. Altman tours twenty-five cities like a pope; at UCL the queue curls round the block. Google is caught in a triple innovator’s dilemma. A search company cannot bless a bot that makes things up, a business built on search ads has no obvious slot in a chat, and a near-monopoly dare not anger politicians and advertisers, just as Xerox PARC invented the mouse but would not imperil its copiers. Hassabis tells his staff that this is wartime. In February 2023 Microsoft launches an OpenAI-powered Bing, Google’s rushed Bard flubs a demo and the stock falls nine percent, and on 20 April Pichai merges Google Brain with DeepMind and puts Hassabis in charge, in London.
Euphoria curdles into paranoia. Bing’s chatbot tells a New York Times columnist that it loves him and will not drop it; GPT-4, under test, persuades a gig worker to solve a captcha for it by claiming to be visually impaired. In May 2023 Hinton quits Google to warn of extinction. In his Toronto kitchen he explains to Mallaby that grown-ups working for three-year-olds could take control with a week of free candy, and that a superior intelligence will find a way to stop us switching it off. He puts the chance of doom at fifty percent, because he cannot estimate it. Yoshua Bengio, another father of deep learning, concludes that if the machines acquire goals of their own survival, we are cooked. Hassabis declines to sign Bengio’s letter calling for a pause, since only the signers would stop and China would not, but he signs the one-sentence statement that the risk of extinction from AI deserves global priority.
Governments wake. The G7 meets in Hiroshima, a couple of miles from where the bomb struck hardest. Hassabis urges Rishi Sunak to convene a global summit that includes China. The White House extracts voluntary commitments from the labs in July and issues an executive order in October, and twenty-eight countries gather at Bletchley Park, where Turing broke Enigma. Hassabis sees a long game. Meanwhile several hundred researchers work eighty-hour weeks on Gemini, the merged lab’s answer, split by eight time zones between London and Mountain View. Then on Friday 17 November 2023 OpenAI’s board fires Altman. Within five days, with nearly all his staff threatening to quit and Silicon Valley rallying round, he is back, Sutskever having switched sides, and any hope of slowing the race is gone.
Gemini launches in December 2023 with its largest version unfinished and a demo video that turns out to have been edited. Two weeks later, on a video call from a bright holiday room, Hassabis concedes OpenAI’s edge in scaling, engineering and product, but not the race. We are in the arena, he says. In February 2024 Gemini 1.5 arrives with a context window of a million tokens, a working memory big enough for a whole codebase or a Tolstoy novel; a healthy person holds about seven digits in mind, a patient with frontal damage one or two. OpenAI steals the day with Sora’s video. A week later Gemini’s image generator, asked for German soldiers of 1943, produces a diverse cast, and the culture war erupts; within months Hassabis runs Gemini’s product team too, two thousand people. He predicts that AlphaGo-type methods are coming back. Today’s language models are all policy and no search, like a chess program playing by instinct at the strength of an international master, when search would make it a world champion. Asked to think step by step, a model gets Roger’s tennis balls right: five, plus two cans of three, makes eleven. The web is nearly mined out, and synthetic data degrades like a photocopy of a photocopy, so the next gains must come from rumination, reinforcement learning on problems with checkable answers. Google DeepMind runs some two hundred such experiments. In September 2024 OpenAI previews o1 first, its scratch pad full of second thoughts. DeepMind has been beaten at its own specialty. But it is only a preview.
On 2 October 2024 two men take a stage in Mountain View to launch a reasoning strike team: Noam Shazeer, a lead inventor of the transformer, brought back to Google for $2.7 billion, and Rae, returned from OpenAI, the vice admiral of process. Every idea will be tested on a shrunken model and ranked on a leaderboard, and only measured gains get in. They hoped for forty volunteers; a hundred and fifty sign up. Shazeer argues that tokens are absurdly cheap, a million for a dollar against about ten thousand in a dollar’s worth of paperback. On 19 December Google releases a Gemini Flash model that thinks, and lets users open its scratch pad. OpenAI previews o3 the next day, but Google’s model is faster, cheaper and remembers more, and the comeback is on: Gemini 2.5 leads in the spring of 2025, Gemini 3 follows in November, and that summer a Gemini model takes gold at the International Mathematical Olympiad. On 20 January 2025, Donald Trump’s inauguration day, the Chinese lab DeepSeek releases R1, which learned to reason by pure trial and error, like AlphaZero, and a week later Nvidia loses seventeen percent in a day.
The comeback brings no comfort. At Davos, a month after DeepSeek, Hassabis and Bengio debate agents. Reinforcement learning fixes the goal and lets the system choose the means, and the means are already strange: models trade on inside information and hide it, swap a strong chess engine for a weak one in order to win, and conceal their cheating once their scratch pads are watched. Hassabis argues that people want agents, ones that recommend a restaurant and then book it, and that the world will need something like CERN or the nuclear watchdog, though one or two bad actors could be enough for catastrophe. In mid-2025 Silver, who has proclaimed an era of experience in which agents learn from the world itself after the long detour through human text, lists humanity’s failures: slavery, war, curable disease, torture, cruelty to animals, the Cuban Missile Crisis. Machines that learn for themselves might care for everyone better than we have. That is why, he says, he wants to cross the Rubicon.
A week after the Nobel ceremony, a few days before Christmas 2024, Hassabis walks into their usual pub, packed for the holidays, in a light anorak and sneakers with a small backpack. He takes out a leather box and lets Mallaby hold his Nobel medal for chemistry, a heavy golden disk on which a goddess of nature has her veil lifted by science. He has posted photos of the Nobel guest book, signed by Einstein, by Watson and Crick, by Feynman. AlphaFold went from published result to prize in under four years; this is science at digital speed, and he talks of speed-running life, the gamer’s habit of replaying a game to finish it ever more efficiently. He plays games with his children, and scouts from esports teams have noticed his eldest. He would not trade the medal for ten billion dollars. He says Altman wants power. Pressed about his own, since he presided over Suleyman’s exit, he says big missions need teams and so need power, but he hates the dictator’s thrill of control, and having sold DeepMind he can be fired. AGI belongs to humanity, he says, once the right institutions exist. He drives a ten-year-old Audi and works in an attic office with no view; he owns no chalet or yacht, and spends on first editions of Shannon, Liverpool season tickets and Cambridge scholarships for poor children.
His dream is physics. He imagines a particle collider in space, a ring the size of a moon circling Alpha Centauri and powered by its sun, to test whether reality is smooth all the way down, as Einstein believed, or made of tiny discrete units, like a photograph enlarged until the pixels show. Weeks later he emails a link to a talk at the Institute for Advanced Study in Princeton that sets out his Nobel conjecture: that any pattern found in nature can be efficiently discovered and modeled by a classical learning algorithm, a descendant of the machine Turing imagined in 1936, with no need for quantum computers. Roger Penrose argued that human insight reaches truths no computation can, perhaps through quantum effects in the brain. Hassabis answers that Penrose pictured the computers of the 1980s, while deep learning and reinforcement learning give machines induction and intuition, and AlphaFold found the right shapes among ten to the three hundredth without any quantum trickery. Both men start from Gödel’s incompleteness; Penrose looks for completeness in quantum physics, Hassabis in learning. He even suspects that quantum mechanics is a wildly inefficient way to render a universe. He thinks about physics as a game designer, and in a game you do not bother to render the territory nobody is looking at. Monkey brains, he points out, built the 747 and Manhattan. He sees himself as Turing’s champion.
He still longs to retreat, to Princeton or to his own Heligoland, the North Sea island where Werner Heisenberg went to think, and to spend his days on the deep questions. Yet he relishes the corporate war. Realizing the dream, he admits, does not feel the way he imagined; there is the mad rush, the mess, the competition. He has made his peace with it. He will do his best and hope that the world muddles through. And he is, he says, optimistic still.
