Ambition

“Solve intelligence, and then use that to solve everything else.” — DeepMind’s founding mission statement

The boy was four years old when he learned chess by watching his father play, and within a few years he was beating grown men across a board. By thirteen he had reached master standard, with a rating that put him among the strongest players in the world for his age, and he had noticed something about the game that would organize the next forty years of his life. Chess was hard. It was hard in a particular way, a way that seemed to ask the whole machinery of the mind to engage at once: memory and pattern and planning and the cold reading of an opponent’s intentions, all running together under the clock. Demis Hassabis, growing up in north London in the 1980s, the son of a Greek Cypriot father and a Singaporean Chinese mother, was the kind of child who, having mastered the game, began to wonder less about how to win and more about the apparatus that was doing the winning. Given his first computer at the age of eight, he taught it to play another board game, Othello, and watched it beat opponents using rules he had written down himself. What was the mind doing, exactly, when it played? And could a machine be made to do the same?

That question never left him, and it explains a career that otherwise looks like a series of swerves. Hassabis did not take the chess prodigy’s usual road into a grandmaster’s life. He taught himself to program on home computers, finished school two years early, and won a place at Bullfrog Productions, the Guildford studio run by Peter Molyneux that was turning out some of the most inventive games in the world, after entering a competition the studio ran. There, still a teenager, he co-designed and did much of the programming for Theme Park, a 1994 simulation game in which the player builds and runs an amusement park and watches a small artificial world respond, its tiny visitors growing happy or sick or bored according to rules the designer set in motion. The game sold in the millions. It was, in its way, his first serious encounter with the thing that would consume him: a system complex enough that its behavior surprised the person who built it.

A games career followed, the kind that would have been the whole story for almost anyone else. Hassabis read computer science at Cambridge, took a double first, then went back to the industry and founded his own studio, Elixir Studios, in London in 1998. Elixir made ambitious, idea-heavy games. Republic: The Revolution, released in 2003, tried to simulate the politics of an entire fictional country; Evil Genius, the year after, let the player run a James Bond villain’s lair. The games were clever and reviewed unevenly, and in 2005, after a larger project was cancelled, the studio closed. To a casual observer this was a talented man whose first act had ended without a clear second.

What the casual observer missed was that Hassabis had never actually been interested in games. He had been interested in intelligence, and games were the most tractable place he could find to study it. So in his late twenties he did the thing that looks, on a résumé, like a non sequitur and was in fact the most direct possible move toward his real subject: he went back to school to study the brain. He took a PhD in cognitive neuroscience at University College London, finishing in 2009, and he did not work on anything as practical as vision or motor control. He worked on memory and imagination, on the hippocampus, on what happens in the brain when a person remembers the past or envisions a future that has not occurred. His research drew a connection between the two that drew real attention: the same neural systems that let you recall where you were last Tuesday let you construct a scene you have never witnessed. Remembering and imagining, in the brain, are close kin. For a man who wanted to build a mind, this was not an idle finding. It was a clue about what a mind is for.

The neuroscience was not a detour from the games and it was not a detour from what came next. It was the middle term in a single argument Hassabis had been making to himself since he was a child at the chessboard. If you want to build general intelligence, the argument went, you should first understand the one example of it that already exists, which is the brain, and you should test your understanding on the kind of problems that exercise the whole of it, which are games. The line ran straight from Theme Park through the hippocampus to the company he was about to start. Almost no one else in technology was thinking this way. The people building artificial intelligence in 2009 were, with few exceptions, building narrow tools: a system to recognize faces, a system to rank web pages, a system to read handwritten digits. The phrase “general intelligence” was not something a serious person put in a business plan.

That was the gap Hassabis walked into. In September 2010, in London, he founded a company with two co-founders and an objective so large it sounded like a joke. Its mission, as the company stated it, was to “solve intelligence, and then use that to solve everything else.” The name was DeepMind.

The two co-founders mattered as much as the premise. One was Shane Legg, a New Zealand-born researcher Hassabis had met during his time around UCL, who had done his doctoral work on the mathematics of machine intelligence under Marcus Hutter and who had been willing, when it was deeply unfashionable, to use the phrase “artificial general intelligence” in the title of his thesis and to argue that it was a coherent and reachable goal. Legg was the company’s intellectual conscience on the question of what they were actually trying to build, and he had spent years thinking rigorously about a thing most academics would not say aloud without a nervous laugh. If Hassabis supplied the conviction that general intelligence was the right target, Legg supplied the case that it was a target at all, something you could define and measure and march toward rather than a science-fiction word. The third founder was Mustafa Suleyman, who had not come up through academia or games but through social enterprise and conflict resolution, and who took on the operational and human side of the company, including, from early on, the question of what it would mean to be responsible while building something this powerful.

The premise was the hard part to sell. “Solve intelligence” is not a deliverable. It does not fit on a product roadmap, and in 2010 it did not fit anywhere in the ordinary financing of a startup, where investors want to know what you will ship and when. DeepMind would not ship anything for years. Its plan was to do fundamental research toward a goal that might take decades and might not arrive at all, and to do it as a private company rather than a university lab. By the conventional logic of venture capital this was close to unfundable. And yet it got funded, by a small group of backers who were betting less on a product than on a thesis and the people holding it. Peter Thiel’s Founders Fund put money in; so did the Hong Kong investor Li Ka-shing’s Horizons Ventures and others on the unconventional edge of the investing world, people comfortable with the idea that the largest returns come from the bets that sound absurd at the time. Hassabis had to convince them of something other than a quick profit: that intelligence itself was an engineering problem, and that his particular blend of neuroscience and games and machine learning was the way to crack it.

The proof he offered, once the company had grown enough to offer one, was a machine that learned to play. DeepMind’s early research bet was on reinforcement learning, the branch of machine learning concerned with an agent acting in a world and learning from the consequences, the way an animal learns: try something, see what happens, do more of what worked. It learns by acting rather than by sorting through a fixed pile of data. Hassabis’s instinct, the one that traced back to the chessboard, was that games were the perfect training ground, because a game is a world small enough to fit inside a computer and rich enough to demand real intelligence to master. So DeepMind’s researchers pointed their methods at the most humbling possible target: the Atari 2600, the home video-game console of the early 1980s, and its catalog of simple arcade games.

The system they built, which they called the Deep Q-Network, did something that had not been done before. It learned to play dozens of different Atari games, from Breakout to Space Invaders, from nothing but the raw pixels on the screen and the score. No one told it the rules. No one told it what a paddle was, or a ball, or that knocking out bricks was good. It saw the same thing a human player sees, a grid of colored dots changing over time, and a single number going up or down, and it learned, by playing the game millions of times, which sequences of joystick movements made the number go up. On some games it reached and then exceeded the level of a skilled human. In Breakout it discovered, on its own, a strategy that good human players know but that no one had programmed into it: tunnel a hole through the side of the wall and loft the ball behind it, where it ricochets along the top and clears bricks without further effort. The machine had not been taught the trick. It had found it, the way Hassabis had once found things at the chessboard, by playing until the structure of the game revealed itself.

The first version of this work appeared as a paper at the end of 2013, and a fuller account was published in the journal Nature in early 2015. To most of the public it was a curiosity, a computer good at old video games. To the small number of people who understood what they were looking at, it was something else: a single learning system, general enough to master many different games it had never seen, learning from raw perception and a reward signal, with no hand-built knowledge of any particular game. That was the shape of the thing DeepMind had promised to build, in miniature. It was very far from general intelligence. But it was a demonstration that the path Hassabis had described was not empty, that you could point one learning method at a wide class of problems and watch it teach itself, and that was enough to change who was paying attention.

Among those paying attention was Google. By 2013 the largest technology companies had begun to understand, in the way the prologue of this story describes, that the people who could make neural networks work were suddenly the most valuable talent on earth, that they were scarce, and that owning them was now a strategic necessity rather than a research luxury. DeepMind had assembled an unusual concentration of that talent in London, around a mission that promised more than any product. There was reportedly more than one suitor; the names of Facebook and others have been attached to the courtship. The company Hassabis chose to sell to was Google, which acquired DeepMind early in 2014.

The price has never been stated with precision in public, and the reports vary, which is worth being honest about. Figures cited at the time ran from around $400 million to upward of $600 million; a sum near £400 million, or roughly $500 million to $650 million depending on the exchange rate and the source, is the range most often given. Whatever the exact number, the salient fact is its size measured against what DeepMind was. This was a company with no product, no revenue, and no near-term plan to have either, founded less than four years earlier, whose central asset was a few dozen researchers and a thesis about the future. Google paid for it on roughly the scale of a substantial product acquisition, and it did so because it had come to believe that the thesis might be real, that general intelligence was a thing one might actually build, and that it would rather own the leading attempt than watch a rival own it. The audacity that had made DeepMind nearly unfundable in 2010 had become, by 2014, the precise thing that made it worth hundreds of millions of dollars.

Hassabis negotiated a condition that revealed how seriously he and his co-founders took their own premise. As part of the deal, Google agreed to establish an internal ethics and safety board to oversee the use of DeepMind’s technology, and DeepMind retained an unusual degree of autonomy, staying in London rather than being absorbed into the Mountain View machine. The details of the ethics board were closely held and its workings were never made fully public, which became its own small controversy in later years. But the impulse behind it was clear and, at the time, almost eccentric. Here was a founder selling his company for a fortune who spent part of his leverage on a constraint, rather than on a higher price: a limit on what the buyer could do with what it was buying. The people building DeepMind believed they were working toward something powerful enough that it would need governing, and they wanted that belief written into the contract before the ink was dry. That instinct, that the technology was dangerous in proportion to how well it worked, would become one of the defining tensions of everything that followed, splitting labs and founders apart in the decade to come. It entered Big Tech, quietly, as a clause in an acquisition agreement.

What Google had actually purchased was harder to name than a product or a patent portfolio. It had purchased a worldview, and the people who held it most coherently. The conviction at the center of DeepMind was that artificial general intelligence, a single system that could learn to do more or less anything a human mind can do, was an achievable engineering goal rather than a distant fantasy, possibly within the careers of the people then working on it, and that whoever built it first would hold something of incalculable consequence. In 2010 that conviction had been rare enough to make the company hard to finance. By putting hundreds of millions behind it, Google made the conviction respectable. It signaled to the rest of the industry that “solve intelligence” was a sentence a serious company could organize itself around, that AGI was now a fundable corporate objective rather than a thing one muttered at conferences. Within two years that signal would help call into being an organization founded expressly as a counterweight to it, on the premise that this goal was too important to leave to a single search-and-advertising giant. The race that defines the rest of this book was, in part, a response to the bet Google placed on a London lab in 2014.

The man at the center of it kept his eyes on the original question. The point of DeepMind, Hassabis maintained, was never the games, just as the point of the chessboard had never been the trophies. The games were a proving ground, the neuroscience a source of clues, the company a vehicle. The actual goal was to understand intelligence well enough to rebuild it, and then to aim that rebuilt intelligence at the problems that had defeated the unaided human mind, disease and climate and the basic riddles of biology and physics. It was a goal grand to the point of grandiosity, and it would have sounded like hubris from almost anyone. From Hassabis it sounded like a research plan, because he had been working toward it, in one disguise or another, since he was a child counting the moves ahead at a board.

That was where the founding cast of the modern story came to rest. The idea that had been buried in 1969 and kept alive in basements and exile had done more than survive. It had become the engine of two opposing strategies inside the largest companies on earth. On one side stood the compute-and-talent machine that the previous chapters described, an industrial apparatus for turning data and processors into working systems. On the other stood a London lab that treated intelligence itself as the product and the prize. Both now belonged, in some measure, to Google, and the question of who would own intelligence, and what they would do once they did, was about to stop being a matter of academic factions and become a matter of money, and a great deal of it. The first move in that contest came from a chief executive, not a researcher, who decided to pick up the phone himself.