British DeepMind founder and CEO Demis Hassabis remains a humble figure in the world of brash tech giants.
In March 2016, a hushed silence settled on the 21st floor of a Four Seasons hotel in Seoul. Hundreds of people had gathered in the South Korean capital, alongside millions more via a global live stream, to witness an epic battle showcasing the evolution of artificial intelligence.
On one side of a wooden board sat Lee Sedol, the local hero with a home advantage. Sedol, then 33, was one of the world’s greatest masters of the fiendishly complex game of Go. This board game, played with black and white stones on a 19-by-19 grid, emerged in China over 2500 years ago.
On the other side was a man whose moves were being dictated by a computer screen. He served merely as the assistant to AlphaGo, an artificial learning machine devised by scientists at DeepMind, a London AI startup, with one simple mission: beat any human at Go.
For six days, Sedol and AlphaGo traded blows in an event even more consequential than the clashes in the late 1990s between IBM’s Deep Blue supercomputer and the world chess champion Garry Kasparov.
In computational terms, Go makes chess look like child’s play. In chess, there are around 35 possible moves at each turn. In Go there are 250, and each of those leads to 250 more, giving billions of possible positions. So while Deep Blue relied on a “brute force” approach to anticipate gameplay, this would not work on a Go board.
To beat Sedol, AlphaGo would have to think like him, mimicking human intuition and mining the database of millions of human moves it had been trained on. Few people thought the machine could win, and Sedol was looking forward to a 5-0 victory.
Meanwhile, the real brains behind Alphago stood nervously in a separate hotel room that had been set up as DeepMind’s mission control. As multiple screens displayed data and graphs connected to a vast array of computing power, Demis Hassabis, who is 47, watched Sedol’s face, occasionally breathing heavily and scratching his head.
Hassabis, an unassuming figure in a world of brash tech titans, had a lot riding on the match. Two years earlier, he had sold his startup to Google for a reported USD 625 million, earning an undisclosed fortune and an uncomfortable level of fame for a man of his natural reticence.
Google had swooped based on the promise of an AI revolution and Hassabis’s talents. So the Go match was more than a stunt: it was a proof of concept. If a machine could defeat a human at Go, what else might it achieve?
Despite his modesty, Hassabis described the work of DeepMind as an “Apollo programme for the 21st century”, such was the scale of its ambition and assembled talent. But what he might not have fully grasped in 2016 were the consequences of victory, and the coming era of existential anxiety about the machines he was helping to build.
Demis Hassabis was born in North London in 1976 to a Chinese-Singaporean mother who worked in a department store, and a father of Greek-Cypriot descent, who “did lots of different things,” including writing and singing songs. Neither were at all technical.
He started playing chess at age four, developing a natural talent for strategy and winning. A chess master by the age of 13, he competed regularly in the Mind Sports Olympiad, whose organisers described him as “probably the best games player in history”.
Hassabis used his first chess prize-winnings to buy a ZX Spectrum, an early British home computer. “I had this feeling that this was a kind of magical device,” he recalled. “Looking back on it now, just as if you think of a car as a machine that amplifies human capabilities physically, the computer to me felt like that for the mind. I think I intuitively understood that as a child.”
From age eight, Hassabis was buying books and teaching himself to program. He went to the local school, where learning came easy. Completing his final school exams early, he was told to take a year off before starting a computer science degree at Cambridge University. So he joined a computer games company, Bullfrog Productions, working on titles that sold millions of copies, including Theme Park, one of the first video games to incorporate AI.
At Cambridge, where Hassabis graduated in 1997, he was often frustrated, particularly on the subject of AI. He thought his tutors were too focussed on a narrow version of the fledgling technology. “I remember distinctly one lecture, where I said to my friends around me, ‘We shouldn’t listen to this, they’re brainwashing us.’” Hassabis recalled in 2015. “I said that slightly too loud and the lecturer called me out and said, If you think you know everything, you shouldn’t come here.”
But he still graduated with a top degree and returned to the games industry, launching his own video games company, Elixir Studios, in 1998. Its ambitious titles involved vast worlds guided by AI.
Hassabis, who describes himself as easily bored, is a natural polymath who enjoys mastery in multiple fields. But it was the brain that fascinated him most, and he soon returned to academia, first at University College London (UCL), where he gained a PhD in cognitive neuroscience in 2009.
His research papers on memory received widespread media coverage, and he became a visiting scientist at the Massachusetts Institute of Technology (MIT) and Harvard University. But Hassabis was playing a long game. Throughout all his academic work, he was seeking inspiration from the workings of the brain for the AI algorithms he dreamed of creating.
In 2010, Hassabis launched DeepMind in London with fellow AI geniuses and entrepreneurs Shane Legg and Mustafa Suleyman. They wanted to combine the latest neuroscience research with advances in machine learning and computing power to create learning algorithms that would lead to artificial general intelligence. Their mission: to "solve intelligence" and then use intelligence "to solve everything else”.
Hassabis continued to use games as a vehicle for developing, testing, and demonstrating progress in what was already an arms race among the major tech firms to realise AI’s full potential.
In 2013, DeepMind published the results of its first landmark work. It had trained an increasingly smart system to play seven classic games made by Atari, including Space Invaders, Boxing and Pong. The machine received no instructions other than to keep getting higher scores – and it aced every game.
Soon after publication of the test results, Google made its big move - and its biggest acquisition of a European startup. In fact, Google was not the first major tech power to bankroll DeepMind, such were the talents of its principal founder. But there was already a hint from an earlier backer of the unease building around the AI arms race.
Elon Musk had first met Hassabis at Musk’s SpaceX rocket factory, outside Los Angeles, where each man argued that he was developing the most important project in the world: interplanetary colonisation for Musk, artificial super-intelligence for Hassabis. Musk argued that we might need to escape to Mars if AI were to destroy humanity. Hassabis suggested AI would simply hitch a ride and continue its march.
The men were joking, but Musk was becoming concerned about the consequences of AI’s rapid advance. In 2014, he said he had bought shares in DeepMind, before Google’s purchase, not to make money but “to just keep an eye on what's going on with artificial intelligence. I think there is potentially a dangerous outcome there… And we should try to make sure the outcomes are good, not bad.”
Musk later claimed he had tried to outbid Google for DeepMind. Then in 2015, he co-founded OpenAI, the research lab and startup that would later launch the ChatGPT language model (Musk left OpenAI in 2018). He was not alone in his concern. In 2014, the late British physicist Stephen Hawking said that AI would be “the biggest event in human history . . . unfortunately, it might also be the last.”
In the months after the Google buyout, Hassabis was anxious to reduce the sense of threat, and highlight AI’s benefits. “At the moment, these are science fiction stories,” he told the Financial Times in 2015. “Yes, there’s no doubt that AI is going to be a hugely powerful technology. That’s why I work on it. It has the power to provide incredible advances for humanity.” He added: “In our research programme, there isn’t anything that says program consciousness. ”
But Hassabis was not naive. He had agreed to sell his company to Google on condition that it remained independent and non-commercial, and he later pushed for a separate legal status with an independent governance board. Notably, as well as comparing DeepMind with the Apollo space program, he has also compared it with the Manhattan Project, the World War II mission to make the first nuclear weapons.
There was a sense, heightened by the huge success of ChatGPT in 2022, that the genie was out of the bottle. In April this year, Google, reportedly blindsided by OpenAI’s coup, announced that DeepMind would merge with Google Brain, its own AI team. Hassabis lost his independence, but in return gained greater influence as the leader of the new Google DeepMind unit.
The move coincided with more warnings of AI’s potential threat, much of it from technology innovators and promoters, with widespread calls to halt its progress. By now Hassabis was keener to engage with such concerns. In May 2023, he put his name to a statement issued by the Center for AI Safety that said: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.”
In the end, Lee Sedol only won a single game against AlphaGo, recording a 4-1 defeat that left him as stunned as everyone else. “It made me question human creativity,” he said in a post-match press conference. Three years later, he retired from the game, saying: "I'm not at the top even if I become the number one… There is an entity that cannot be defeated."
It was a momentous victory, but Hassabis was characteristically reserved. He said that Sedol had exposed weaknesses in AlphaGo. Above all, he wanted to use the attention gained by the clash to tell the world that the algorithms used for AlphaGo “one day can be used in all sorts of problems, from health care to science.”
That day came quickly. In November 2020, DeepMind revealed AlphaFold, a project using AI to predict the structures of proteins, and making them freely available to science. In July 2022, DeepMind announced it had deciphered the structure of almost every known protein, opening the door to the creation of new technologies and treatments to tackle challenges like sustainability, food insecurity, and disease. Researchers at the University of Oxford were already using it to develop new malaria treatments.
Then in June 2023, in response to OpenAI’s ChatGPT triumph, Hassabis announced that DeepMind was working to create Gemini, its own language model. He promised it would be enhanced by the work on AlphaGo, and offer “some new innovations that are going to be pretty interesting.”
More than 30 years after he first got his hands on a computer, Hassabis is at a critical juncture in a long career marked by a rare combination of modesty, laser-like determination, and a string of startling successes.
Even after the sale of his baby to Google had made him a multi-millionaire, he continued to live a relatively ordinary life. In 2016, he was still taking the train home after work to his house in North London, shared with wife Teresa, an Italian molecular biologist, and their two sons.
Apart from slightly more expensive glasses and jackets, he has resisted a Silicon Valley makeover. He still looks like the genius geek from North London who was equally at home on a chessboard or exploring the limits of his ZX Spectrum.
As anxiety has mounted about the potential of AI to upend everything we know, Hassabis has used his quiet wisdom to walk a fine tightrope, seeking to reassure us that the brains of the AI architects are big enough to contain the machines they are creating.
And while he signed the AI safety statement, he was also keen to resist calls for a pause in AI research. Doing so would be impractical, and would also threaten the increasingly evident benefits of the technology. “If done correctly, it will be the most beneficial technology for humanity ever,” he said of AI earlier this year. “We’ve got to boldly and bravely go after those things.”