Demis Hassabi- AI's Deepest Mind Believes the World Is About to Change Radically

YOU CAN LEARN a lot about the people racing to build the world’s most powerful technology by asking what keeps them up at night.

In the case of Demis Hassabis, it’s not just the complexities of AI. He’s wide awake when he should be asleep because the middle of the night also happens to be his favorite time to work. In the small hours of the morning, his brain is loose, the world is silent, and he can read, think deeply and really focus on the problems he’s spent a lifetime trying to solve.

Ever since he was a boy, Hassabis dreamed of building a form of artificial intelligence that would be smart enough to crack the mysteries of the universe. When he started the research lab that became Google DeepMind, he turned his life’s obsession into the mission of his company.

“The ultimate goal was to build AI and understand the mind, then use it to understand science,” Hassabis tells me. “Step one, solve intelligence. Step two, solve everything else.”

AI can already write code and rewrite emails and summarize meetings and prove abstract math conjectures and figure out why your washing machine keeps making that strange beeping noise. But in Hassabis’s mind, the whole point of intelligence is to do something with it. What he wants to do with it is make the kind of discoveries that actually change our lives.

The clarity of his ambition makes him an outlier among the titans of his industry. Then again, Hassabis has always stood apart. Now 50 years old, he lives and works in England, where he was born, raised and knighted, and the distance between Sir Demis Hassabis and Silicon Valley gives him a different perspective altogether. “I feel like it helps me think original thoughts,” he says from London.

In this moment of AI anxiety, his most original idea is offering a vision of the future that doesn’t make us want to pull the plug on every data center. He understands the profound risks of this technology, but he feels it’s possible to address them without overhyping them—“It’s not all disaster, and it’s not all nirvana”—and he finds it irresponsible to bring up the problems without trying to solve them.

“I’m very much in the solutions business,” he says. “If there’s a problem, how do we solve that with collective human ingenuity? And if you don’t have a solution, then you probably need to work harder.”

These days and nights, he’s taking on a new role after a recent leadership shuffle at Google DeepMind, but he’s working as hard as ever on artificial general intelligence. He thinks of AGI as a system with the full cognitive range of the human brain—not just some of our capabilities but all of them. Earlier this year, he felt AGI was still three to five years away. Now he expects AGI within two to three years.

“It’s soon,” he tells me. “I’m updating nearer and nearer.”

The people who know him best tend to use the same words to describe Hassabis: brilliant, relentless, so ferociously competitive that he can’t stand losing a game of foosball. So as it gets nearer and nearer, I ask someone who has spent almost his entire life preparing for this moment what it will take to win.

“Winning,” he says, “means helping humanity steward AGI into the world safely and securely for the benefit of everyone. I think that’s the most important thing I can do over the next few years.”

HASSABIS FIRST started thinking about thinking as a 4-year-old chess prodigy who needed to sit on a phone book just to see the board. By the time he was 12, he was one of the best chess players in the world for his age. But he suspected there was more to life than moving “inanimate lumps of plastic.” With earnings from chess tournaments, he bought his first computer.

He turned out to be just as talented at programming games as he was at playing them. When he was 16, he took a gap year before college to work at a videogame production company and experimented with primitive forms of AI. But he already had reason to believe AI would reshape the world.

“Not to ascribe too much to my childhood,” he says, “but one of the reasons I got into this whole field is that I read way too much sci-fi.”

His boss offered Hassabis a million pounds to skip college, but he turned down the money and went to Cambridge. At the university that produced Newton and Turing, he studied computer science—and more games. He taught himself Go, a board game older and vastly more complex than chess. He also went to the pub and mastered a simpler game: foosball. “I was Cambridge champion,” he tells me. He later set up a foosball table at work and dared colleagues to beat him, according to The Infinity Machine, Sebastian Mallaby’s biography of Hassabis. When he finally lost, he stood up and declared: “It’s like my soul is on fire!”

In 2010, after earning a Ph.D. in neuroscience and deepening his study of the mind, he started DeepMind. When he set out to apply his knowledge of human intelligence to the artificial kind, it was only natural to start with games.

At a company that he envisioned as a cross between Bell Labs and the Manhattan Project, the research began by teaching AI to play Pong. For months that felt more like an eternity, the AI didn’t win a point. But once it started winning, it couldn’t stop. The same thing happened in Breakout. After 100 games, the AI wasn’t very good. After 300, it was as good as any human. After 500, it was superhuman. When their study of AI conquering Atari games was published, Hassabis stood up at a DeepMind meeting and toasted what they all thought was a huge accomplishment.

By the way, he added, one day we’re going to win a Nobel Prize with this technology.

Google bought DeepMind in a $650 million deal in 2014, and the lab had its breakthrough in 2016, when its revolutionary system AlphaGo beat a top-ranked Go player in a series of matches that stunned humans around the world. The highlight of AlphaGo’s victory was a move so counterintuitive and terrifyingly creative that Move 37 became a watershed moment in AI history.

When it came time for his next challenge, a scientific puzzle that had fascinated Hassabis ever since he learned about it from a biologist friend at a Cambridge foosball table was at the top of his list: the protein-folding problem.

In this one problem, Hassabis saw an opportunity to find solutions to a whole lot more, since predicting the shapes of proteins—“unbelievably exquisite things,” he calls them—held the promise of understanding diseases and curing them. When DeepMind’s AlphaFold system effectively solved this problem in 2020, it was AI’s most impressive feat yet. But it wasn’t enough to satisfy Hassabis. In a meeting to discuss what else they could do, his colleague John Jumper mentioned that they could predict the structures of millions and millions of proteins in a month, and—

“Why don’t we just do that?” Hassabis said. “We should just do that.”

So they did. Before long, they had predicted the structure of nearly all known proteins—more than 200 million in total—and released them to scientists free of charge. And in 2024, he nailed one more prediction.

For their pioneering work on AlphaFold, Hassabis and Jumper won the Nobel Prize in chemistry.

A DECADE AGO, before the Nobel, before his lifetime of interests came together like a protein taking shape, so long ago that he still had a full head of hair, Hassabis took a long walk with a researcher he was trying to hire.

“I was expecting to talk to a CEO,” says Pushmeet Kohli, now DeepMind’s vice president of science. “I left with the impression that I was talking to a scientist.”

From the company’s early days, Hassabis’s priority was using AI to advance science and improve human health. If step one is solving intelligence, he was always more interested in step two. “You have to have the tool first, and building that tool is fascinating,” he says, “but it’s even more fascinating what you can do with those tools.”

But in America, the tools are deeply unpopular. This bothers Hassabis, who sees it as another problem to solve. For years, AI leaders have teased the possibility of curing cancer. To change people’s minds about AI, he says it’s time to actually cure cancer.

“Then,” Hassabis says, “I think people could really start feeling the benefits of AI directly in their lives.”

Does he feel that’s possible?

“I don’t see any insurmountable obstacles currently,” he says. “I see hard things, but not impossible things.”

To overcome them, he’s in the process of reorienting his own life. After running Google DeepMind, he recently took on the job of Google parent Alphabet’s chief scientist and DeepMind’s chair. For his next act, he’s shifting his attention to accelerating scientific progress even more than he already has.

A few years ago, after the AlphaFold milestone, Hassabis launched a company inside Google called Isomorphic Labs, a nod to the mathematical term describing objects with the same form, shape or structure.

As the CEO of Isomorphic, he isn’t satisfied by simple progress. When the DeepMind team entered a competition for predicting the shape of proteins, for example, he didn’t just want to win. He wanted to make the entire premise of the competition obsolete. He wanted to beat the game. He still does.

“We’re not just trying to solve one disease,” he says. “We’re trying to create many models, and then a whole process, that can revolutionize the drug discovery process itself.”

To him, the difficulty of the problem is precisely what makes it worth tackling. When I ask if he expected AI-designed drugs to take so long to be discovered, tested and turned into medicine, he says of course he did. Then he sounds less like a CEO than a scientist in awe of his subject.

“It’s just sort of mind-blowing that it works at all,” he says. “That’s what draws me to biology—and life. It’s the most complex system we know of.”

If only it were as simple as a board game with more positions than the number of atoms in the universe. But he knows how it feels to watch progress unfold.

“I’ve seen this before,” Hassabis says. “It’s slow, slow, slow—then suddenly all at once.”

As we discussed the possibility of AI saving lives, the rest of the world was more focused on AI killing us all. We happened to be speaking days after one Anthropic researcher put an alarmingly high probability of human extinction in the next decade. My conversations with Hassabis and people around him kept returning to what else could happen in the next decade. “I’d be disappointed if we haven’t solved many diseases,” says Max Jaderberg, Isomorphic’s president.

Meanwhile, there’s the small matter of building the smartest machine ever made.

Google was the early leader in AI and brought the technology into the lives of billions of people, but now it’s trying to keep up with OpenAI and Anthropic. Hassabis, though, is looking beyond today’s models to general intelligence, which he says will be more consistent and capable of original thought. To him, the G in AGI matters as much as the AI itself.

He’ll know AGI is here when it passes what he calls the Einstein test: Could a system trained only on human knowledge up to 1901 produce Einstein’s work from the annus mirabilis of 1905?

We’re still a few years from AI that can invent modern physics, he says, but that’s sooner than he would have predicted when he won the Nobel and inscribed his name in a book signed by his scientific heroes: Feynman, Curie, Bohr, Planck, Crick—and Einstein.

NOW BACK to the part where everyone dies.

For all the scientific upside of AI, there’s also the very real possibility of apocalyptic downside. After years of safety warnings from doomers, even the leaders of frontier AI labs came together in recent months and called for a slowdown. By then, Demis Hassabis had already sketched out a design for a brake system.

This summer, he spent his nights agonizing over every word of a proposal for developing the technology responsibly. In his essay, Hassabis wrote that he considers AI miraculous (“we’ve essentially found a way to make sand think”) and so consequential that it’s more like fire or electricity than the internet. To make sure it doesn’t burn down civilization, he pitched an industry-funded body that would operate under federal oversight and test frontier models before they’re released. His plan was meant to be adaptable, updated in real time and ratcheted up if necessary, nimble enough to keep pace with the technology it’s trying to control.

After all, yet another risk of AI improving so rapidly is that anyone drafting regulations ends up writing rules for a world that no longer exists.

“What would be wrong is to try and regulate things now that we don’t even fully understand,” Hassabis says. “I think you’re almost certainly going to get the wrong regulation—and wrong regulation is worse than no regulation.”

His proposal was praised by archenemies Sam Altman and Elon Musk, which I believe qualifies him for a Nobel Peace Prize. When we spoke, it was unclear how—or if—the industry would be regulated. Hassabis says the leaders are in regular conversation about what they see inside the labs, despite the vicious personal rivalries of the humans leading the most valuable AI companies.

Do the AI CEOs have a group chat?

“No,” he says. “I mean, I speak to everybody. But some of them don’t get on with each other.”

NOT LONG AGO, Demis Hassabis appeared onstage at a developer conference and managed to give a keynote speech that was actually memorable.

“When we look back at this time,” he said, “I think we’ll realize that we were standing in the foothills of the singularity.”

Until now, Hassabis has been focused on everything that must happen before the advent of AGI.

“Not enough time has been put into what should happen after,” he says. “What do we want to value? How is this going to affect the human condition and our institutions? How’s everyone going to benefit from this? What is the purpose going to be?”

He goes on: “I believe we can have an amazing future. But we have to intentionally build toward it.”

Building that future and preparing society for its arrival doesn’t leave him time for much else. Does he read? “A lot, but mostly work books or science books. It’s been a while since I’ve read something for pleasure.” What about TV? He liked the first season of True Detective, which aired in 2014, and Chernobyl, which came out before the pandemic. And his gaming? Also work. He likes to test AI agents by re-creating the games he designed as a teenager and vibecoding new ones.

There are only so many hours in a day, which is why he’s awake late into the night. When he’s in a flow state at 3 or 4 a.m., he looks at his morning schedule and decides if he can handle it on less sleep. “If I can, and if things are going well creatively, I’ll just stay up,” he says. He aims for six hours but often settles for less. “It’s not good for your brain,” he says. “I know that, having studied the brain.”

I ask Hassabis if he ever lets himself imagine a day when all of those sleepless nights pay off, whether it’s the end of a Go match or the end of a disease.

“When you’re in the midst of it, it’s so hard to make any progress that you don’t really worry about the top of the mountain,” he says. “You’re just focusing on the next foothold and not falling off the mountain.”

So while he’s still in the foothills, I ask him to picture the summit. He’s solved intelligence. He’s on to everything else. It’s the middle of the night. The first sign of daylight is hours away. He pours himself a cup of Yorkshire tea and walks up to the attic, just another human trying to figure out life in the age of AGI. What’s on his mind?

“I want to use AGI to understand the nature of reality,” he says. “I’m hoping that I’ll be sitting there with my AGI systems and some great human scientists, trying to work out the next level of physics. That’s probably one of the first things.”

But it’s only the first thing.

“There’s a lot of work to do,” he says. “Even if I get all my nights back for thinking, it’s still not going to be enough.”