How Will AI Transform Human Life in the Next 20 Years?

It’s a beautiful spring day and six floors below us people are scurrying to work, worrying about mortgages, jobs, schools — their future. High above them in his office, the CEO of Google DeepMind is telling me how that future will unfold.

There is no one else in the UK like Demis Hassabis, 48. He is probably the only person in the country who can sit as equals with the billionaire tech bros of the US. He can also — unlike them — speak credibly to the scientists; he did, after all, just win the Nobel prize.

He advises government, is referenced by ministers, is arguably Britain’s only serious dog in the artificial intelligence race — but is also a dog very much owned, in the corporate sense, by the US.

And today, looking down on King’s Cross, he is an odd mix of Pollyanna and Cassandra.

What those people below don’t quite realise yet, going from home to work to home again, is just how much the AI revolution, he says, is “going to change everything”. There will be the good: in the next decade, perhaps, we could cure all diseases and “solve energy”.

Then there will be the admittedly more challenging: there will be nothing a human can do that a computer can’t do better. Except, perhaps, wrestling with what “humanity” is then for. “I think we will need some new philosophers,” Hassabis says. “This would be the perfect time for a new Kant to arrive.”

What would that philosopher philosophise on? “The question is about human meaning and purpose.” Hassabis’s life has lacked neither.

• Should we fear AI? The British scientist who says don’t panic

Almost 40 years ago Hassabis was the second-best chess player of his age in the world. He bought his first computer, which was funded from chess winnings, and taught himself how to program. Then he quit competing at chess — to apply his brain more usefully.

Home-schooled as a teenager, he completed his A-levels two years early at 16, and within a year he helped to program one of the most successful computer games of the 1990s. The games company Bullfrog had run a programming competition where the prize was to work for them. When Hassabis turned up, he wasn’t what they expected. “He looked about 12,” recounts Peter Molyneux, the Bullfrog founder, in The Thinking Game, a new documentary about DeepMind. “I thought, what the hell are we going to do with him?”

What they did with him was get him to write the code for a game called Theme Park, in which players designed rides and food concessions, before “all these little people came in”. These pixelated punters might then “leave in a huff”, Hassabis explains. “If your roller coaster had too many G-forces they’d come off, feel sick, throw up, and no one would buy from the food stands.”

His code is what made those little people happy, or sad, or vomity. The world loved it. Theme Park was a blockbuster.

Demis Hassabis in 1998 with his co-founder of the Elixir Studios video game developer, David Silver

Molyneux offered Hassabis a million pounds to stay. He turned it down. One explanation as to why is that he wanted to go to Cambridge University. Another is that he wanted to “solve” intelligence — and bring about a revolution he considers comparable to the discovery of fire. You can’t do that with a computer game.

How did he break it to his parents, a Greek Cypriot toy salesman and a Singaporean department store worker living beside the North Circular, that he’d turned down a million pounds?

“You know, I’m not sure I told them, actually. By that point they were already a bit confused and bemused by what I was doing. I still have the cheque somewhere. I found it about ten years ago, I should have framed it.”

After graduating from Cambridge and doing a PhD in neuroscience, Hassabis co-founded DeepMind with a friend from UCL, Shane Legg. Theirs was then a fringe view — that an AI with “general intelligence” (AGI) was coming, that it would apply its brain to any task, and do so better than us.

We had long been used to computers excelling at complicated things — no human is better at multiplication than a Casio calculator. But they wanted to make a general intelligence that was clever in the way a clever human is clever.

And then they wanted to make it clever in a way far beyond how a human is clever. We are still not there. “With today’s systems we can all see the flaws,” Hassabis says. “They are good at some things, but you push them a bit and they are clearly terrible.”

It took unusual investors to gamble on DeepMind, and one of them was Elon Musk. Hassabis liked him. “He’s an incredible, impressive person … Then I think he only had one factory. He was building one rocket. And now he’s got his empire.”

Does he still like him? “You know, I get on with most people. Given where AI’s got to go, how it’s going to affect all of humanity, I really think that’s important.”

This is the fourth time I’ve interviewed Hassabis. The first time was in 2018, when almost all his company had done was design programs that did well at games. I asked him when it would do something useful. He mentioned his aspiration to solve a problem in biology called protein folding. It was, arguably, the biggest longstanding challenge in the field. He also mentioned that within ten years an advance reliant on AI would win the Nobel prize.

Somehow it was clear that the recipient of the Nobel would be him. Somehow he said this without sounding like an arse. In the years since, as his predictions came true, he still hasn’t sounded like one. He is very “normal”. He doesn’t wear an intimidating suit or Steve Jobs-style poloneck.

Neither does he wear the grungy casual of the successful computer scientist, where the subtext is simultaneously “I don’t care about your human conventions” and “I don’t care about you”.

• Anthropic chief: ‘By next year, AI could be smarter than all humans’

DeepMind itself, made to a large extent in his image, treads the same line. It isn’t one of those US tech offices with folksy, overfriendly signage and smoothies on tap (although I suspect there are smoothies to be had).

On the walls there is memorabilia, chess boards signed by grandmasters, and reminders of past glories — framed covers from the many scientific journals that have published its advances. It feels exciting, in the way that working for Nasa in the 1960s must have been exciting. It should probably also feel more intimidating than it does.

“I’m quite British in my outlook,” he says. “I like to be ambitious, competitive. I think we can compete at the top table. But I like to do that quietly and with a bit of humility.”

The tale of DeepMind, and Hassabis, could have gone very differently. The early investors assumed the company would relocate to California. Hassabis, a north London boy who still lives in north London with his wife and two sons, pushed back.

“I like the values we have here and I think we’re a very innovative country,” he says. “Maybe we don’t translate it enough into commercial gains, but in terms of invention, innovation we punch well above our weight. Our universities are incredible.”

He also thought those university graduates were underserved. “We thought there was a lot of untapped talent … if you’re a PhD in physics from Cambridge and you don’t want to work in a hedge fund but you want to do something really intellectually interesting, there aren’t that many choices in the UK.”

So even after Google bought it outright for a reported £400 million, and even after DeepMind took over Google’s AI operations — the acquisition eating the acquirer — Hassabis stayed put. “AI is going to affect the whole world; it’s important there are other voices and cultures involved, not just 100 square miles of California.”

And thus it transpired that it was in a UK office that, in the mid-2010s, the entire staff of DeepMind fell into a slough of despond over their inability to win the world’s simplest computer game: Pong. A few pixels made up each bat. A smaller number of pixels made up the ball. Their opponent, in this computerised version of tennis, was the computer — a simply coded script designed to offer a mild single-player challenge.

Yet their own, considerably more sophisticated code could not beat it. “I remember for weeks, it felt like it might have even been months, we couldn’t get a point,” Hassabis said. Of course they could have written a program to beat it. But that wasn’t the goal — they wanted to write a program that learnt to beat it by itself.

They were using a niche kind of programming: agent-based reinforcement learning. It can trace its routes, in part, to Alan Turing — who in 1948 proposed a “pleasure-pain system” to train intelligent machines. At first the idea was their program moved randomly. If it did well, it was rewarded — and remembered. Then, like a dog given treats, it learnt through trial and error to do the right thing.

Their program was not a good dog. “We couldn’t even get the ball back,” Hassabis recalls. He began to wonder, was the idea a dud? Then one day the bat connected with the ball and they got a point. “And then, finally, we won a game.” Soon after it started winning 21-0 every time.

Technology is exponential and exponentials are deceptive. It can feel that not a lot happens before, suddenly, incremental gains multiply. Then, whoosh, suddenly you are winning not just at Pong, but chess and Go. And then you’re solving a 50-year-old challenge in biology …

Living creatures run using proteins — molecular machines that are behind everything that works in the body, and everything that doesn’t. We know that chemistry determines their shape. We also know the shape determines their function. But we cannot predict the shape from the chemistry.

Or, rather, we couldn’t until 2021 — when DeepMind turned “protein folding” into a game. Their program, not that dissimilar in principle from its Pong program, predicted and then published every protein in nature.

And still the exponential goes whoosh. Last year, a decade after defeating Pong and just before Hassabis shared the Nobel for the protein program with his colleague John Jumper, DeepMind’s robotics team designed a system that could win not at Pong but ping-pong — Pong in the real world.

Hassabis in 1983, when he was a chess prodigy

Even so, Hassabis thinks, sometimes people succumb to hype. Some of his competitors think we will have AGI while Trump is president. Him? He puts the chances in that period at less than even. He wants to dampen expectations.

“I think there are still quite a few things missing, in terms of creativity and inventive capability.” Something new is needed, he says, to make an AI that could, say, come up with the theory of general relativity. “Given the information Einstein had at the time, I consider that one of the greatest leaps of human imagination.”

So he is not so bullish. He thinks merely that within the next ten years AI may be good enough to cure all diseases. He has set up a company, Isomorphic, to do just that. The AI can tell us what the proteins that are causing disease look like. Maybe that AI can also tell us how to fix them — what drugs to use to return people to health.

“I know that sounds crazy today,” he says, “but ten years ago it would have been crazy for me to tell you we could fold all 200 million protein structures.”

But eventually, with all the inevitability of exponentials, he thinks AGI will come — if Google DeepMind doesn’t get us there, OpenAI or Grok or the Chinese will. Ideally he would like a body to oversee the advance, like the International Atomic Energy Agency does for nuclear. But it probably won’t happen.

And once that comes, he says, “we will be in a new era”. Parts of that new era, I tell him, feel bleak. It can feel like an era without human agency or purpose.

What does Hassabis tellhis two sons? Do they bother with school? They very much do, he says. “For the next ten years I think there is going to be an amazing flourishing of creativity, empowered by AI. For those who understand and are native with these tools, it’s going to be a sort of superpower.

“Then there’s the question of what happens after that. That’ll be the next era, as AGI gets better and better. That is more difficult to predict.” It will not be a world, he thinks, of work as we understand it, but it will bring “radical abundance” and, perhaps, radical human obsolescence.

Maybe, he suggests, some people will find meaning elsewhere — through extreme sports, say. Maybe we will fuse our brains with computers, like Musk believes. Maybe we will explore the stars — as his friend also believes. “If you have radical abundance, you’re freed from resource constraints. It could be maximum human flourishing.” It is, he says, risky but exciting. “We’re extremely adaptable as a species.”

All the rooms in Google DeepMind’s headquarters are named after scientists. Hassabis’s office is the Leonardo da Vinci room. Next to it is the Alan Turing room. These two scientists are his favourites. Turing because “obviously”. Leonardo because “he’s the last all-rounder”.

What Hassabis envisions is a change of such magnitude that the world of today would be more recognisable to someone from Leonardo’s time than the world of 2050 will be to us. It feels dizzying. Afterwards, when I rejoin the scurriers in the street below, it feels something else: silly.

Then I think that, from above, we must look like those little pixelated people he created all those years ago, with their wants, desires and occasional vomiting. And then all I can think is: he has been right so far.
The Thinking Game previews March 17 across Picturehouse and Curzon cinemas, then UK-wide March 21, thinkinggamefilm.com