AI's superhero: Can this genius keep it in check?

Demis Hassabis has a modest demeanour and an unassuming countenance, but he is deadly serious when he tells me he is on a mission to “solve intelligence, and then use that to solve everything else”. Coming from almost anyone else, the statement would be laughable; from him, not so much.

Hassabis is the 39-year-old former chess master and video-games designer, whose artificial intelligence research startup DeepMind, was bought by Google in 2014, for a reported $625 million. A “visionary” manager, according to those who work with him, Hassabis also reckons he has found a way to “make science research efficient” and says he is leading an “Apollo programme for the 21st century”.

Artificial intelligence is already all around us, of course, every time we interrogate Siri or get a recommendation on Android. And in the short term, Google products will surely benefit from Hassabis’ research, even if improvements in personalisation, search, YouTube, and speech and facial recognition are not presented as AI as such. In the longer term, though, the technology he is developing is about more than emotional robots and smarter phones. It’s about more than Google. More than Facebook, Microsoft, Apple, and the other giant corporations currently hoovering up AI PhDs and sinking billions into this latest technological arms race.

If it sounds wildly ambitious, it is. Most AI systems are ‘narrow’, training pre-programmed agents to master a particular task and not much else. So IBM’s Deep Blue could beat Gary Kasparov at chess, but would struggle against a three-year-old in a round of noughts and crosses. Hassabis, on the other hand, is taking his inspiration from the human brain and attempting to build the first “general-purpose learning machine”: a single set of flexible, adaptive algorithms that can learn — in the same way biological systems do — how to master any task from scratch, using nothing more than raw data.

In his vision of the future, super-smart machines will work in tandem with human experts to potentially solve anything. “Cancer, climate change, energy, genomics, macroeconomics, financial systems, physics: many of the systems we would like to master are getting so complex,” he argues. “There’s such an information overload that it’s becoming difficult for even the smartest humans to master it in their lifetimes. How do we sift through this deluge of data to find the right insights? One way of thinking of AGI is as a process that will automatically convert unstructured information into actionable knowledge. What we’re working on is potentially a meta-solution to any problem.”

Computers, for obvious reasons, have traditionally been terrible at making such judgments. Go has, therefore, long been considered one of the “outstanding grand challenges” of AI, and most researchers expected at least another decade to pass before a machine could even hope to crack it.

But here was the rigorously peer-reviewed evidence that DeepMind’s new artificial algorithm, AlphaGo, had thrashed the reigning three-times European champion, Fan Hui, 5-0 in a secret tournament last autumn, and was being lined up to play the world champion, Lee Sedol, in March. “A stunning achievement” is how Murray Shanahan, professor of cognitive robotics at Imperial College. “It’s pretty cool, yeah,” Hassabis agrees, sans drama, when we meet in his office to discuss this latest triumph.

“Go is the ultimate: the pinnacle of games, and the richest in terms of intellectual depth. It’s fascinating and beautiful and what’s thrilling for us is not just that we’ve mastered the game, but that we’ve done it with amazingly interesting algorithms.” Playing Go is more of an art than a science, he maintains, “and AlphaGo plays in a very human style, because it’s learned in a human way and then got stronger and stronger by playing, just as you or I would do.” Hassabis may look like a student, but he is beaming like the proudest of parents.

AlphaGo is the most exciting thing he’s achieved in his professional life. “It’s an order of magnitude better than anyone’s ever imagined,” he enthuses, “but the most significant aspect for us is that this isn’t an expert system using hand-crafted rules. It has taught itself to master the game by using general-purpose machine learning techniques. Ultimately, we want to apply these techniques to important real-world problems like climate modelling or complex disease analysis, right? So it’s very exciting to start imagining what it might be able to tackle next.”

My first encounter with Hassabis was back in the summer of 2014, a few months after the DeepMind acquisition. Since then, I’ve observed him at work in a variety of environments and have interviewed him formally for this profile on three separate occasions over the past eight months.

In that time, I’ve watched him evolve from Google’s AI genius to a compelling communicator, who has found an effective way to describe to non-scientists like me his vastly complex work — about which he is infectiously passionate — and why it matters. Unpretentious and increasingly personable, he is very good at breaking down DeepMind’s approach; namely their combining of old and new AI techniques — such as, in Go, using traditional “tree search” methods for analysing moves with modern “deep neural networks”, which approximate the web of neurons in the brain – and also their methodical “marriage” of different areas of AI research.

The way animals learn

In DeepQ, they combined deep neural networks with “reinforcement-learning”, which is the way that all animals learn, via the brain’s dopamine-driven reward system. With AlphaGo, they went one step further and added another, deeper level of reinforcement learning that deals with long-term planning. Next up, they’ll integrate, for example, a memory function, and so on — until, theoretically, every intelligence milestone is in place. “We have an idea on our road map of how many of these capabilities there are,” Hassabis says. “Combining all these different areas is key, because we’re interested in algorithms that can use their learning from one domain and apply that knowledge to a new domain.”

This sounds a bit like the man himself. “I get bored quite easily, and the world is so interesting, there are so many cool things to do,” he admits.(He also holds a world record as five-times winner of the Mind Sports Olympiad’s elite ‘Pentamind’, in which competitors challenge each other across multiple games.) “If I was a physical sportsman, I’d have wanted to be a decathlete.”

His company, 50-strong when Google bought it, now employs almost 200 people from over 45 countries and occupies all six floors of a building, in a regenerated corner of King’s Cross. Hassabis was determined that his company should remain close to his roots, despite pressures to move elsewhere (including, presumably, Mountain View in Silicon Valley.)

“We’re really lucky,” says Hassabis, who compares his company to the Apollo programme and the Manhattan Project for both the breathtaking scale of its ambition and the quality of the minds he is assembling at an ever increasing rate. “We are able to literally get the best scientists from each country each year. So we’ll have, say, the person that won the Physics Olympiad in Poland, the person who got the top maths PhD of the year in France. We’ve got more ideas than we’ve got researchers, but at the same time, there are more great people coming to our door than we can take on. So we’re in a very fortunate position. The only limitation is how many people we can absorb without damaging the culture.”

Insisting that the Google acquisition has not in any way forced him to deviate from his own research path, Hassabis reckons he spends “at least as much time thinking about the efficiency of DeepMind as the algorithms” and describes the company as “a blend of the best of academia with the most exciting start-ups, which have this incredible energy and buzz that fuels creativity and progress.” He mentions “creativity” a lot, and observes that although his formal training has all been in the sciences, he is “naturally on the creative or intuitive” side.

I ask Hassabis to outline what he thinks the principal long-term challenges are. “As these systems become more sophisticated, we need to think about how and what they optimise,” he replies. “The technology itself is neutral, but it’s a learning system, so inevitably, they’ll bear some imprint of the value system and culture of the designer so we have to think very carefully about values.”