The Man Who Thought He Could Keep AI Safe

At first glance, Demis Hassabis, a co-founder of the DeepMind AI lab, seems a familiar type: the missionary entrepreneur and out-of-the-box scientist who emerges as the right person for a particular moment. In this case, that moment was when hardware and software and data aligned to make superintelligence possible. But Hassabis is hardly a conventional figure. He has devoted his life to creating a technology that he thinks has the potential to destroy the world.

Hassabis agreed to talk with me about his quest, because he believes that societies will never trust inventors unless they understand what makes them tick. For almost three years, as I worked on my book The Infinity Machine, we met regularly at a pub near his home, in North London. We would climb a shabby wooden staircase to a room on the second floor, sit with cappuccinos under a once-grand chandelier, and spend two hours talking: me with an obsessively detailed list of topics to get through; Hassabis with his sparky riffs on intelligence and life, computer science and neuroscience, philosophy and movies. Through every conversation, the question of motivation hung in the air like the image of the mushroom cloud over Los Alamos.

Hassabis is fluent in the full gamut of AI doom scenarios. He met one of his DeepMind co-founders, Shane Legg, at a lecture on AI safety. He buttonholed his first financial backer, Peter Thiel, at a Singularity summit, where futurists shared visions of machines that outsmart people. Once, Hassabis impressed a bumptious rocket builder, Elon Musk, by telling him that space colonization would not ensure human survival. Superintelligent systems could also build rockets, Hassabis observed. If the systems turned out to be malign, a colony on Mars would offer no protection.