Demis Hassabis: The case that AI's real test is science, not chatbots

Demis Hassabis was born in London in 1976 to a Greek Cypriot father and a Chinese Singaporean mother. A chess prodigy from age four, he reached master standard by thirteen with a rating around 2300 and captained England junior teams. Chess winnings bought his first computer, and he taught himself to program from books.

He finished his A-levels at sixteen and, asked by Cambridge to wait a year, spent the gap at Bullfrog Productions, where at seventeen he co-designed and lead-programmed the 1994 hit Theme Park with Peter Molyneux. He then took a double first in computer science at Queens' College, Cambridge, worked as lead AI programmer on Black and White at Lionhead, and in 1998 founded his own studio, Elixir Studios, which shipped Republic: The Revolution and Evil Genius before closing in 2005. The through line was simulation: building systems in which agents make decisions, and watching what humans find compelling about them.

After Elixir, Hassabis did something few technology founders would consider: he went back to university to study the brain. His UCL PhD in cognitive neuroscience, completed in 2009 under Eleanor Maguire, produced a landmark 2007 PNAS paper showing that patients with hippocampal damage, known to cause amnesia, were also unable to imagine new experiences. The finding linked memory and imagination as one constructive process and was listed by Science among the year's top ten breakthroughs.

The intellectual yield was a working theory he later called a simulation engine of the mind: the brain as a system for imagining scenarios in order to plan. Postdoctoral work at the Gatsby Computational Neuroscience Unit, where he met Shane Legg, turned that theory into a research agenda. If imagination and planning are the core of intelligence, then a learning system should be tested on exactly those capacities. That is a description of what DeepMind would spend a decade doing.

Hassabis co-founded DeepMind in London in 2010 with Shane Legg and Mustafa Suleyman, on a mission to solve intelligence and then use it to solve everything else. The method was deep reinforcement learning: general algorithms learning from reward, not hand-coded rules. In 2013 the lab showed a single system learning to play dozens of Atari games from raw pixels, and in 2014 Google acquired the company for a reported 400 million pounds.

AlphaGo made the laboratory famous. In October 2015 it beat European champion Fan Hui five games to nil, and in March 2016 it defeated Lee Sedol, one of the greatest Go players alive, four games to one in Seoul. Go had been considered decades away from machine mastery. The match was less a stunt than a demonstration that learning systems could find strategies no human had articulated, in a domain long treated as the benchmark of intuitive intelligence.

From 2016, DeepMind turned the same approach on protein structure prediction, a fifty-year grand challenge in biology. AlphaFold won the CASP13 assessment in 2018, and in November 2020 AlphaFold 2 effectively solved the core problem at CASP14, achieving accuracy competitive with experimental methods. Within two years the lab, working with EMBL-EBI, had predicted and published the structures of more than 200 million proteins, essentially every protein known to science, in a free open database.

In 2021 Alphabet spun out Isomorphic Labs, led by Hassabis, to apply the approach to drug discovery. In October 2024 he and John Jumper received the Nobel Prize in Chemistry for AlphaFold, sharing the year with David Baker for computational protein design; Hassabis was also knighted that year for services to AI. His Nobel lecture in Stockholm, delivered in December 2024, framed the work plainly: AI as an engine for accelerating scientific discovery.

Hassabis's position on the present is more measured than the discourse around it. He signed the 2023 statement that mitigating extinction risk from AI should be a global priority, while arguing that a pause is unenforceable and that the benefits, in health and climate, justify continuing under serious evaluation. In 2023 Google merged DeepMind with Google Brain under his leadership to form Google DeepMind, the unit behind Gemini. In August 2026, in an Alphabet reshuffle, he stepped down as Google DeepMind's CEO to become chair of the lab and chief scientist of Alphabet, continuing to lead Isomorphic Labs while Koray Kavukcuoglu took over day-to-day operations.

The honest assessment cuts both ways. He was right early about the two things that mattered most: that learning systems would beat engineered ones, and that AI's first civilizational payoff would come in science rather than in consumer software. The contested ground is timelines and concentration. His estimates of when AGI arrives have moved closer over the years and remain disputed, and the model he proved, frontier AI as a subsidiary of a platform giant, is now the industry default, with the governance questions that implies. DeepMind's own health work, absorbed into Google Health after a UK regulator ruled in 2017 that an early hospital data-sharing arrangement was unlawful, is the standing reminder that scientific ambition does not exempt anyone from institutional discipline.