Hassabis has spent his entire life asking one question: how does general intelligence work, and can we build it? Every result — the games, the neuroscience, AlphaFold, the AGI wager — is an answer to the same question. The protein folding solution was not a productivity gain. It was a new kind of mind entering the room.
What does it mean to solve a problem no human could?
In 2020, AlphaFold predicted protein structures with atomic-level accuracy. The protein folding problem — how an amino acid chain collapses into its three-dimensional shape — had resisted global scientific effort for five decades. AlphaFold didn't inch past the previous best result. At CASP14, it exceeded the cumulative progress of the prior decade in a single submission.
Drug discovery timelines measured in decades began to compress. Diseases with no structural foothold suddenly had one. The Nobel Committee awarded the Prize in Chemistry in 2024 — to a machine learning system, accepted by the man who built it.
Hassabis said at Stockholm: "The goal has always been to use AI to accelerate scientific discovery to benefit humanity. AlphaFold is a proof of concept — but it is nowhere near the end of the road."
That last sentence is the one worth sitting with.
If AlphaFold is the proof of concept, what is the experiment?
AlphaFold didn't solve protein folding incrementally. It made fifty years of prior progress look like a warm-up.
Was the chess prodigy always building toward this?
Hassabis became a chess master at thirteen. Not a talented junior — one of the strongest under-14 players in the world. The skill wasn't raw genius. It was pattern recognition compounded by obsessive practice. He was already doing what DeepMind would later formalize: extracting structure from complexity, faster than any peer.
At seventeen, he entered Cambridge. He finished his computer science degree in two years. Before graduating, he had co-designed Theme Park at Bullfrog Productions — a commercial hit, and an early experiment in what he would later call emergent rule-governed complexity. A game where small rules generate large, unpredictable behaviors. The same logic he would apply to Go. Then to biology.
He didn't go to Silicon Valley. He went back to London, to University College London, and started a PhD in cognitive neuroscience. That choice was not a detour. It was the research question he needed answered before he could build what he intended to build.
Every chapter of Hassabis's life was pointed at the same question. He just needed the neuroscience before he could write the code.
What did the neuroscience tell him that the computer science couldn't?
His doctoral research at UCL established something that should have been more disruptive than it was. Hippocampal damage doesn't only erase memory. It destroys the ability to imagine the future.
Patients with hippocampal lesions cannot picture tomorrow. They cannot construct a hypothetical scene, simulate a conversation, or visualize themselves somewhere they have never been. The machinery for reconstructing the past and simulating the future is the same machinery. Memory and imagination are one system.
This is not a minor cognitive science footnote. It reframes consciousness, planning, and hope as a single computational act. The mind does not retrieve the past and then separately predict the future. It runs one process — episodic simulation — in two directions.
For Hassabis, this meant something precise: any machine capable of genuine reasoning would need something analogous. Not retrieval. Not pattern matching against a database. Simulated experience. The ability to construct scenarios that have never happened.
He published. Then he went to build it.
The hippocampus reconstructs past events from fragments. It does not play back recordings — it reassembles. Each recall is a new construction.
AlphaFold does not retrieve known structures. It constructs predictions from sequence data, using learned physical constraints. Each output is generated, not retrieved.
Hippocampal patients cannot simulate futures. Intact minds run the same system forward — building scenes that haven't happened yet.
AlphaGo does not look up moves. It simulates futures — running thousands of hypothetical game states forward to evaluate the present position.
What happened when the games stopped being games?
DeepMind was founded in London in 2010. Co-founders: Hassabis, Shane Legg, Mustafa Suleyman. The explicit mandate was artificial general intelligence — a system that reasons across every domain the way humans do. Not a narrow tool. Not a better search engine. The whole thing.
The industry considered this naive. Most AI companies were building specialized systems — better at one task, useless at others. Hassabis was betting on generality. He believed games were the fastest path to get there.
Games are closed worlds with perfect rules and measurable outcomes. They are laboratories for learning. AlphaGo proved it.
In March 2016, DeepMind's Go-playing AI played world champion Lee Sedol. The result: 4-1. The field had not expected this for at least another decade. Go's search space dwarfs chess by orders of magnitude. The branching factor is too large for brute-force computation. AlphaGo didn't brute-force it. It learned to see the board — to evaluate positions through something closer to intuition than calculation.
Move 37 in Game 2 became famous. A play so unexpected that Sedol left the room. Commentators didn't know whether to call it brilliant or broken. It was neither. It was outside the tradition. AlphaGo had found territory that millennia of human play had never entered.
Then came AlphaZero — which taught itself chess, Go, and shogi from scratch, with no human game data, in hours. It developed what grandmasters called alien chess. Strategies no tradition had produced. Sacrifices that paid off twenty moves later. Positional ideas that looked wrong and weren't.
Something had shifted. The games were no longer just games.
AlphaGo's Move 37 unsettled observers in a way chess victories never had. Not because the machine won — because it went somewhere humans hadn't thought to go.
What did Google's acquisition actually cost?
In 2014, Google acquired DeepMind for more than $500 million. Hassabis negotiated structural protections unusual for any acquisition, let alone one at that price. An independent ethics board. Commitments against certain military applications. Restrictions on how DeepMind's work could be used.
Those commitments have since been contested. The ethics board was dissolved. The boundary between DeepMind's research agenda and Google's commercial interests has blurred. The tension between building AGI for humanity and building it inside a corporation optimized for shareholder return is not resolved. It may not be resolvable.
Hassabis has said publicly that he loses sleep over the risks. He has also said AGI could arrive within his lifetime — and that it could be the greatest benefit or the greatest threat humanity has ever faced. He does not present these as separate possibilities.
He keeps building.
The question of what it costs to do this work inside a structure like Google is not answered by good intentions. Intentions don't survive institutional pressure. The ethics board proved that. What survived is the science — and the question of who it ultimately serves.
The ethics board Hassabis insisted on was dissolved. What remains is the science, inside a corporation, pointed at AGI.
What comes after the hardest problems are no longer ours to solve?
AlphaZero developed chess strategies no human tradition had produced. AlphaFold solved a problem no human had solved. If the pattern holds, the next system Hassabis builds will enter cognitive territory humans cannot follow.
This is not a metaphor. It is the stated goal. AGI — a system that reasons across every domain — would by definition be capable of scientific problems that exceed human cognitive reach. Not faster at our problems. Capable of problems we cannot formulate.
Hassabis's neuroscience research showed that imagination and memory are one system. The ability to simulate futures that haven't happened is what makes planning, meaning, and hope possible. Those capacities are what distinguish directed inquiry from reflex. They are what make the search for knowledge feel like ours.
If a machine acquires that capacity — and AlphaFold is, by Hassabis's own account, a proof of concept — the question of who is doing the knowing becomes urgent. Not philosophically. Practically. The drug was discovered. The disease may be cured. The insight belongs to no one.
Whether that is liberation or loss is not a question Hassabis has answered. He has said he thinks about it constantly. He has said the stakes are existential in both directions. He has said the work must continue.
He is probably right that it will. The question is whether the civilization doing this work has thought carefully enough about what it is reaching for — and who gets to decide when enough has been reached.
If imagination and memory are one system, and machines acquire both, who is left doing the imagining?
If AlphaFold solved a novel problem by constructing predictions it had never seen — was it imagining, in any meaningful sense, or only retrieving at a scale that mimics imagination?
AlphaZero found chess strategies that five centuries of human play had missed. If AI is already exploring cognitive territories humans cannot reach, who decides which territories are safe to enter — and by what authority?
Hassabis insists the ethics of AGI must be resolved before the technology arrives. He is also building as fast as he can. Is that a contradiction, or is it the only honest position available to someone who believes the alternative is someone else building it worse?
The hippocampal research showed that the same system that remembers also hopes. When machines can simulate futures better than we can, what happens to hope as a human act?
DeepMind's ethics board was dissolved inside a decade. If institutional protections cannot survive commercial acquisition, what structure could actually constrain a technology moving this fast?