Engineering Philosophy: Demis Hassabis, Solve Intelligence to Solve Everything

Key Takeaways

• The plan was two steps: solve intelligence, then use it to solve everything else. That sentence – Hassabis’s actual founding pitch for DeepMind – is the whole strategy. Build a general learning system rather than a narrow one, then point it at the hardest problems in science. It sounds like hubris until you watch him execute it in order.¹⁷
• Games were the proving ground, not the goal. AlphaGo beat Lee Sedol 4-1 in March 2016 – a decade ahead of expert predictions – and game two’s “Move 37,” a play with a roughly 1-in-10,000 chance of being chosen by a human, showed the system inventing rather than imitating.⁵ Hassabis chose games because they are self-contained, with clear objectives, the ideal sandbox for reinforcement learning.⁷
• Then he turned that intelligence on biology and won a Nobel. AlphaFold2 solved the 50-year-old protein-structure-prediction problem at CASP14 in 2020; the AlphaFold database now holds structures for ~200 million proteins used by over two million researchers.⁸⁹ In 2024 Hassabis shared the Nobel Prize in Chemistry with John Jumper (and David Baker) – step two, delivered.¹⁰¹¹
• From chess prodigy to game designer to neuroscientist to founder. Born in London in 1976, master-strength at chess by thirteen, co-designer of Theme Park at seventeen, a Cambridge double-first, then a UCL PhD in cognitive neuroscience studying memory and imagination – every chapter fed the next.¹²³⁴

The Principle

“Step one, solve intelligence; step two, use it to solve everything else.” – Demis Hassabis, on DeepMind’s founding plan⁷

Most ambitious technology is built the other way around. You pick a problem – search, recommendations, fraud detection – and you build the narrowest, most reliable system that solves that. Generality is treated as a luxury you earn later, if ever. Hassabis inverted the order. His bet was that the right first move is not to solve any particular problem but to solve the general capacity that solves problems – to build a learning system that can master domains it was never specifically engineered for – and only then to aim it at the targets that matter most.⁷

That is why the order in the sentence is load-bearing. “Solve intelligence” comes first not because the applications don’t matter but because, in his view, a sufficiently general intelligence is the highest-leverage tool you can build: solve it once, and you get a key that opens many locks rather than a tool shaped for one. The phrase can read as grandiose, and it would be – if he hadn’t then spent fifteen years walking the two steps in sequence, in public, and arriving at a Nobel Prize for the second one.¹⁰

The method underneath the slogan has two unusual ingredients. The first is neuroscience as inspiration: Hassabis studied the brain precisely to mine it for algorithmic ideas, on the theory that the one general intelligence we know of is worth understanding before you try to build another.⁷ The second is games as the proving ground: self-contained worlds with clear objective functions, where a learning system can be trained, measured, and pushed to superhuman play before it is ever trusted with something real.⁷ Games to general learning to scientific discovery – that is the whole arc, and it is the principle in one line: solve intelligence, then use it to solve everything else.

Context

Demis Hassabis was born in London on 27 July 1976, to a Greek Cypriot father and a Chinese Singaporean mother.¹ He was a chess prodigy from the age of four; by thirteen he had reached master standard with an Elo rating around 2300, and he captained England junior teams.¹ Chess is not a footnote here – it is the origin of the whole worldview. A child who spends years calculating lines, evaluating positions, and choosing the best continuation has internalized, before puberty, the exact loop that would later sit under AlphaGo: look ahead, evaluate, choose.

The next chapter was games – the building of them. After winning a competition for a job at Bullfrog Productions, Hassabis co-designed and lead-programmed the simulation classic Theme Park at seventeen, working alongside Peter Molyneux; it sold millions and helped spawn the management-sim genre.² He went on to be lead AI programmer on Black & White at Lionhead, then founded his own studio, Elixir Studios, shipping Republic: The Revolution and Evil Genius.¹ A decade of his life went into making software that simulated intelligent behavior – crowds, creatures, opponents – which is exactly the apprenticeship you would design for someone about to try to build the real thing.

Then he went back to first principles. He had taken a double-first in Computer Science at Cambridge in 1997, and he returned to academia for a PhD in cognitive neuroscience at UCL (2009), working on episodic memory and imagination – showing that patients with hippocampal damage struggle not only to remember the past but to imagine novel future scenes.¹³⁴ That is a profound finding for an AI builder: the machinery of memory and the machinery of imagination are the same machinery. He did postdoctoral work at the Gatsby Computational Neuroscience Unit, and in 2010 he co-founded DeepMind with Shane Legg and Mustafa Suleyman; Google acquired it in 2014.¹ Chess taught him search; games taught him simulation; neuroscience taught him what general intelligence looks like from the inside. DeepMind is where the three converge.

The Work

Games as the proving ground: AlphaGo and Move 37

Start with the engine itself, in miniature. Long before neural networks, the core idea behind a game-playing machine was lookahead: before committing to a move, imagine the opponent’s replies, your replies to those, and so on down the tree of possibilities, score where each line ends up, and choose the move that leads to the best outcome assuming both sides play well. The widget below is that idea stripped to its smallest form – tic-tac-toe against an opponent that searches every continuation before each move, so it cannot be beaten. Play it, and you are playing the great-grandparent of AlphaGo.

Tic-tac-toe is small enough to search to the end. Go is not. A Go board has more legal positions than there are atoms in the observable universe, which is why brute-force lookahead – the technique that cracked chess – had failed at Go for decades.⁵ DeepMind’s move was to keep the search but make it intelligent: AlphaGo combined Monte Carlo tree search with two deep neural networks – a policy network that proposes promising moves (so you don’t waste search on bad ones) and a value network that judges how good a position is (so you don’t have to read every line to the end). The networks were trained first on human expert games and then sharpened through reinforcement learning, playing millions of games against versions of itself.⁵

In March 2016, in Seoul, AlphaGo beat Lee Sedol – one of the greatest players alive – 4-1, before an audience of more than 200 million, roughly a decade earlier than experts had predicted.⁵ The moment that defines it is game two’s Move 37: a play the system itself estimated a human would choose with about a 1-in-10,000 probability, initially read as a mistake by commentators, and later understood as brilliant. It was the first time the world watched a machine produce, in a domain humans had refined for millennia, not an imitation of human genius but something genuinely new.⁵ That is the whole point of the games phase: not to win, but to demonstrate that a learning system could discover.

AlphaZero: learning from nothing but the rules

AlphaGo still leaned on a crutch: it bootstrapped from a database of human games. The next step removed even that. AlphaZero (introduced in 2017, full results in Science in 2018) was handed only the rules of a game and then taught itself entirely through self-play, starting from random moves – no human games, no opening books, no handcrafted heuristics.⁶ Tabula rasa. From that blank start it mastered Go, chess, and shogi at superhuman level, outplaying the strongest specialized engines – and it did so with a single, general algorithm applied to three very different games, the clearest evidence yet for the “general” in general intelligence.⁶

The detail I find most instructive is how it won. AlphaZero searched only about 60,000 positions per second in chess, against the 60 million of the traditional engine Stockfish – a thousandfold fewer – and still came out ahead, because its learned intuition told it which lines were worth reading at all.⁶ It also played with an alien, sacrificial, deeply positional style that surprised grandmasters. This is the same lesson as Move 37, generalized: when you stop forcing a system to imitate human knowledge and let it learn the structure of a problem from scratch, it does not merely match us – it finds things we missed. That is the bridge from “solve intelligence” to step two, because most of science’s hard problems have no human game-book to copy from in the first place.

AlphaFold: turning intelligence on science (Nobel 2024)

Then Hassabis spent the key. For fifty years, biology’s grand challenge had been the protein-folding problem: a protein is a chain of amino acids that folds into a precise 3D shape, that shape determines what the protein does, and predicting the shape from the sequence had resisted every approach. Experimental methods to determine one structure could take years and cost a fortune.⁸⁹

At CASP14 in 2020 – the field’s blind, biennial assessment – DeepMind’s AlphaFold2 predicted structures to within roughly an atom’s width of the experimental answer, about three times more accurate than the next-best system and, for most proteins, comparable to the lab.⁸ The CASP organizers declared the 50-year problem essentially solved.⁸ DeepMind then did the thing that turns a result into infrastructure: it released the AlphaFold Protein Structure Database, growing from the human proteome to ~200 million structures – nearly every catalogued protein known to science – freely available, now used by over two million researchers in 190 countries.⁹¹¹

In October 2024, the Nobel Committee awarded Hassabis and his colleague John Jumper half of the Nobel Prize in Chemistry “for protein structure prediction,” sharing the prize with David Baker, recognized “for computational protein design.”¹⁰¹¹ Hassabis was knighted the same year for services to artificial intelligence.¹ The thing to sit with is the order: this is not a separate career pivot from games to biology. It is step two of the original sentence. Build a general learning system, validate it where the objective is clean (a game), then aim it at a problem where the objective is the same shape – a vast search space, a clear scoring function – but the payoff is a tool for all of biology.

The mission and Isomorphic: what’s next

If AlphaFold proved the thesis, Isomorphic Labs – which Hassabis founded and leads, spun out of DeepMind in 2021 – is the attempt to industrialize it: using AI to reinvent the drug-discovery process, treating the whole pipeline from target to molecule as a problem an AlphaFold-style system can attack.¹ It is step two, again, on a bigger target: from predicting a protein’s shape to designing the molecule that binds it. Meanwhile DeepMind’s stated mission has broadened to building AI responsibly to benefit humanity, and Hassabis has become one of the more cautious senior voices on AGI – urging care precisely because he takes the “solve everything else” half literally.⁷ The pattern holds in both directions: the ambition is enormous, and the discipline about how to get there is just as serious.

The Method

Hassabis’s method is unusually legible, because he has narrated it. Strip away the press and it is a repeatable recipe.

Aim at the general capability, not the specific task. The defining choice was to “solve intelligence” first and applications second – to build a learning system that generalizes, rather than the narrowest thing that ships. Most teams cannot afford this. The discipline is knowing when generality is the actual leverage and when it is procrastination.⁷

Mine the one working example. There is exactly one general intelligence in existence – the brain – so Hassabis studied neuroscience to steal its ideas. When a problem has a single known solution in nature, understanding it deeply beats inventing from scratch. The same instinct runs through this series: encode the structure that already works rather than reinvent it, as with LeCun baking translation invariance into convolution.⁷

Build the proving ground before the product. Games gave him self-contained worlds with clear objective functions where a system could be trained and measured honestly before it touched anything that mattered. The general lesson: invest in the environment where you can get a clean, fast signal of whether the thing works. A benchmark you trust is worth more than an opinion you don’t – the evidence gate applied to a whole research program.⁷

Remove the human crutch when you can. AlphaGo learned from human games; AlphaZero learned from none, and got better. When a system has matured enough to learn the structure of a problem directly, the human-supplied scaffolding can become a ceiling. Knowing when to kick it away is its own skill.⁵⁶

Then spend the capability on something that matters. The discipline that makes the whole thing more than a stunt is the second step actually happening. Game-playing was never the point; AlphaFold was. Capability without a worthy target is incomplete – the Steve test of whether the work deserves to exist, applied to intelligence itself.⁸¹⁰

Influence Chain

Who Shaped Him

Chess. Before computer science, before neuroscience, there was the board. The loop of lookahead, evaluation, and choice that a master internalizes is the same loop that sits under AlphaGo and AlphaZero. Hassabis learned the algorithm as a child by living it. (Formative influence)

The brain, studied deliberately. Hassabis took a PhD in cognitive neuroscience explicitly to learn from the only general intelligence we know of, working on the shared machinery of memory and imagination. The bet that AI should be brain-inspired is not a metaphor for him; it was a research plan. (Direct influence)

The deep-learning revolution. AlphaGo and AlphaFold are deep neural networks at their core, and that lineage runs straight through Geoffrey Hinton, whose work made the networks trainable, and Yann LeCun, whose convolutional architectures taught networks to see structure. Hassabis built the search and the system; they built the substrate it learns on. (Direct influence)

Who He Shaped

AI for science. AlphaFold did not just solve one problem; it established a template – that a general learning system, aimed at a hard scientific question with a clean objective, can outrun decades of specialized effort. Every “AlphaFold for X” project is downstream of that demonstration.

Reinforcement learning at scale. AlphaGo and AlphaZero are the canonical proof that deep reinforcement learning with self-play can reach and exceed human expertise in vast search spaces, reshaping what an entire subfield believed was possible.

The public imagination of AI. Move 37 and the Lee Sedol match were, for hundreds of millions of people, the moment a machine stopped imitating and started creating. That cultural marker is part of his influence too.

The Throughline

Hassabis is where this series’ deep-learning branch turns from perception into action and discovery. Fei-Fei Li supplied the data that taught networks to see; Geoffrey Hinton made the learning machine actually work; Yann LeCun gave it the architecture to find structure. Hassabis takes those same networks and wraps them in search and self-play – a system that doesn’t just classify the world but acts in it, plans, and discovers. The forward line runs naturally to Andrej Karpathy’s “Software 2.0,” the idea of a program compiled from data rather than written by hand, which is exactly what AlphaZero is: no rules of strategy programmed in, only the rules of the game and a reward, with everything else learned. LeCun says learn to see; Hinton says the learning works; Li says here is the world to learn from; Hassabis says: now use it to do something – and points it at a fifty-year-old problem in biology. (Series bridge)

What I Take From This

The lesson I keep from Hassabis is about sequencing ambition. The “solve intelligence, then solve everything else” sentence is easy to mock as a founder’s grandiosity, and it would be – except he treated it as a literal two-step plan and executed the steps in order, in public, for fifteen years, with a Nobel Prize at the end of step two. The discipline isn’t the size of the ambition; it’s the refusal to skip the proving ground. He didn’t claim to cure disease on day one. He built a system, validated it superhumanly somewhere the score was unambiguous, and only then aimed it at the target that mattered. That reorders how I think about big goals: state the audacious end, but earn the right to it on a clean benchmark first. It is quality is the only variable applied to a roadmap – the question is “is the capability real?” before “is the application impressive?”

The second lesson is quieter and runs through the whole arc: the best ideas often come from studying the one example that already works. Hassabis didn’t theorize about intelligence in the abstract; he went and studied the brain, because it is the existence proof. When I’m stuck, the move is rarely to invent from first principles in a vacuum – it’s to find the system that already solved a version of this and understand why it works well enough to steal the idea. Chess gave him search, the brain gave him architecture, games gave him a sandbox, and biology gave him a worthy target. Nothing was wasted, because each chapter was him mining a working example for the next. Solve intelligence, then use it – but first, go learn from the thing that already has it.

FAQ

What is Demis Hassabis’s engineering philosophy?

Solve intelligence first, then use it to solve everything else. Rather than build the narrowest system that solves a specific task, Hassabis bet on building a general learning system – inspired by how the brain works, validated in games where the objective is clean – and then pointing that general capability at the hardest problems in science.⁷ The strategy is legible because he executed it in order: AlphaGo and AlphaZero proved the intelligence was real and general, and AlphaFold spent it on biology’s protein-folding problem, winning a Nobel.⁵⁶⁸¹⁰

How does AlphaGo work, and what was Move 37?

AlphaGo combined Monte Carlo tree search with two deep neural networks: a policy network that proposes promising moves and a value network that judges how good a position is, trained first on human games and then sharpened by reinforcement learning through self-play.⁵ Because Go has more legal positions than there are atoms in the universe, exhaustive search is impossible – the networks let the system search intelligently instead. In March 2016 AlphaGo beat Lee Sedol 4-1 in Seoul, and game two’s “Move 37” – a play with roughly a 1-in-10,000 chance of being chosen by a human – was the moment a machine produced a genuinely novel idea in a game humans had refined for millennia.⁵

What is the difference between AlphaGo and AlphaZero?

AlphaGo learned partly from a database of human expert games before improving through self-play. AlphaZero (2017) removed the human data entirely: given only the rules of a game, it taught itself from random play through pure self-play reinforcement learning – tabula rasa.⁶ From that blank start, a single general algorithm mastered Go, chess, and shogi at superhuman level, beating the strongest specialized engines while searching far fewer positions, because its learned intuition told it which lines were worth reading.⁶ AlphaZero is the stronger evidence for “general” intelligence because the same method worked across three different games with no domain-specific tuning.

What is AlphaFold, and why did it win the Nobel Prize?

AlphaFold is DeepMind’s AI system for predicting a protein’s 3D structure from its amino-acid sequence – the “protein-folding problem” that had resisted solution for about fifty years.⁸⁹ At the CASP14 assessment in 2020, AlphaFold2 predicted structures to roughly atomic accuracy, comparable to experimental methods, and the organizers declared the problem essentially solved.⁸ DeepMind released ~200 million predicted structures – nearly every known protein – free to researchers.⁹ In 2024 Hassabis and John Jumper were awarded half the Nobel Prize in Chemistry “for protein structure prediction,” sharing it with David Baker, recognized “for computational protein design.”¹⁰¹¹

Sources