Hassabis Codex

Six frontier labs, six axes

Score Google DeepMind, OpenAI, Anthropic, Meta AI, xAI and Microsoft AI across scientific output, product velocity, openness, safety investment, compute access and talent density. The polygons diverge precisely where the labs disagree on what AI is for. Hassabis's lab occupies a distinctive corner: highest in scientific output, very high in compute and talent, and unusually high — for a frontier lab — on safety investment.

In October 2024 the Royal Swedish Academy of Sciences awarded a share of the Nobel Prize in Chemistry to Demis Hassabis and John Jumper for solving a problem biology had carried for fifty years — predicting the three-dimensional structure of proteins from their amino-acid sequence. It was the first time an AI system had been singled out by a Nobel committee as the instrument of a fundamental scientific advance, and the first time the head of a private AI lab was credited as one of its scientific authors. The award did not come from nowhere. It was the culmination of an unusual career — chess master at thirteen, video-game designer at seventeen, neuroscientist at thirty, AI founder at thirty-four, Google's AI head at fifty — and of an unusual institution, DeepMind, that has spent fifteen years arguing that the right way to build artificial general intelligence is to take the brain seriously as an existence proof and the game as the natural training ground. This Codex reads that career and that institution as a single structural object — not biographical, not technical, but the kind of reading you would give to a body of work whose meaning is not yet settled.

Hassabis's career has the unusual property of not being one career. By his early twenties he had already lived through two: a competitive chess career — he peaked at master strength at thirteen and was, by twelve, the second-highest-rated player under fourteen in the world — and a video-game career — Theme Park at Bullfrog as a teenage co-designer, then Black & White at Lionhead with Peter Molyneux, then his own studio Elixir, which built ambitious simulation games before closing in 2005. Then he went back to school. The Cambridge undergraduate computer-science years were followed by a doctorate at UCL in cognitive neuroscience, studying memory and imagination with Eleanor Maguire — the work that produced a Science paper on episodic memory and that earned him a 2007 Top Ten Breakthroughs listing. DeepMind followed in 2010, then Google's acquisition in 2014, then the Atari/AlphaGo years, then the AlphaFold years, then in 2023 the formal merger of DeepMind with Google Brain that gave him institutional command of one of the two or three labs that will determine what intelligence the next decade ships. Five lives, each a small mastery; the Codex argues that the unusual combination — chess pattern recognition, games as simulated reality, neuroscience as the brain's existence proof, DeepMind as the engineering vehicle, Google as the distribution — is the real signature.

Hassabis's most-quoted line — used in talks, papers, internal documents — is that the brain is 'the only existence proof we have of general intelligence'. The methodological commitment that follows is structural and the source of DeepMind's distinctive shape. Where most AI labs at the time of DeepMind's founding (2010) treated cognitive science as decoration, DeepMind treated it as a starting point: identify a neuroscientific principle (episodic memory, reward prediction, hippocampal replay, predictive coding, world-model learning), formalise it computationally, build a system that implements it, and use the system's performance as a test of the principle. The hippocampal-replay → experience-replay → Deep Q-Network chain is the cleanest example: a finding about how the rodent brain reactivates past trajectories during sleep became, after several remove, the trick that let DQN learn Atari games from pixels. The brain hypothesis is not vague inspiration. It is a research programme that decided, in 2010, that the right next move in AI was to bring cognitive neuroscientists into the same room as deep-learning engineers and to insist they design systems together. Whether that programme has been vindicated by AlphaFold and the others is a question the Codex returns to throughout; that it remains DeepMind's distinguishing methodological commitment is not in doubt.

DeepMind was founded in London in September 2010 by three people whose combination is part of what the company became: Demis Hassabis (the neuroscientist-engineer), Shane Legg (the AI safety theorist who had been working on machine super-intelligence at Singularity Institute / Gatsby), and Mustafa Suleyman (the systems-and-people operator, later founder of Inflection and then head of Microsoft's consumer-AI division). The pitch to early investors — Founders Fund, Horizons, Elon Musk, Peter Thiel — was unusual: 'build the world's first artificial general intelligence', framed not as a product company but as a research lab with a thirty-year horizon. In 2014 Google acquired DeepMind for a figure reported in the high hundreds of millions, with the unusual condition that an independent ethics board be established (which Google subsequently dissolved into broader internal governance). The acquisition gave DeepMind compute, runway, and access to Google's data; it cost DeepMind some of its institutional independence and some of Suleyman, who eventually left. The shape of post-acquisition DeepMind — research-led, paper-heavy, brand-conscious, with a sustained focus on hard scientific problems rather than direct consumer product — is the architectural signature of the founding triad and the bargain they struck with Google.

In late 2013 DeepMind published a workshop paper showing that a single neural network, given only the raw pixels of an Atari 2600 screen and a score, could learn to play seven games at human or near-human level using deep Q-learning. In 2015 Nature published the full result — 49 games, a single architecture, no game-specific engineering, exceeding human professional play on more than half. This was the demonstration that bought DeepMind its credibility. The technique stack — convolutional neural network seeing the screen, experience replay borrowed from hippocampal research, target network for stability, a single Q-learning update rule — has been picked apart in ten thousand follow-up papers; the institutional fact is that it was the first sustained demonstration that the same learning algorithm could acquire many distinct competencies without being told what each competence required. The Atari project also set DeepMind's pattern for the next decade: pick a hard benchmark the field already cared about, beat it with a clean method that has neuroscience roots, publish in Nature, move on. The Atari paper made DeepMind the lab that would, three years later, take on Go.

Go had been the standing problem of the field — too large for tree search, too dependent on pattern intuition for hand-coded heuristics. In March 2016, in a five-game match in Seoul, AlphaGo defeated Lee Sedol four games to one. The technical structure combined convolutional value and policy networks trained on professional games with Monte Carlo Tree Search and reinforcement learning from self-play; the cultural moment was the second game's move 37 — a play on the fifth line, in the opening, that no top professional would have considered, which AlphaGo's policy network had estimated as having a one-in-ten-thousand probability of being a human move and which decided the game. Lee Sedol's response — that he had felt for the first time that he was playing against an opponent rather than a program — became the moment the public understood that something different had happened. AlphaZero (2017) went further: the same architecture, with no human game data, learned chess, shogi and Go from self-play alone, defeating Stockfish at chess and elite engines at shogi. The pair of results showed that reinforcement learning from self-play, at scale, could match or exceed centuries of accumulated human play in three different game traditions. The Go programme is, in retrospect, the moment Hassabis's gaming and chess years became part of the technical thesis: games are the natural training ground.

Protein folding had been called a fifty-year problem. Each of the body's proteins folds in milliseconds into a unique three-dimensional shape that determines its function; predicting that shape from the linear amino-acid sequence is computationally intractable by brute force. In December 2020 DeepMind's AlphaFold 2 won the CASP14 protein-structure-prediction competition by a margin so large that the organisers said the problem was, in effect, solved for single-chain proteins. AlphaFold 2's architecture combines evolutionary information across related sequences with a deep attention-based model and a structural module that iteratively refines a predicted shape. DeepMind then made the choice that turned the result into a public-good infrastructure: in 2021 it released the AlphaFold Protein Structure Database, containing predicted structures for over 200 million proteins — essentially every protein known to science — free to use. Pharmacology, structural biology, vaccine design, basic biology research were all accelerated. AlphaFold 3 (2024) extended the architecture to predict interactions between proteins, DNA, RNA, and small molecules. The Nobel Committee for Chemistry recognised AlphaFold in October 2024, awarding Hassabis and Jumper a share of the prize with David Baker. The recognition was not for an AI achievement; it was for a chemistry achievement that happened to be enabled by AI. That distinction matters.

The arrival of GPT-3 and then ChatGPT changed the field's centre of gravity in a year. Google, despite having invented the transformer in 2017 and having Google Brain working on large language models, was structurally too cautious to ship a competitor in 2022. In April 2023 Google merged Google Brain and DeepMind into a single division — Google DeepMind — with Hassabis at the head. The new division shipped Gemini 1 in December 2023, then Gemini 1.5 with its million-token context window, then Gemini 2 across 2024–25; in parallel it shipped AlphaCode, AlphaProteo, AlphaGeometry, Project Astra, Genie, and Veo, and ran a continued AlphaFold programme. The merger was the institutional transformation that ended DeepMind's protected research-lab status: the lab now had to produce shippable foundation models for Google's product surfaces, on a quarterly cadence, while still doing the long-horizon science that had earned it the reputation Google was leveraging. Hassabis's distinctive contribution to the merger has been to insist that the foundation-model work continue to be coupled to the scientific work — that Gemini is not in a separate building from AlphaFold — and that safety, evaluation and capability research remain in the same organisation. The Codex treats Gemini not as Hassabis's true subject but as the structural test of whether the DeepMind model can survive being a product organisation.

Hassabis's public posture on AI risk is distinctive. He has consistently described AI as 'like fire' — civilizationally transformative, dangerous if mishandled, and not something humanity can choose to leave undeveloped. He signed the Center for AI Safety's 2023 'mitigating the risk of extinction from AI' statement alongside other lab heads. DeepMind has produced a sustained body of safety work — debate, scalable oversight, dangerous-capability evaluations, the AGI Safety paper (Shah et al., 2025) — that is unusual in being internal to a leading capabilities lab rather than separate from it. The tone is also distinctive: where some lab heads have leaned into alarm and others into reassurance, Hassabis speaks with what colleagues describe as the careful voice of someone who genuinely does not know how AI will go and is trying to keep all the options open. He has been clear about disagreeing with Mustafa Suleyman on certain governance proposals, with Sam Altman on shipping cadence, and with the safety-doomer wing on the magnitude of near-term existential risk; he has also been clear that the work needs international coordination of a kind no current institution can provide. The Codex treats safety as the dimension on which Hassabis's career is most exposed: he is the head of a lab whose products may, in his own framing, change the conditions of civilization. The question of whether the careful voice is enough is open.

Step back. What kind of intelligence is Hassabis trying to build? The Codex's reading of the public record is that the answer has been remarkably consistent for fifteen years and is in three pieces. (1) Modelled on the brain — not in the literal sense of simulating neurons, but in the principled sense of taking cognitive-neuroscience findings as a privileged source of architectural hints. (2) Trained in environments rich enough to require general capabilities — games at first, then molecular biology, then the open-ended environment of the internet and code, then, eventually, the physical world via robotics and embodied agents. (3) Built inside a single organisation that holds capability and safety work together — a structural commitment Hassabis has defended through every reorganisation. The synthesis is unusual in three respects. It is more biological than the language-model wing of the field — Hassabis is not betting that scale plus next-token prediction will produce the whole answer. It is more rigorous in scientific output than its peers — the Nature papers, the Nobel, the released AlphaFold database are not a marketing strategy. And it is more institutionally cautious about the political implications than the public conversation suggests — the careful voice is not a posture, it is a working assumption that the technology will only be usable by humanity if its governance keeps pace. Whether all three commitments survive the next decade of pressure — competitive, financial, political — is the open question the Codex stops at.

Ask the Codex

Five open questions the public record raises and does not close, read in turn by a cognitive neuroscientist, a chess master, a game designer, a mathematician, a safety researcher, and a business strategist. Where they agree is solid ground; where they diverge is the open frontier — and where Hassabis's career is most legible.

Neuron → civilization

Run the brain hypothesis at every scale, from a single neuron to a civilisation deciding what intelligence is for. Toggle the scales at which the methodology is operative, and watch the work's surface area grow. The most ambitious version of the bet is that the same principles that let a hippocampus reactivate past trajectories will let an AI lab build something that earns a Nobel and shapes the next century. That bet is still being made.