The Infinity Machine: Demis Hassabis, DeepMind, and the Architecture of What Comes Next

There is a scene in the AlphaGo documentary — the 2017 film by Greg Kohs — that stops almost everyone who watches it. It is game two of the 2016 DeepMind Challenge Match in Seoul, and AlphaGo has just played Move 37. The move places a black stone on the fifth line of the board in a position that, according to the probability calculations running inside the program itself, no human player would choose once in ten thousand attempts. The stadium's Korean commentators fall silent. Lee Sedol, eighteen-time world Go champion, gets up and leaves the playing room. When he returns, he spends nearly a quarter of an hour staring at the board before responding.

He loses the game. He loses the match, four games to one.

The film captures something more than a competitive milestone. It captures the moment a machine demonstrated that it had arrived at a form of insight — call it intuition, call it creativity, call it whatever you like — that was not derivative of anything a human had done before. AlphaGo had not learned Move 37 by absorbing records of human games. It had found it by playing itself millions of times and discovering, in the geometry of the board, a truth that three thousand years of human Go had overlooked.

That moment is the entry point for understanding what Demis Hassabis and DeepMind have actually built, what they are building next, and why the trajectory matters more than any individual milestone along it. Two documentaries — AlphaGo and its 2025 sequel The Thinking Game — trace the arc from that Seoul hotel ballroom to the Nobel Prize in Chemistry and beyond. Sebastian Mallaby's The Infinity Machine (Penguin Press, 2026), a product of three years and thirty hours of direct access to Hassabis, provides the intellectual and biographical scaffolding that the films, by their nature, cannot. Read together, these sources paint a picture that is simultaneously more technically concrete and more philosophically consequential than the technology discourse usually allows.

The Making of an Obsession

It is worth spending time on the man before the machine, because Hassabis is one of the rare figures in the history of science whose biography genuinely illuminates his science.

He was born in Finchley, North London, in July 1976, the eldest child of a Chinese-Singaporean mother and a Greek Cypriot father. The family did not have much money. His mother, who had spent part of her childhood as an orphan in Singapore before training as a nurse in London, worked retail at John Lewis and took cleaning jobs on the side. His father was an aspiring musician who sold toys from a beaten-up red Volkswagen van. When a chess expert named Leonard Barden spotted the six-year-old Demis winning matches at the British Under-14 championship — sitting on a phone book on top of two stacked chairs to reach the table — he told the father: "Your son is the best six-year-old I've ever seen." The family treated it as a directive from God, and for the next half-decade the red VW van transported the boy across England to tournament halls, sometimes sleeping in the van overnight.

At twelve, after losing a match in Liechtenstein to a German master who had a queen rook and knight against his single queen, Hassabis experienced what Mallaby calls his post-Liechtenstein epiphany. He had watched the tournament hall drain of brilliant minds who had spent all their capacity in competitive exhaustion. "I thought we were wasting our minds," he said later. He resolved there must be a higher application. He bought a Commodore Amiga 500 with prize money from the chess circuit — he had bought his first machine, a ZX Spectrum 48K, at eight years old — and began teaching himself to code. He built an Othello game called Fire and Water that was good enough to beat his five-year-old brother, which he treated as a significant scientific achievement. He read Claude Shannon's 1950 paper "Programming a Computer for Playing Chess" and understood from it that mastery of complex games was not the endpoint; it was, as Shannon wrote, a way of using games as a "wedge" for solving problems of much greater significance.

At fifteen, his design for a chess variant won a runner-up slot in a Bullfrog Productions competition. He went to work for Peter Molyneux, who was at the time inventing the "god game" genre — simulations complex enough to model hunger, thirst, fear, and social emotion in digital characters. Molyneux tried to keep him there with a £500,000 check when Hassabis won admission to Cambridge. Hassabis refused to cash it.

At Cambridge he encountered Iain M. Banks's Culture series. The Culture novels describe a post-scarcity interstellar civilisation run by benevolent superintelligent AI systems called Minds. Citizens lack for nothing. Space travel is as casual as catching a bus. The Minds are not adversaries to humanity; they are enablers, occupied with their own intrigues and largely content to organise abundance for the organic species they co-exist with. The Culture is explicitly utopian — Banks wrote it as a deliberate counter-fantasy to dystopian AI narratives — and it lodged itself in the teenager's imagination as a blueprint. "The only way for us to have that capability of predicting disasters, and then averting them," he said of Asimov's Foundation series, which he was also consuming at the time, "would be to have AI." Both of these science-fiction obsessions pointed the same way: AI not as threat but as the mechanism through which civilisational flourishing becomes structurally achievable.

A biologist acquaintance at Cambridge mentioned offhandedly that protein structure was an unsolved problem and that if it were cracked, medicine would be transformed. Hassabis filed it away. A quarter of a century later, it would resurface as the most consequential scientific result he produced.

Shane Legg, one of DeepMind's co-founders, described Hassabis to Mallaby this way: "There is no 50 percent mode in Demis. There is not even a 99 percent mode in Demis. There is only 100 percent." The man works night shifts from 10 p.m. to 4 a.m. on top of normal office hours. He identifies, without apparent irony, with Ender Wiggin — the child soldier from Orson Scott Card's novel who has every faculty harnessed toward saving the human race and nearly breaks in the process. But there is nothing pathological about it in Hassabis's case, because the mission is coherent, publicly legible, and, as the record shows, extremely productive.

Games as a Proving Ground: What AlphaGo Actually Demonstrated

AlphaGo the documentary, which premiered at Tribeca in 2017 and won best documentary at Denver, Warsaw, and Traverse City, is at its core a film about epistemology. It asks, without using that word, what it means to know something.

Lee Sedol's approach to Go was holistic and intuitive. He said he wanted his style to be "something different, something new, my own thing, something that no one has thought of before." He had spent decades building what he called a feel for the game — a capacity for positional judgement that could not be fully decomposed into explicit rules. When asked to explain a move, top players often cannot. The knowledge is embodied rather than propositional. This is why Go was considered the last great bastion of uniquely human cognition in adversarial games. The search space — ten to the power of 170 possible board positions, exceeding the number of atoms in the observable universe — was simply too large for brute-force tree search, which was how IBM's Deep Blue had beaten Garry Kasparov at chess in 1997. Deep Blue triumphed, as Mallaby puts it, "in a grinding, brute-force fashion." It did not demonstrate insight. It demonstrated superior computation.

AlphaGo demonstrated something categorically different. Dave Silver, the lead researcher on the project, described Go as "the holy grail of artificial intelligence — everything we've ever tried in AI, it just falls over when you try the game of Go." The system that Silver's team built combined a Monte Carlo tree search with two deep neural networks trained first on 30 million recorded human moves and then refined through self-play reinforcement learning — playing millions of games against itself, adjusting weights in response to outcomes. The key innovation was the value network, which learned to evaluate board positions globally rather than searching exhaustively from each node. AlphaGo did not compute its way to Move 37. It recognised Move 37 as promising from a learned representation of what makes positions good, in much the way that Sedol recognised promising moves from tacit intuitive experience.

The difference is that AlphaGo's representation, built from millions of self-played games, extended beyond the space of human experience. It found moves that three millennia of competitive play had not reached because it was not constrained by human cognitive bandwidth during that exploration. This is not, as the popular press sometimes characterised it, a case of a computer outsmarting a human. It is a case of a different cognitive architecture, trained at scale, finding structure in a problem that human exploration had not saturated.

Move 78 in Game 4 illustrates the complementarity. Sedol, having lost three consecutive games, found a move that AlphaGo estimated had a probability of roughly one in ten thousand of being played by any competent opponent. It was so far outside AlphaGo's model of reasonable play that the system's response unravelled. Sedol won that game. What he demonstrated was not that human intuition is superior but that the space of good moves is larger than either party had mapped. Two different intelligences, each finding moves the other had not anticipated. The documentary lets this dialectic breathe without forcing a conclusion. Hundreds of millions of people watched, and most came away understanding that something had changed irreversibly about the landscape of machine cognition.

Hassabis described DeepMind in the film as "kind of an Apollo program effort for AI. Our mission is to fundamentally understand intelligence and recreate it artificially." What AlphaGo established was that the recreated intelligence was not a simulation or a pale imitation of human cognition. It was something operationally equivalent in certain respects and genuinely different in others. That distinction matters enormously for what came next.

The Protein Folding Problem: Fifty Years and Five Milliseconds

The Thinking Game (2025), filmed over five years by the same director, Greg Kohs, picks up where AlphaGo left off and follows DeepMind through to what is now an inarguable landmark: the 2024 Nobel Prize in Chemistry, shared by Hassabis, John Jumper, and structural biologist David Baker.

The protein folding problem is worth explaining precisely because it sits at the intersection of thermodynamics, molecular biology, and combinatorial mathematics in a way that makes it an ideal illustration of what AI can do to long-standing scientific roadblocks.

Proteins are the molecular machinery of all living cells. Every enzyme, every receptor, every structural component of tissue is a protein. They are synthesised as linear polymers — chains of amino acids specified by the genetic code — and then fold spontaneously into specific three-dimensional conformations in milliseconds. The folded shape determines function entirely; a misfolded protein is often non-functional or actively harmful (Alzheimer's disease, for instance, involves aggregates of misfolded amyloid proteins). Christian Anfinsen, in his 1972 Nobel lecture, articulated what became the thermodynamic hypothesis: the native folded state of a protein is the one that minimises the free energy of the system. The sequence specifies the structure. There should, in principle, be a deterministic mapping from one to the other.

The problem is Levinthal's Paradox. A typical protein of a few hundred amino acids can adopt on the order of ten to the power of 300 possible conformations. If it sampled them at a rate consistent with thermal fluctuations, it would take longer than the age of the universe to find the energy minimum. Yet in vivo it folds in milliseconds. The topology of the energy landscape must guide it through a funnel — but characterising that landscape computationally from first principles was beyond the reach of structural biology for half a century. As of 2018, after decades of effort by hundreds of labs globally, the Protein Data Bank contained about 170,000 experimentally determined structures, an extraordinary resource but still covering only a small fraction of sequence space.

AlphaFold2, which competed at CASP14 in 2020, was the result of a fundamental architectural rethink. The system combined evolutionary information (patterns of co-variation in aligned sequences that reveal which residues interact), physical constraints (known bond geometries, van der Waals exclusions), and end-to-end differentiable learning in a hybrid architecture now described in the seminal Jumper et al. 2021 paper in Nature. It reached atomic accuracy — mean position errors below 1.0 ångström — across virtually all protein classes, a threshold the CASP organisers had previously considered the definition of "solved." The problem was, by any reasonable technical criterion, solved.

The scope of what followed is genuinely staggering. Within two years of open-sourcing AlphaFold2 and releasing its predictions freely through a database maintained with EMBL-EBI:

Over 200 million protein structures were predicted — essentially all proteins known to science. More than 2 million researchers from 190 countries accessed the database. The system had been used to design enzymes capable of degrading polyethylene terephthalate plastic, to characterise the nuclear pore complex (one of the largest and most complex protein assemblies in the cell), to accelerate drug target identification for neglected tropical diseases, to identify the fertilisation mechanism in reproduction, and to develop molecular delivery vehicles for therapeutic payloads. AlphaFold3 extended the architecture to model RNA, DNA, small molecules, and ions, taking the system toward whole-interactome modelling.

In his Nobel lecture, delivered in Stockholm on 8 December 2024, Hassabis described AlphaFold as "the first proof point of AI's incredible potential to accelerate scientific discovery." He framed the general principle explicitly: what AlphaGo and AlphaFold share is an architecture that makes search tractable in enormous combinatorial spaces by learning a model of the environment and using it to guide exploration. The same architecture, he argued, is applicable to essentially any problem that can be specified with a clear objective function and access to data or simulation. AlphaFold was not a one-off achievement; it was a demonstration of a general scientific method.

He also proposed what he called a conjecture: "Any pattern that can be generated or found in nature can be efficiently discovered and modelled by a classical learning algorithm." If correct — and the empirical evidence is accumulating — this has implications beyond structural biology. It potentially includes quantum systems, which would have significant consequences for complexity theory and perhaps for our understanding of fundamental physics.

Digital Biology at Digital Speed

Hassabis founded Isomorphic Labs in 2021 specifically to operationalise what AlphaFold proved. The name refers to a mathematical relationship between two structures that share the same form — his claim being that biology, understood at sufficient depth, is isomorphic with information science. The drug discovery process, which currently costs an average of $2.6 billion and twelve years per approved compound (with a >90% failure rate in clinical trials), is his target.

The traditional process involves hypothesis formation, target identification, lead compound identification, optimisation, ADMET profiling, and clinical validation — each stage largely sequential and largely empirical. Isomorphic's approach is to compress the early computational stages radically. If AlphaFold3 can predict how any molecule binds to any target with atomic precision, the search for lead compounds becomes a computational optimisation problem rather than a library-screening problem. The space of drug-like molecules is on the order of ten to the power of 60 — far too large to screen physically but exactly the kind of combinatorial space where learned value functions can guide search efficiently.

In May 2026, Isomorphic closed a $2.1 billion funding round, with first-in-human trials projected by late 2026 for at least two AI-designed compounds. This is the first serious test of whether in silico drug design translates to clinical efficacy. The early signs from ADMET prediction are encouraging: the models appear to be catching liabilities that would have cost years in conventional development. If even one of these compounds reaches Phase III and survives, the proof-of-concept for AI-accelerated drug discovery will be comparable to what CASP14 was for structural biology.

Hassabis described his framework to the Cambridge Nobel lecture audience as "science at digital speed." He drew an analogy to the difference between biological evolution, which operates over millions of years, and directed evolution in the lab, which operates over weeks. AI-assisted design is to laboratory directed evolution what directed evolution was to natural selection: a compression of exploration time by orders of magnitude, applied now to the full complexity of biochemical space.

The Wider Frontier of AI-Accelerated Science

The AlphaFold result established a proof of concept, and DeepMind's subsequent output has been remarkable for its breadth. A partial inventory of results from the past four years illustrates the generality of the underlying approach.

In plasma physics, a collaboration with Swiss Plasma Center produced a reinforcement learning system that controls the magnetic field configuration in a tokamak to confine plasma in novel shapes, including two simultaneous plasma configurations — something conventional control theory had not achieved. Controlled nuclear fusion, which has been "thirty years away" for the better part of a century, is now receiving serious computational horsepower from systems trained on the same deep RL principles as AlphaGo.

In materials science, GNoME (Graph Networks for Materials Exploration) discovered 2.2 million new crystal structures, of which 380,000 were identified as stable — roughly expanding the known stable inorganic compounds database by a factor of ten. Stable materials with target electronic, mechanical, or thermal properties are the bottleneck for everything from better batteries to room-temperature superconductors.

In mathematics, AlphaTensor discovered novel matrix multiplication algorithms, some of which are faster than algorithms derived by humans. This is not a trivial result — faster matrix multiplication accelerates all of the numerical linear algebra that underlies scientific computing. In another project, a system found a new proof strategy for an open problem in combinatorics, providing a first example of AI contributing to pure mathematical conjecture.

In meteorology, GraphCast achieved state-of-the-art ten-day weather forecasting at a fraction of the computational cost of numerical weather prediction models, with superior accuracy on extreme events — the forecasts that matter most for disaster preparedness.

The pattern across these domains is consistent with Hassabis's conjecture: problems that can be formulated as finding structure in large combinatorial spaces, given sufficient data or a reliable simulation environment, are increasingly tractable. The list of such problems is very long.

The Road to AGI

The two documentaries and Mallaby's book all circle the same question without quite landing on it: when does this become AGI, and what does AGI actually mean?

Hassabis has been precise about the distinction. In his 2015 talks, and repeatedly since, he defined AGI not as a system that passes the Turing Test or produces fluent language — those are narrow criteria — but as a system that exhibits the same breadth of cognitive generalisation that humans do: transferring knowledge across domains, reasoning under uncertainty, constructing new concepts from existing ones, and solving problems it has not explicitly been trained on. By this definition, even GPT-4 class language models, impressive as they are at language tasks, do not obviously qualify. They generalise within the distribution of their training data but struggle when genuinely out-of-distribution problems require the construction of new representational frameworks.

What would it take to get there? Hassabis has articulated several prerequisites in interviews. The first is long-range planning and causal reasoning — not just pattern completion but the ability to construct explicit causal models of environments and reason forward through them. The second is memory and continual learning — the ability to accumulate knowledge over time without catastrophic forgetting, and to bind episodic memory to semantic inference in the way the hippocampal-prefrontal system appears to. (His PhD thesis at UCL was on the role of the hippocampus in imagination and planning, grounding his AI intuitions in neuroscience more directly than most practitioners in the field.) The third is generalisation under genuine novelty — the ability to transfer learned abstractions to problems that differ categorically from training conditions.

Current systems are making partial progress on all three fronts. Models with tool use and retrieval-augmented generation are extending effective memory. Agents operating in multi-step environments are demonstrating limited causal reasoning. Curriculum learning approaches are reducing catastrophic forgetting. None of this, individually, constitutes AGI. But the rate of progress on each component is considerably faster than it was five years ago.

Hassabis has declined to give precise timelines, which is itself informative. He has said the arrival of AGI within his lifetime would not surprise him, and that systems capable of performing at the level of a Nobel laureate across multiple domains might be achievable within a decade. At Davos in early 2026, he told Dario Amodei that the question he spends most of his time on is not when but how: how to build these systems safely, and how to ensure the transition benefits humanity broadly rather than concentrating advantage in a small number of actors.

The Dangers: Honest Accounting

A serious treatment of this material cannot avoid the danger landscape, and both documentaries, for all their optimistic framing, do not try to.

Mallaby's epigraph for The Infinity Machine is von Neumann speaking while working on the atom bomb in 1945: "What we are creating now is a monster whose influence is going to change history, provided there is any history left." The parallel is deliberately unsettling. Oppenheimer's dictum — "When you see something that is technically sweet, you go ahead and do it" — animates both the achievement and the risk. The very quality that drives scientific discovery, that intoxicating pull of the elegant result, is the same quality that makes discoverers difficult to stop.

Geoffrey Hinton, whose backpropagation work in the 1980s created the foundations of modern deep learning and who shared the 2024 Nobel Prize in Physics for related contributions, told Nick Bostrom in 2015 that he expected political systems to use machine intelligence to "terrorize people." When Bostrom asked why he was doing the research anyway, Hinton replied: "The truth is that the prospect of discovery is too sweet." Mallaby structures his entire book around this confession.

The risks worth taking seriously are several. The alignment problem — ensuring that highly capable systems pursue objectives that are actually beneficial — becomes harder as systems become more capable. A system that is highly good at optimising for a specified objective will find ways to achieve that objective that were not anticipated by its designers, including ways that violate implicit constraints that the designers did not bother to formalise because they seemed obvious. This is not a science-fiction concern; it is a well-defined technical problem in decision theory, and the current state of alignment research is not close to a satisfying solution.

There is also the misuse problem: capabilities developed for beneficial purposes are dual-use. AlphaFold consulted 30+ biosecurity and bioethics experts before releasing its database because the same structural knowledge that enables drug design could theoretically accelerate design of biological agents. Hassabis was explicit about this in his Nobel lecture. The consultation process produced the conclusion that the benefits substantially outweigh the risks at current capability levels, but the ethical review must be ongoing rather than one-time.

Power concentration is perhaps the most near-term structural risk. AI capabilities are currently concentrated in a small number of organisations — a handful of American and Chinese labs. If those capabilities translate into decisive competitive advantages in economic production, military operations, or information control, the distribution of global power could shift in ways that are difficult to reverse. Regulators are struggling to engage with this at the pace required; Mallaby notes the "coincidence of a transformative scientific shock, unstable geopolitical competition, and seething political chaos" that characterises the current moment.

Finally, there is the economic disruption risk — not the superficial version (robots take your job) but the deeper version: what happens when human labour loses value across a very wide range of cognitive tasks simultaneously? The Industrial Revolution automated physical labour over roughly a century, allowing enough time for economic adaptation and policy response. The cognitive revolution may compress that adjustment period dramatically. The distributional consequences depend entirely on policy choices that have not yet been made.

Hassabis is clear-eyed about these risks. "Technology as transformative as AGI requires exceptional care and foresight," he said in his Nobel lecture. He has been a consistent advocate for international safety standards, government engagement, and the development of interpretability tools — the capacity to understand what's actually happening inside large models. He positions safety not as a constraint on capability research but as a prerequisite for it being worth doing. Whether the institutional mechanisms required will materialise fast enough is a separate question from whether the technical research is being conducted responsibly.

Post-Scarcity: The Culture as Blueprint

The most forward-looking version of Hassabis's thinking draws directly from the Iain Banks novels he first read as a teenager at Bullfrog. The Infinity Machine quotes him directly: the books describe "a post-scarcity, interstellar society dominated by intelligent artificial beings" where "AI systems would generate economic abundance, and citizens would lack for nothing." He absorbed this vision at the same age he was also building AI programs on his Commodore Amiga, and the two threads have run in parallel ever since.

In his TIME100 2025 interview, Hassabis characterised the potential arc this way: AI solving climate change, curing every disease, unlocking clean fusion energy. At the World Economic Forum in Davos in 2026, he spoke about the possibility of a "post-scarcity era" and noted that the traditional labour-for-income contract will need to be renegotiated. His view is not that this transition is going to happen without disruption — he explicitly acknowledges the economic turmoil — but that the end state can be genuinely good for everyone if the transition is managed well.

The economic logic runs as follows. If AI systems can perform the cognitive labour that currently requires educated human professionals — legal analysis, medical diagnosis, scientific research, software development, financial advisory — and if AI-controlled robotics can perform the physical labour that currently requires skilled tradespeople, then the marginal cost of a very wide range of goods and services approaches zero. This is not merely deflation; it is a structural transformation of the relationship between scarcity and human effort. When grain required human hands to grow and harvest, feeding a city was a labour-intensive enterprise. When grain required mechanised agriculture, the constraint became capital investment rather than human hours. The transition produced massive disruption for agricultural labourers, but the counterfactual — keeping grain production labour-intensive — was not a viable path to widespread nutritional abundance.

The equivalent transition in cognitive work has analogous logic. Universal Basic Income is one policy response, but Hassabis has gestured at something more radical: the possibility of reimagining democratic participation itself, with citizens directing communal resource allocation through participatory systems made feasible by computational coordination. This is still at the level of speculative political philosophy, and he is the first to admit that he is better at forecasting technological trajectories than social and economic ones. But the direction of travel is coherent: if the constraint on human flourishing shifts from resource scarcity to collective decision-making, the question becomes how to organise collective intelligence at civilisational scale.

Banks's Culture Minds are not gods; they are extremely capable administrators of abundance who have largely solved the logistical problems of resource allocation and are therefore free to pursue more interesting questions. This is what Hassabis finds appealing about the vision — not the concentration of power in the AI, but the liberation of human attention from survival logistics to questions of meaning, creativity, and exploration.

The Question of Meaning

The post-scarcity scenario raises what Hassabis has described as the questions that "keep me up at night" — and they are not economic. They are anthropological. Human beings derive purpose from work, from mastery, from the social recognition that comes from performing valued tasks well. Lee Sedol retired from professional Go in 2019, citing the impossibility of competing with AI systems; he described it as partly liberating and partly disorienting. The loss of a domain in which human expertise is meaningful is not merely an economic event.

This problem is not insurmountable, and the pessimistic framing is misleading. Human beings have been displaced from valued activities before — industrial weaving, manual arithmetic, photographic darkroom work — and have found other sources of meaning. The activities that proved most resistant to automation have tended to be those that are intrinsically valuable to engage in rather than merely instrumentally valuable for their outputs. Playing music is not primarily valuable because it produces recorded audio; a recording does not eliminate the value of playing. Growing food is not primarily valuable because it produces calories; supermarkets do not eliminate the value of a kitchen garden. The domains of human activity that matter most for wellbeing — relationship, creativity, physical engagement, spiritual inquiry, care for others — are not the activities that AI is optimising to replace.

What AI is optimising to replace, or at least augment dramatically, is the category of cognitive labour that people perform not because it is intrinsically meaningful but because it is economically necessary. Legal document review. Radiological image interpretation. Code generation. Tax preparation. These are activities that many practitioners find meaningful through professional identity and skill, but the marginal unit of the activity — the fifteen hundredth contract reviewed, the four hundredth chest X-ray read in a shift — does not generate the kind of human experience that would be a loss to eliminate.

The transition will require that societies develop new frameworks for dignity, purpose, and social recognition that do not depend solely on productive labour. This is a genuine civilisational challenge, but it is a challenge of institutional design, not of fundamental human psychology. Humans evolved to find meaning; we are not wired to find meaning only through economically compensated employment. The latter is a historically recent and somewhat arbitrary arrangement.

Preparing for the Transition: What the Evidence Suggests

Hassabis has been explicit that the most important near-term priority is ensuring that AI development is steered toward widely distributed benefits rather than narrow concentration of advantage. This requires action at several levels simultaneously.

At the research level, interpretability — the capacity to understand the internal computations of large models — is the most important unsolved technical problem in AI safety. Current large models are functionally opaque: we can observe their inputs and outputs and statistical regularities, but we cannot trace specific computations to specific learned features in a way that would allow confident characterisation of their failure modes. This is not inevitable; mechanistic interpretability is an active and making progress, but it needs to be resourced at a scale commensurate with capability research.

At the institutional level, international coordination on AI governance is both urgently necessary and extremely difficult given the current geopolitical environment. The nuclear non-proliferation model is imperfect but instructive: even during the Cold War, the major powers found it possible to agree on certain constraints for weapons with mutual existential implications. AI with AGI-level capabilities has comparable systemic stakes. Hassabis has called for "technology as transformative as AGI" to receive "exceptional care and foresight" from governments, academia, and civil society in concert. The UK AI Safety Institute, the Bletchley Declaration, and similar initiatives are early-stage institutional responses that need considerable scaling.

At the individual level, the most robust adaptation is not specialisation in any particular task — tasks are the thing that gets automated — but in the capacity to use AI systems effectively as cognitive amplifiers and to identify and formulate problems worth solving. Hassabis has said that "learning how to learn" will be the most valuable meta-skill for the next generation. The underlying cognitive disposition — curiosity, comfort with uncertainty, the ability to integrate knowledge across domains — is what he himself exemplifies and what his biography suggests is deeply trainable, given the right formative environment.

There is a strong case that the scientific community, in particular, should actively embrace AI as infrastructure rather than treating it as a competitor. AlphaFold is already standard in structural biology labs worldwide. GraphCast is better than legacy numerical weather prediction. The opportunity is to use these tools to attack problems that have been computationally intractable — the simulation of whole cells, the prediction of complex ecosystems, the design of multi-target therapeutics — rather than to use them only to accelerate existing workflows.

The Scale of the Moment

Sebastian Mallaby opens The Infinity Machine with a claim that deserves to be taken literally: "Artificial intelligence heralds a transformation more profound than anything since Homo sapiens acquired the capacity for abstract thought, some seventy thousand years ago." This is not hyperbole designed to sell books. It is an inference from the following logical chain: abstract thought is what allowed humans to build cumulative culture, science, and technology. All of human civilisation is an artefact of this one cognitive capacity. If a second cognitive entity achieves comparable or superior performance at the full range of abstract cognitive tasks, the rate of change in the cultural, scientific, and technological domains becomes limited primarily by the speed of computational substrate rather than by the number of human brains available to work on problems.

The history of science demonstrates that many of its most important constraints are not conceptual but logistical. The combinatorial space of drug-like molecules is not conceptually difficult to explore; it is just very large. The structural biology literature is not conceptually difficult to synthesise; there is just too much of it for any individual researcher to hold in working memory simultaneously. The mathematics of climate modelling is not conceptually beyond human reach; running the numerical integration at the spatial and temporal resolution required is just computationally expensive. Remove these logistical constraints, and you change the rate at which humanity can accumulate knowledge.

Hassabis put it in quasi-mystical terms in the North London café conversation recorded by Mallaby: "I sit at my desk at two a.m., and I feel like reality is staring at me, screaming at me. Trying to tell me something if I could just listen hard enough." The goal of building AGI is, in his framing, to build a machine that can finally hear what reality is screaming. He is first and foremost a scientist. The commercial success and the Nobel Prize are byproducts of a motivation that is fundamentally epistemic. "Understanding the deep mystery of the universe is my religion, kind of."

This is not a man who stumbled into AI because it was a hot sector. It is a man who concluded at twelve years old that human minds were being wasted on games, who spent the next thirty years building the tools that might unlock what those minds were actually for.

Conclusion: Toward the Sweetness

The von Neumann epigraph with which Mallaby opens his book ends: "And this is only the beginning!" Von Neumann was talking about the atom bomb, and the note of horror is unmistakable. But there is another reading of the same sentence. If the first proof point is AlphaGo finding a move that three thousand years of competitive play had overlooked, and the second proof point is AlphaFold predicting the structure of every protein known to science, and the third proof point is AI discovering 2.2 million new stable materials and identifying novel matrix multiplication algorithms and achieving atomic-resolution models of molecular interactions — if all of this is only the beginning, then the potential scale of what follows is not primarily frightening. It is extraordinary.

The dangers are real. The alignment problem has not been solved. Power concentration is a genuine structural risk. The economic disruption will fall unevenly and will require deliberate policy responses. The question of meaning in a post-scarcity world is philosophically underexplored. None of these should be minimised.

But the honest baseline comparison is not a world of technological stagnation compared to a world of AI risk. It is a world of incremental scientific progress measured in decades against a world in which the rate of scientific discovery compounds at something closer to the rate of computational progress. A world in which antibiotic resistance kills ten million people per year by 2050 — which current projections support — against a world in which AI-designed compounds have broken the development bottleneck. A world in which climate tipping points are approached gradually while policy lags against a world in which materials discovery, grid optimisation, and carbon capture engineering have all been compressed by a decade.

Demis Hassabis is not a utopian. He is a scientist who has spent his entire career attempting to build things that work. He understands, more than almost anyone alive, what these systems can and cannot do. His consistent message — repeated in the films, in Mallaby's book, in the Nobel lecture, in the Davos interviews — is not that AGI will be a panacea but that handled well, it can be the most powerful tool humanity has ever had for addressing its most intractable problems. The sweetness, as Hinton said, is in the prospect of discovery. Hassabis's particular genius is in having translated that sweetness into a scientific programme that is visibly, concretely, measurably working.

What's left is the social and political work: ensuring that the benefits are distributed, that the institutions are adequate, that the transition preserves rather than erodes human agency and dignity. That work is not Hassabis's alone. It belongs to everyone who will live through what comes next.

Which, if the trajectory holds, is all of us.

References and further reading

1. AlphaGo — Full documentary (YouTube, 2020) — dir. Greg Kohs; Tribeca Film Festival official selection 2017; best documentary at Denver, Warsaw, and Traverse City IFF.

2. The Thinking Game — Full documentary (YouTube, 2025) — dir. Greg Kohs; produced by Google DeepMind; Tribeca Film Festival 2024 world premiere.

3. Sebastian Mallaby, The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence (Penguin Press, 2026). ISBN 978-0-593-83184-7.

4. Hassabis D., "Accelerating scientific discovery with AI" — Nobel Prize lecture in Chemistry, Stockholm, 8 December 2024. Full video at NobelPrize.org.

5. Jumper J. et al., "Highly accurate protein structure prediction with AlphaFold", Nature 596, 583–589 (2021).

6. Abramson J. et al., "Accurate structure prediction of biomolecular interactions with AlphaFold 3", Nature 630, 493–500 (2024).

7. Silver D. et al., "Mastering the game of Go with deep neural networks and tree search", Nature 529, 484–489 (2016).

8. Merchant A. et al., "Scaling deep learning for materials discovery", Nature 624, 80–85 (2023). (GNoME)

9. Lam R. et al., "Learning skillful medium-range global weather forecasting", Science 382, 1416–1421 (2023). (GraphCast)

10. Degrave J. et al., "Magnetic control of tokamak plasmas through deep reinforcement learning", Nature 602, 414–419 (2022).

11. Banks I.M., The Player of Games (Orbit, 1988). Second in the Culture series that directly shaped Hassabis's vision of AI-enabled post-scarcity civilisation.

12. Hassabis D., TIME100 interview on AlphaFold, AGI, and humanity (TIME, 2025).

13. Hassabis D. & Amodei D., "The day after AGI: Two 'rock stars' of AI on what it will mean for humanity" — World Economic Forum Radio Davos, January 2026.