The Infinity Machine (Sebastian Mallaby)
In this Book Essential
INTRODUCTION
In November 2022, Sebastian Mallaby secured something no other writer had: deep, on-the-record access to Demis Hassabis and the leadership of DeepMind. A week later, ChatGPT launched. The book he had begun as a study of a scientific institution became, by accident of timing, the record of how that institution was overtaken by the race it had helped to start. Mallaby reports more than thirty hours of conversation with Hassabis and hundreds of interviews across the industry, and the result is not a corporate history. It is a portrait of an ambition and of its price.
The ambition is stated plainly in the book's title. An infinity machine is a machine that can do induction, not deduction: a system that finds patterns in a search space so large that no human could enumerate it. Hassabis spent his childhood as a chess prodigy learning that calculation alone loses to intuition, his twenties designing games that simulated whole worlds, and his doctorate proving that memory and imagination share the same machinery in the brain. Mallaby uses that biography to explain a technical bet: that general intelligence would come from combining fast, learned intuition with slow, deliberate search, the two systems that AlphaGo fused in 2016.
What makes the book worth your time is that Mallaby refuses to write a hero story. He grants Hassabis the scientific achievement, including the 2024 Nobel Prize in Chemistry for AlphaFold, and then documents the compromises that came with it: an ethics board that never functioned, an independence charter that Google absorbed, a company that missed two consecutive stages of the language model race because its leaders found language less interesting than the problems they had chosen. Mallaby is sympathetic and unsparing at the same time, and the tension between those two registers is the book's real subject.
Read this if you lead a long-horizon project inside a larger institution, if you are deciding how much of your mission you will trade for resources, or if you want to understand why the people building the most consequential technology of the century keep telling us they are worried and keep shipping anyway. The book does not settle that question. It does something more useful: it shows you the specific decisions, in sequence, that produced it.
U365'S VALUE PROPOSITION
WHO THIS IS FOR
Founders and operators of deep-tech ventures, who will recognize the sequence from visionary pitch to funding dependency to governance compromise, and who need to see how each stage was actually negotiated.
Academics and researchers moving between disciplines, because Hassabis's path from computer science to neuroscience to a Nobel Prize is the strongest available case study in how a sideward move can become a strategic advantage.
Product and engineering leaders at AI companies, who will find the clearest published account of why DeepMind's research-first culture lost the assistant market twice, and what it cost to recover.
Educators and curriculum designers, because the book is a sustained argument about what intelligence is, how it is trained, and why the difference between induction and deduction should change what you teach.
Policy and governance professionals, who will see, in detail, why voluntary commitments, ethics boards and independent charters failed in the face of competitive pressure.
KEY TENSIONS
Scientific purity against commercial survival: Hassabis wanted to solve intelligence, and solving it required compute that only a giant could supply. Every funding decision narrowed the range of futures his company could choose.
Ambition against patience: DeepMind's mission was maximal from the first day. Its method was incremental, one game, one protein, one benchmark at a time. Mallaby shows how much discipline the second half of that formula required.
Charisma against control: Hassabis told the author repeatedly that the worst thing you can do to a person is to control them. He also built one of the most concentrated research organizations in the world and presided over his cofounder's removal. The book lets both facts stand.
Safety against speed: DeepMind asked Google for a promise that its technology would never be used for weapons or surveillance. By 2025 Google was selling AI to defense agencies, and Hassabis was arguing that this was inevitable.
Language against everything else: The transformer, the architecture that produced ChatGPT, was invented by Google researchers in 2017. DeepMind's leadership saw it and kept working on agents, games and biology. That judgment cost them the most valuable product category in the industry.
One lab against a race: DeepMind was founded on the premise that a single, well-run institution could steer AI toward safety. The premise collapsed as soon as OpenAI, Anthropic, xAI and Chinese labs entered, because a singleton cannot survive competition.
WHY IT MATTERS NOW
Every organization now faces a version of the decision Mallaby documents at Google. You can build with general models that someone else controls, or you can build your own capability and accept the cost and the dependency. The book is the most detailed account available of what that choice feels like from inside a lab that had the talent, the money and the prestige to take either path, and chose the slower one twice.
The book also arrives at the moment when the safety argument has narrowed to a claim about people rather than structures. Hassabis's own conclusion, as Mallaby reports it, is that governance mechanisms matter less than the values of the small group of leaders who control the technology. That is a defensible position and a troubling one. It places the future of the technology in the hands of maybe a dozen individuals, and it asks the rest of us to trust their judgment. Mallaby does not resolve the tension; he documents how it was reached.
For anyone teaching or studying AI, the book supplies something the technical literature rarely provides: a causal account of which ideas actually mattered. Reinforcement learning plus deep learning won Go. Transformers plus scale won language. Biology yielded to both, combined with a decade of stubborn domain work. Those are teachable conclusions, and they are more reliable when you have watched them being discovered.
OVERVIEW
The Infinity Machine is a biography of Demis Hassabis written through the history of the lab he founded, and a history of the lab written through the choices of its founder. Mallaby structures it as a chronological narrative from a North London childhood in a family of modest means, through chess, Cambridge, a games studio that nearly destroyed him, a neuroscience doctorate at University College London, and the founding of DeepMind in 2010, to the Google acquisition, AlphaGo, AlphaFold, ChatGPT, the merger into Google DeepMind, and the Nobel Prize.
The book's central claim is technical before it is moral. Classical AI tried to encode logic, and logic cannot handle a world of infinite variation. Intelligence, on Mallaby's account, is the capacity to induce patterns from examples, and the way to build it is to combine a learned model of the world with a search process that tests possibilities against that model. AlphaGo is the demonstration: a policy network that proposes moves by intuition, a value network that judges positions, and a tree search that explores ahead. AlphaFold is the second demonstration, applied to a search space larger than Go's by an enormous margin.
Mallaby's method is access plus reconstruction. He is in the room for the arguments, and he has the emails. The most memorable passages are not the triumphs but the negotiations: Peter Thiel's lieutenants sizing up a company they did not understand, a colleague describing Larry Page's indifference to hummingbirds and whales, Mustafa Suleyman fighting for an independence charter and losing, Sundar Pichai settling a corporate structure over a series of negotiations that produced fifteen billion dollars and no real autonomy.
The structure rewards attention. The early chapters establish the intellectual bet. The middle chapters show the bet paying off in games and biology. The final third documents the cost: the missed language models, the botched Bard launch, the safety letter, the board crisis at OpenAI, and a founder who admits, in the last pages, that the moment does not feel the way he imagined. It is a book about achievement and disappointment, and it is unusually honest about the relationship between them.
KEY IDEAS
SUPERINTELLIGENCE BY LADDER
The infinity machine: induction instead of rules. Mallaby's organizing idea is that AI escaped its first half-century of stagnation when researchers stopped writing rules and started feeding examples. A rule-based system fails on the edge cases of the edge cases. An inductive system learns a distribution and generalizes. The name is deliberately grand, and the book earns it by showing how much of DeepMind's success came from finding domains where induction could be tested at scale.
Intuition plus search: the two-system architecture. AlphaGo's policy network supplied the flashes of judgment that chess masters call intuition. The value network and the Monte Carlo tree search supplied deliberation. Mallaby returns to this pairing throughout, and it functions as the book's explanation of both DeepMind's triumphs and its blind spot. Systems built around search excel at bounded problems with clear objectives. Language is not bounded, and its objective is not clear.
Games as controlled infinity. Chess, Atari, Go and StarCraft gave DeepMind something real laboratories rarely provide: an unlimited supply of training data, an unambiguous score, and a simulator that costs nothing to run. Mallaby is careful to note the limitation. A game is an infinity you designed. The world is an infinity you did not.
The ladder, not the leap. Hassabis's most transferable management idea is that a maximal mission needs intermediate rungs. He chose Atari after Go was still out of reach, then Go once the deep reinforcement learning stack worked, then protein folding once the search machinery was proven. Each rung was legible to funders and to researchers, and each produced a public result that bought the next interval of patience.
Scale rewrote the argument. The book is honest about the intellectual defeat that scale administered to hand-crafted knowledge. Hinton's Toronto group, a handful of researchers and a pair of consumer graphics cards, beat entire research programs in computer vision in 2012. DeepMind's own success depended on compute as much as on theory. Mallaby's account makes the point without celebrating it: the ideas that won were often the ones that could absorb more of everything.
The language detour. DeepMind did much of the intellectual groundwork for the transformer, and then failed to make language its priority. Mallaby traces the reasons: a research culture that prized scientific problems over product surfaces, a founder who found conversation less interesting than cognition, and an institutional confidence that the next breakthrough would come from agents. The result was that a Google invention became an OpenAI product.
The safety bargain. DeepMind extracted a promise from Google that its technology would not be used for weapons or surveillance, and created an ethics board to enforce it. Mallaby documents what happened: the board's composition was contested, its independence was limited, and its oversight never became operational in the way its architects intended. The promise did not survive contact with the market.
Charisma as a control system. The book's sharpest psychological observation is that Hassabis's persuasion and his aversion to control are not opposites. A leader who can make people believe a future is real will organize their lives around it, whatever his stated intentions. Mallaby uses the collapse of Elixir Studios, and Shane Legg's exhaustion, to show the mechanism from the inside.
Institutional power is the real constraint. The most consequential decisions in the book are not algorithmic. They are the acquisition, the charter negotiation, the merger, and the decision to release. Mallaby's implicit argument is that the history of AI in this period is a history of corporate governance, and that the technical literature has almost nothing to say about it.
The Nobel as evidence, not decoration. Hassabis's 2024 prize, shared with John Jumper for protein structure prediction, is treated by Mallaby as proof of a method rather than a personal honor. AlphaFold worked because a general learning approach, applied with domain expertise and years of unglamorous iteration, solved a problem that specialists had attacked for fifty years.
The trust substitution. By the end of the book, the governance structures have failed and the argument has moved to character. If safety depends on the values of the people in the room, then the relevant questions are biographical, not procedural. Mallaby does not say this is sufficient. He shows that it is what the participants themselves concluded.
ULM ALIGNMENT
At University 365, the CI-First (Co-Intelligence-First) doctrine teaches you to always invite AI into your reflection and work while remaining the orchestrator. Human Intelligence leads, AI amplifies. This Book Essential connects the book's account of long-horizon ambition to U365's proprietary methods: ULM+EVA (University 365 Life Management powered by the Explore-Visualize-Action Plan cycle) helps you map a mission across six life domains; LIPS+CARE (your digital second brain with the Collect-Action Plan-Review-Execute cycle) captures the decisions and constraints as you learn them; and SL-OS (Successful Life Operating System) integrates all of these with UP-Context-Engineering (context engineering for AI) into a unified life and learning system.
SCHEMA PRIME: ULM Domain: Career and Finance
DOMAIN MAPPING: Primary: Career and Finance. Secondary: Spirit and Mind, Quality of Life. The book is about building an institution around a scientific mission, and about the personal cost of that choice.
EVA PARAGRAPH: You want to commit to a mission large enough to organize a decade of work. The obstacle is that large missions require resources from institutions that will attach conditions, and those conditions quietly rewrite the mission before you notice. The plan: write down, before any negotiation, the three commitments you will not trade, and the order in which you will trade the rest. Then, when an offer arrives, check it against the list instead of against the offer's own framing. Do this before the conversation, because during it you will be persuaded.
LIPS+CARE CAPTURE CARD:
Book: The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence
Primary ULM Domain: Career and Finance
3 Key Takeaways:
1. A maximal mission survives only if it is broken into rungs that funders and peers can see.
2. Every resource you accept narrows the set of futures you can still choose.
3. Governance written at the moment of maximum bargaining power is the only governance that will still exist later.
Apply It Action: Draft your own three non-tradable commitments in the Career and Finance domain, and date the document.
Next CARE Step: Collect, by listing every condition attached to your current funding, employer or partnership.
EXPLAINER LINK: For more on ULM, EVA, LIPS, and CARE methods, see the [ULM page](https://www.university-365.com/ulm) and the [LIPS page](https://www.university-365.com/lips). For the CI-First doctrine, see the [CI-First page](https://www.university-365.com/ci-first). For the full SL-OS framework, see the [SL-OS page](https://www.university-365.com/slos).
APPLY IT (domain-tagged): In the Career and Finance domain, book a ninety-minute session this week and write a one-page mission ladder: the ten-year objective, three rungs you can complete in the next twelve months, and the evidence each rung would produce.
SUMMARY
MINDMAP SKELETON
Center: The Infinity Machine, by Sebastian Mallaby.
Branch 1, Chess and Games (Chapters 1 to 3). Leaves: intuition over calculation, Elixir Studios, charisma and control.
Branch 2, Founding DeepMind (Chapters 4 to 6). Leaves: induction, reinforcement learning, the Atari result.
Branch 3, Investors and Google (Chapters 7 and 8). Leaves: Founders Fund, the six hundred and fifty million dollar acquisition, the promise.
Branch 4, AlphaGo (Chapters 9 and 10). Leaves: policy network, value network, Move 37, the ethics board.
Branch 5, Health and the Transformer (Chapters 11 and 12). Leaves: Streams, the NHS inquiry, self-play, attention.
Branch 6, The Language Detour (Chapters 13 and 14). Leaves: the language blind spot, Project Mario, the diluted charter.
Branch 7, AlphaFold to ChatGPT (Chapters 15 to 17). Leaves: CASP14, the race, Bard, the merger.
Branch 8, Safety and Comeback (Chapters 18 to 21). Leaves: the pause letter, Bletchley, Gemini, the Nobel Prize.
Reconstruction prompt: "Draw a mindmap with this structure. Center node at top, eight branches arranged vertically below, three or four leaves per branch, soft modern palette, clean lines, white background."
Chapter 1: Destiny
Hassabis grows up in North London with a Chinese Singaporean mother and a Greek Cypriot father, playing chess before he can remember choosing to. Leonard Barden tells his father that the six-year-old is the best he has seen, and for the next decade weekends are spent in church halls across England. By the time he reaches Cambridge he is already a chess master, and at fourteen he answers an advertisement inviting readers to reinvent Space Invaders for a job at Peter Molyneux's Bullfrog.
Mallaby's analytic move here is to treat the chess career as an education in the limits of calculation. Young Demis learns that the players who beat him are not computing more; they are seeing differently. That lesson becomes the intellectual foundation of everything he later builds, and the book is careful to show it forming before it becomes a theory.
The chapter's decisive scene is not a tournament but a book. Reading Orson Scott Card's Ender's Game during his doctorate, Hassabis finds a fictional character whose situation matches his own sense of himself: a gifted young person asked to carry a responsibility that no one else can carry. Mallaby uses the identification without endorsing it, and the reader is left to decide how much of the later mission was designed and how much was a story he had already chosen.
Chapter 2: Deep Philosophical Questions
Cambridge in 1994 is hostile to the idea that intelligence can be built. The dominant view holds that neural networks are a dead end and that symbolic logic is the serious path. Hassabis argues with his teachers, reads Gödel, Escher, Bach, and concludes that the symbol-grounding problem is the real obstacle: a system that defines symbols in terms of other symbols never touches the world.
Molyneux tries to keep him by writing a cheque for five hundred thousand pounds, roughly one point seven million dollars in today's money. Hassabis does not cash it. Mallaby uses the refusal to establish the pattern that recurs at every later stage: he will take resources, but he will not be owned.
The chapter introduces Shane Legg and David Silver, the two people who will matter most to the technical project. Legg is thinking about how to measure intelligence, and Silver about learning from experience. Between them, Mallaby shows, the three students already hold the components of DeepMind: a definition of the target, a method for reaching it, and someone willing to start a company.
Chapter 3: The Jedi
Hassabis founds Elixir Studios in 1998 and promises a game, Republic: The Revolution, that will simulate politics, crowds and revolution at a scale no studio has attempted. London venture capital behaves badly: meetings conducted in pubs, an offer to become a currency trader instead, and eventually a term sheet demanding half the company. He refuses, and funds it another way.
The chapter is where Mallaby first names the quality that will define the rest of the book. David Silver calls it the Jedi mind trick: the ability to make people believe a future so vividly that they reorganize their lives around it. Elixir's demo wins best upcoming game at a trade show after a hardware failure, because Hassabis talks over the broken machine and makes the audience see the game he intends to build.
Elixir collapses. Silver burns out and leaves. The analysis Mallaby offers is the most useful passage in the book for anyone who leads creative teams: charisma oscillates between inspiration and control, and the oscillation itself is what hooks people. Hassabis's stated horror of controlling others is genuine, and it is not the same thing as being unable to.
Chapter 4: The Gang of Three
By 2009 Hassabis is finishing his doctorate and feeling the entrepreneurial pull again. Molyneux tries once more, and Hassabis tells the emissary that he wants to be the person who solves AI. He begins meeting Legg regularly, and they argue about timing: Legg thinks it is too early to raise money, and Hassabis thinks academia is too slow.
Mustafa Suleyman enters, and Mallaby gives him a full portrait rather than a supporting role. The son of a Syrian taxi driver and an English nurse, raised in a devout household, he arrives with a record of activism and an appetite for scale. The three founders are, in the book's reading, a complete company: the scientist, the theorist and the operator.
The name DeepMind is chosen to point three ways at once, at deep learning, at Deep Blue and at Deep Thought. Mallaby treats that triple reference as a small confession of the founders' ambition, and as a warning about how much of the plan existed as narrative before it existed as engineering.
Chapter 5: Founding DeepMind
DeepMind is founded in 2010 and launches at the Singularity Summit, a conference whose attendees Mallaby describes with evident relish. The company has no product, no revenue and a mission that most investors consider unserious. Hassabis raises money by convincing people that the mission is the asset.
The early office in Russell Square is described as a library, with people speaking in whispers and phone calls conducted in the stairwell or beside the servers. Mallaby uses the physical detail to make a structural point: DeepMind was designed as a research institute that happened to be a company, and that choice determined what it would later be good and bad at.
The chapter also establishes the talent strategy that made the lab formidable. Hassabis hired neuroscientists, physicists and mathematicians before the field had a name for what they would do, on the theory that general intelligence would require general intelligence. Mallaby treats this as a genuine insight rather than a recruiting story, and the later chapters vindicate him.
Chapter 6: Atari
In 2012 Alex Krizhevsky's ImageNet result detonates the field, and Hinton's trio sells itself to Google for forty-four million dollars after two months. The episode sets the terms of the next decade: talent, compute and speed would decide who led.
Vlad Mnih's Atari agent is the chapter's centerpiece. A deep network reads the pixels, Q-learning estimates the value of actions, and one system learns dozens of games without being told the rules. When the results are shown, the room goes silent, and Mallaby captures why: the same code had learned different things.
The analysis is precise about what the result proved and what it did not. It proved that a general learning method could master diverse tasks. It did not prove that the method would scale to the open world, or that playing well was the same as understanding. Those distinctions matter later, when the language models arrive and DeepMind's agents look less general than they seemed.
Chapter 7: Thiel Trouble
Peter Thiel's Founders Fund becomes DeepMind's most consequential early investor, and Mallaby reconstructs the courtship with dry comedy. Hassabis hooks Thiel with a chess observation about the bishop and the knight, two pieces of equal value with entirely different powers, and the conversation moves from there.
The due-diligence visits unsettle the founders. Palantir, a defense contractor, and a military robotics company with a gun on treads are presented as peers. Suleyman, whose politics are explicitly anti-war, comes away frightened about what an investor with those interests might ask the company to build.
The relationship deteriorates when the science requires more capital than venture investors want to supply for an open-ended mission. Mallaby's point is structural rather than personal: venture capital is a poor instrument for a project whose payoff is a scientific result rather than a market position, and DeepMind's near-death experience taught Hassabis that lesson permanently.
Chapter 8: Get Google
The Google acquisition is negotiated through a series of demonstrations and a single decisive inspection. Jeff Dean asks to see the Atari code because a demo can be staged and a codebase cannot. What follows is a bidding process that ends in January 2014 with a price of six hundred and fifty million dollars.
Mallaby makes the personal arithmetic explicit, because it explains a great deal about what came later. Hassabis receives a hundred and thirty-six million, Suleyman thirty-four, Legg twenty-nine. The founders become wealthy, and the company becomes able to pay researchers at a rate no independent lab could match.
The chapter's real subject is the promise. DeepMind asked for, and believed it had received, commitments about how its technology would be used. Mallaby shows the commitment being made in good faith and then shows how little institutional machinery was created to enforce it. The gap between the two is the seed of the book's final argument.
Chapter 9: Intuition
AlphaGo is built by David Silver's team from two networks and a search. The policy network narrows the possibilities the way a strong player's eye does. The value network judges positions. The search explores. In October 2015 the system beats Fan Hui, the European champion, and in January 2016 Nature puts it on the cover.
The match against Lee Sedol in Seoul is the book's most sustained set piece. Mallaby has the room, the timings and the emotional register: the Korean press, the million-dollar prize, the professionals who expected a rout in the other direction. AlphaGo wins the first three games.
Then Move 37. The move looks wrong to almost everyone, including the commentators, and it turns out to be decisive. Lee's comment, that he thought AlphaGo must be creative, is the human admission the whole book has been building toward. In the fourth game Lee wins with a move of his own, and Mallaby resists the temptation to turn it into a moral. He simply records that three years later Lee retired, saying he no longer found joy in the game.
Chapter 10: Out of Eden
The Google relationship works, for a while. Hassabis keeps DeepMind's London culture, its separate identity and its privileged access to compute, while Google gets the prestige of owning the most interesting lab in the field. Mallaby documents the arrangement's benefits and its fragility.
The ethics and safety board is the chapter's through line. Its members are serious people with genuine concerns, and they disagree about fundamentals: whether AI risk is real, whether a lab should publish, whether a single company should hold the technology. Those disagreements are not resolved, and the board's authority never becomes operational.
The founding of OpenAI is presented as the break. Musk and Hoffman attend a DeepMind gathering, learn the plans, and launch a competitor months later. Mallaby does not need to editorialize; the sequence is enough. The singleton strategy was over before the technology that justified it existed.
Chapter 11: P0 Plus Plus
Suleyman's Applied division is the most human part of the book. It builds Streams, an alert system for acute kidney injury, at the Royal Free Hospital in London, and the early results are extraordinary: patients who might have died are seen in time. Mallaby lets the clinicians describe what it felt like, and one of them calls it the best moment of his career.
The project then collides with politics. A data-sharing agreement drafted on a standard NHS template becomes a national scandal, two official inquiries follow, and neither finds that DeepMind abused data. Mallaby's account is scrupulous about this: the failure was one of transparency and political framing, not of technical conduct.
The chapter's title comes from an internal confusion about a product code that nobody could explain, and Mallaby uses it as a small emblem of the division's problem. Suleyman launched too many things, changed direction too often, and left his team unable to say what it was building. The health work was eventually absorbed elsewhere. DeepMind's most direct attempt to do good in the world ended as a management failure rather than an ethical one.
Chapter 12: The Agent and the Transformer
Two papers change the field within months of each other. AlphaGo Zero and then AlphaZero show that self-play alone can surpass human knowledge in bounded domains, and the result is beautiful. Then the transformer paper appears, from Google researchers, and shows that attention over sequences could learn from unlabeled text at a scale nobody had attempted.
Mallaby's account of the divergence is the most important analytical passage in the book. DeepMind's culture was built around agents that act and search. The transformer suggested a different route to generality: absorb the written world and predict what comes next. The lab understood the paper and did not reorganize around it.
Ilya Sutskever leaves Google for OpenAI, and Mallaby uses the departure to mark the moment. One of the people closest to the transformer believed the future was language models, and the institution that had the best claim on his work let him go. The strategic consequences take another decade to become visible, but the decision happens here.
Chapter 13: On Language and Nature
This chapter is the book's confession. Mallaby describes years of pub conversations in North London in which Hassabis explained why language models were not the path to general intelligence. They learn from the surface of the internet. They do not build a model of the world. They cannot invent relativity.
The objections are technically serious, and Mallaby reports them fairly. He then reports what happened. GPT-2, GPT-3 and GPT-4 arrived in sequence, each one closer to professional competence, and the argument that language models were a distraction became harder to maintain. Hassabis admits to the author that he was slow to see it.
The chapter also shows why the mistake was rational for him. DeepMind's identity depended on the superiority of search and agency, and the lab had spent years proving it in games and biology. A leader who abandons his own demonstrated advantage on the strength of a paper is asking his organization to disbelieve its own evidence. Mallaby's portrait is of a judgment error that was also a coherent strategy.
Chapter 14: Project Mario
Between 2015 and 2017 Hassabis and Suleyman negotiate with Google for a structure that would protect DeepMind's mission from commercial pressure. The discussions are given the internal name Project Mario, and Mallaby reconstructs them with unusual access: the meetings, the term sheets, the competing plans.
The proposal that emerges would have made DeepMind a separate global interest company with fifteen billion dollars of funding and an independent board. For a moment it looks as though the founders will get what they asked for. Then the structure is diluted, the independence becomes advisory, and the governance that was supposed to outlast the founders does not survive the year.
Mallaby's verdict is measured and devastating. The founders compromised their stated values, and they were also right that their youthful ideals had been unrealistic. A technology this consequential was never going to be governed by one company's charter. The chapter leaves the reader with the harder question: if the charter was naive, what would have worked?
Chapter 15: Fermat for Biology
Hassabis has wanted to solve protein folding since Cambridge, and Mallaby explains why the problem suited DeepMind so well. The search space is astronomical, the objective is unambiguous, and the ground truth can be measured in a laboratory. It is Go with a scientific payoff.
AlphaFold 2 scores 92.4 in the 2020 CASP assessment, more than fifty percent better than any previous result, and the organizers suspect cheating before they confirm the result. Mallaby tells the story through the crystallographers who checked it, which makes the verification as dramatic as the prediction.
The chapter is where the book's title pays off most clearly. A classical learning algorithm found patterns in a space that no human could search, and the answer was useful. Mallaby is careful to note that the achievement required years of domain work and a large team, and that the Nobel Prize recognized a method as much as a discovery.
Chapter 16: The Power and the Glory
The mood changes. DeepMind has Gopher, then Sparrow, and a credible path to a conversational product. Google has the infrastructure. The pieces are in place, and Mallaby documents the internal confidence that the lab will define the next era.
Then OpenAI releases ChatGPT. The chapter's account of the weeks that follow is the most painful in the book: the internal realization that a competitor had shipped the product DeepMind had been preparing, the scramble, the merger negotiations that create Google DeepMind in April 2023, and the loss of the separate identity the founders had spent a decade protecting.
Mallaby is good on the psychology. Hassabis is not defeated. He is, in the author's phrase, in his element, because a tournament is a format he has understood since childhood. The book does not let him off the hook, but it does explain why the next phase suited him better than the previous one.
Chapter 17: RaceGPT
ChatGPT's release is presented as a case study in technological determinism. OpenAI did not plan to start a race it could not control, and the product made gradualism impossible. Anthropic announced Claude two weeks later. Microsoft committed ten billion dollars. Every lab's timeline compressed.
Google's response is the chapter's cautionary tale. Bard was announced before it was ready, demonstrated badly, and the stock fell. Mallaby reports the internal fury and the defections that followed, and the sense among researchers that the company had been humiliated into action.
The analysis is that the race rewarded shipping over understanding, and that everyone in it knew this and shipped anyway. Mallaby does not present this as hypocrisy. He presents it as a coordination problem, and the book's later chapters show how completely it defeated the safety institutions that had been built to solve it.
Chapter 18: "We're Cooked"
The safety debate arrives in public. Hinton and Bengio speak about risk, the open letter calling for a pause circulates, and Hassabis signs a statement about risk while declining to sign the moratorium. Mallaby asks him why, and gets an answer about timing: a six-month pause would not help, and being inside the process was the only way to influence it.
The Bletchley Park summit in November 2023 is the high-water mark. Twenty-eight countries, a declaration, a new institute in Britain and one in the United States. Mallaby lets a participant describe the atmosphere, and the description is of talking points rather than commitments.
Then OpenAI's board fires Sam Altman, and five days later reinstates him. Mallaby treats the episode as the moment the safety argument lost its institutional vehicle. If the one company structured to prioritize safety could not remove its own chief executive, no governance mechanism was going to slow the race. The chapter ends with the sentence that the rest of the book has been earning: Hassabis had no alternative but to race forward.
Chapter 19: Step by Step
The industry moves to reasoning. Google DeepMind's Gemini 1.5 extends the context window to a million tokens, which allows a model to hold an entire book, or an entire codebase, in working memory. Mallaby connects the advance to Hassabis's neuroscience: working memory is what lets a mind hold several steps of an argument at once.
AlphaProof wins a silver medal at the International Mathematical Olympiad in 2024, and the result is a genuine synthesis of deep learning and reinforcement learning. Mallaby uses it to show what DeepMind is still uniquely good at, and then shows OpenAI's o1 preview arriving with a reasoning product before Google ships one.
The chapter's judgment is severe. DeepMind had the deepest reinforcement learning expertise in the world and lost the reinforcement learning stage of the language race. Mallaby does not blame a single decision. He describes a pattern: research excellence without product urgency, repeated across three consecutive transitions.
Chapter 20: Comeback, and Beyond
Google DeepMind rebuilds. Gemini 2.5 outperforms its rivals on most benchmarks in the spring of 2025, GPT-5 often lags in blind comparisons, and Gemini 3 sets new standards in coding and multi-step planning in November 2025. The Gemini app's user numbers nearly double in seven months.
Mallaby is careful not to declare a winner. The DeepSeek shock, the arrival of capable Chinese models at a fraction of the cost, and the permanent possibility of a competitor's next release keep the outcome open. What has changed is that Hassabis now runs a lab that competes on product as well as on science.
The chapter's most revealing material is about motive. Hassabis tells the author that he wants to cross the Rubicon, and describes a future in which every person has a powerful AI assistant as a right, and the machines cooperate to look after the species. Mallaby records the vision without mocking it, and without resolving whether it is a plan or a hope.
Chapter 21: Turing's Champion
The epilogue returns to the pub. It is December 2024, the Nobel has been awarded, and Hassabis brings the medal in a leather box to show the author. He talks about the Nobel Foundation's guest book, and about sitting in the room where the prize was founded.
The intellectual argument of the epilogue is with Roger Penrose, who claimed that classical computation cannot produce conscious understanding. Hassabis's Nobel lecture made the opposite case in the form of a conjecture: any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm. Mallaby explains why the claim matters. It is a statement about the universe, not just about computers.
The book ends on a note that is neither triumph nor defeat. Hassabis says that the moment should feel amazing and does not, that the rush has made the achievement hard to inhabit, and that he has made his peace with the mess. Then he says he is still optimistic. Mallaby lets him have the last word, and the reader is left holding the contradiction.
IN PRACTICE
1. Write the mission ladder before you write the pitch.
Hassabis's advantage was not the size of his ambition but the sequence of demonstrable rungs beneath it: Atari, then Go, then protein folding. Each rung produced evidence that bought patience and money for the next one.
Action: Draft a one-page ladder with a ten-year objective, three twelve-month rungs, and the artifact each rung will produce.
2. Separate the promise from the paperwork.
DeepMind asked Google for commitments about weapons and surveillance, and did not build the institutional machinery that would have made those commitments binding. The lesson is not that the founders were naive. It is that an unwritten condition is not a condition.
Action: List every promise attached to your current funding or employment and mark which ones exist in a signed document.
3. Treat the technology you did not choose as a competitor.
DeepMind's research culture made language models feel secondary, and that judgment cost it the most valuable product category in the industry. Strong prior evidence can still be a strategic error when the field's direction changes.
Action: Name the approach you have dismissed, and write the specific result that would force you to change your mind.
4. Build the boring capability that makes the exciting one possible.
AlphaFold worked because the team spent years on data pipelines, evaluation and domain detail before the model was good. Mallaby's account gives the unglamorous work its proper weight.
Action: Identify the least visible task in your project and give it a named owner and a deadline this month.
5. Design for the organization you will become.
DeepMind's founding structure assumed a small, independent lab. Every later crisis came from operating a large institution with the governance of a small one. Structures written for the current size fail at the next one.
Action: Review your decision rights and ask which ones will not survive a doubling of headcount.
6. Decide in advance what you will not trade.
Hassabis and Suleyman negotiated hard and still gave up the charter. The difference between a compromise and a surrender is whether you wrote the line down beforehand.
Action: Write three non-tradable commitments and date the document. Re-read it before any negotiation.
7. Measure the cost of the race on the people running it.
The book is full of exhaustion: Silver after Elixir, Suleyman after the health crisis, researchers after each defection. Sustainable pace is a strategic variable, not a welfare extra.
Action: Ask each person on your critical path what they have stopped doing to keep up, and record the answers.
QUIZ: TEST YOUR UNDERSTANDING
1. Mallaby calls the book's subject an infinity machine. What does the phrase mean technically?
It refers to a system that performs induction, finding patterns in a search space too large for any human to enumerate. Classical AI relied on deduction from rules, which fails on novel cases.
2. Why did AlphaGo's architecture combine a policy network, a value network and a tree search?
The policy network proposes plausible moves quickly, the value network judges positions, and the search explores ahead. Mallaby presents the combination as the practical form of intuition plus deliberation.
3. DeepMind invented much of the intellectual groundwork for the transformer. Why did it not lead the language model market?
Its research culture prioritized agents, games and scientific problems, and its leadership judged that language models would not produce genuine understanding. That judgment delayed product work while OpenAI shipped.
4. What happened to the ethics board and the independence charter that were meant to constrain Google's use of DeepMind's technology?
Neither became operational in the way its architects intended. The board's authority was never clearly established, and the charter was diluted during the 2017 negotiations and did not survive.
5. Transfer question: you lead a research team inside a large company and your funding depends on a quarterly product cadence. Using the book's evidence, name one structural safeguard you would negotiate before accepting the money, and explain what would make it enforceable.
A defensible answer names a written commitment with a named enforcement body, a defined trigger, and a cost for breach. Unwritten assurances and advisory boards failed at DeepMind because neither carried a consequence.
How many did you get right? Which ones surprised you?
CAN THIS BOOK ESSENTIAL REPLACE THE ORIGINAL?
No. This Essential gives you the argument, the sequence of decisions and the analytical frame. The book gives you the scenes, the dialogue and roughly four hundred and eighty pages of reporting that make the judgments credible, and it is worth reading in full if you intend to act on them.
QUOTES
"It doesn’t collapse our brains. We can converse. So first order logic is clearly not the whole story."
"But AI may be the right description language for biology, because biology is so messy, emergent, dynamic, and complex."
"Intelligence measures an agent’s ability to achieve goals in a wide range of environments."
"I already had the kernel of the idea for what eventually became DeepMind."
"For every game, the same agent had learned something completely different. People were just blown away. It was a turning point."
"Any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm"
"My Nobel medal is my most valuable possession."
"I was on the stage at the Nobel ceremony, and I was thinking I wouldn’t have swapped this for any amount of money."
"There is too much negativity about AI. People need to see the benefits. That changes the conversation."
"So you have to be adaptable. You have to move from idealist to realist, but hopefully still with your values."
"There should be a new human right to AI assistance."
"Recognize that it’s going to be messy, and I’ll just have to do the best I can."
AUTHOR'S EXPERTISE
Sebastian Mallaby is the Paul A. Volcker Senior Fellow for International Economics at the Council on Foreign Relations, a former contributing editor of the Financial Times, a contributing columnist for The Washington Post, and a two-time Pulitzer Prize finalist. He co-hosts the CFR podcast The Spillover and writes the Substack Machine Readable.
He is the author of six books, including More Money Than God, a history of hedge funds that became a standard reference, The Man Who Knew, which won the Financial Times and McKinsey Business Book of the Year Award, and The Power Law, a study of venture capital that shaped how a generation of investors describe their own industry. His method is consistent across all three: prolonged access, reconstructed scenes, and a willingness to let the subject's own account stand next to the evidence.
That method suits this subject particularly well. Mallaby began negotiating access in November 2022, before the field became the center of public attention, and his interviews run through the period in which DeepMind's strategy was overtaken. The book was shortlisted for the Financial Times and Standard Chartered Business Book of the Year 2026. Readers who want the technical literature should go elsewhere; readers who want to understand how a scientific institution makes commercial decisions will find no better account.
Mallaby writes as a business historian rather than as a technologist, and the choice is deliberate. He is more interested in the negotiation than in the architecture, and the book's judgments about AI are consequently judgments about institutions: who decides, under what constraints, and with what recourse.
RESOURCES
The book: The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence, by Sebastian Mallaby, published by Penguin Press in March 2026. https://www.penguinrandomhouse.com/books/752231/the-infinity-machine-by-sebastian-mallaby/
Companion reading from the same author: https://www.penguin.co.uk/books/309693/the-power-law-by-mallaby-sebastian/9780141988948 , on how venture capital actually allocates risk, which explains much of what happens to DeepMind between 2010 and 2014.
Adjacent books in the U365 library: Superintelligence by Nick Bostrom, for the conceptual case that worried DeepMind's founders before they had evidence; Empire of AI by Karen Hao, for a more critical account of the industry's incentives; Co-Intelligence by Ethan Mollick, for the practical question of how individuals should work with these systems.
Primary sources for the achievements discussed: the AlphaGo paper in Nature (2016), the AlphaFold 2 paper in Nature (2021), and the Nobel Prize in Chemistry 2024 scientific background, which together give the technical substance behind the narrative.
NEXT STEPS
Write your mission ladder this week: A ten-year objective, three rungs for the next twelve months, and the evidence each rung will produce. Keep it to one page.
Audit your conditions: List every promise attached to your funding, your employer or your partnership, and mark which ones exist in writing. The unwritten ones are the ones that will be renegotiated without you.
Name your dismissed option: Write down the approach you have decided is not worth pursuing, and the specific result that would change your mind. Review it quarterly.
Protect the invisible work: Give the least glamorous task on your critical path a named owner and a date. AlphaFold's success was built there.
Write your three non-tradables: Date the document, and re-read it before any negotiation. The purpose is not rigidity. It is to know, in advance, what a compromise would actually cost.
Read the primary papers: The AlphaGo and AlphaFold papers are more accessible than their reputation suggests, and reading them beside Mallaby's narrative will change how you judge the claims you hear about AI.
U365'S RECOMMENDATIONS TO LEARN MORE
This list is curated for readers who want the argument in other formats. Links were checked and verified as of 2026-11-07. Where a source sits behind a paywall, the free alternative is given.
Official learning resources
Penguin Random House book page : publisher description, praise and editions.
Sebastian Mallaby at the Council on Foreign Relations : author profile, books and podcast.
Machine Readable : the author's Substack, where he writes about AI economics and the book's themes.
Video tutorials and channels
CFR Fellows' Book Launch Series: The Infinity Machine : Mallaby presents the book and takes questions on DeepMind, Google and the race.
Sebastian Mallaby: The AI Race Nobody Can Win : Foreign Affairs interview on competition, the AI trilemma and what governments can actually do.
Accelerating Scientific Discovery with AI : Hassabis's own lecture on the scientific method behind AlphaFold, delivered at Cambridge.
Mallaby presents the book and answers questions on DeepMind, Google and the race for AGI, by Council on Foreign Relations, Apr 13, 2026, 1:02:49
Foreign Affairs Interview on competition, the AI trilemma and the limits of national policy, by Foreign Affairs, Jul 02, 2026, 53:53
Hassabis on the research method behind AlphaFold and science at digital speed, by CambridgeComputerLab, Mar 24, 2025, 1:04:10
Written tutorials and deep-dive articles
The story of the man who changed the world : the Guardian's skeptical review, useful precisely because it disagrees with the book's scale claims.
Kirkus review : a short, reliable summary of the book's scope and argument.
A dispatch from the last sane era of AI : a critical essay on what the book reveals about AI governance.
Demis Hassabis and John Jumper awarded the Nobel Prize in Chemistry : the primary account of the award and the AlphaFold program.
Community and social
Sebastian Mallaby on X : the author's own account, where he posts about the book and AI policy.
AI Book Review : an independent reviewer's assessment, useful for a second opinion on the book's strengths and limits.
Community walkthrough: the LessWrong review , a detailed critical reading that focuses on the scientist versus engineer tension in the book.
Resources on X
Dedicated X channels:
Demis Hassabis : the subject's own account, used sparingly and mostly for research announcements.
Google DeepMind : first-party announcements on Gemini, AlphaFold and the lab's research.
Sebastian Mallaby : the author's commentary on AI economics and the book's reception.
X posts with video content:
Sebastian Mallaby on writing The Infinity Machine : the author's pinned post explaining how the access was negotiated and what changed when ChatGPT launched.
Curation policy: every link above was opened and checked on 2026-11-07. Videos are listed only where the channel is the publisher, the author or the subject's own institution. Reviews are included whether or not they agree with the book.
IMPORTANT NOTICE
This Book Essential is an original summary and critical analysis of The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence by Sebastian Mallaby (Penguin Press, 2026). Short quotations from the book are attributed and cited for purposes of criticism, review, and education. All rights in the original work belong to its author and publisher; this Essential is not a substitute for the book: read the original at [Penguin Random House](https://www.penguinrandomhouse.com/books/752231/the-infinity-machine-by-sebastian-mallaby/).
This book is part of University 365's learning library. Explore INSIDE, our publications, and our programs. The best summary is not a substitute for the book. Read the original. Discuss this book with a U.Coach.
Note: figures, prizes and product claims in this field change quickly. Information in this Essential was verified against the book and public sources as of 2026-11-07.
Comments