The infinity machine

The bet behind the book

Sebastian Mallaby has spent his career writing biographies of men who shape money — Alan Greenspan, hedge-fund titans, the venture capitalists of The Power Law. With The Infinity Machine he turns to someone who shapes something stranger: the architecture of machines that may one day match the human mind. Demis Hassabis, co-founder and chief executive of DeepMind, is the protagonist. The promise of the book is that telling Hassabis’s story properly tells the story of how artificial intelligence stopped being a research curiosity and became the central technology of our age.

Mallaby mostly delivers, with caveats. The book is a dual narrative: a personal arc from chess prodigy to Nobel laureate, and a chronicle of DeepMind’s two decades from a London startup to a Google subsidiary at the centre of the AGI race. If you want to understand why Google Brain was eventually merged into DeepMind, why AlphaFold won a chemistry Nobel, why the company missed the transformer, and why Hassabis still believes reinforcement learning will return to the throne — this is the right book.

It is also, repeatedly, a fan letter. Mallaby admires his subject, and his prose admires every other subject in the room a little too breathlessly to qualify as straight reportage. Read it for the access and the technical history. Read something else for the harms.

The chess kid

Hassabis grew up in North London to a Greek Cypriot father and a Singaporean Chinese mother. The father once sold toys from a beaten-up van; the mother had grown up in Singaporean poverty, orphaned at points along the way. The boy picked up chess at age four. He was beating grown men shortly after. By thirteen he was a chess master and ranked second worldwide for his age.

The chess chapters are some of Mallaby’s best. At his tournaments organisers fixed panels beneath the tables to stop the boys taking shots at each other’s shins. His father told him to do his best, and the boy took the instruction so literally that he treated every game as a near-death stress test. That pattern — total exertion as the only proof of effort — never quite leaves him. By his early thirties he is running progress meetings at three in the morning across multiple time zones, holding the same posture toward AGI that he once held toward openings and endgames.

The intellectual seeds matter too. Mallaby spends real time on Douglas Hofstadter’s Gödel, Escher, Bach, which Hassabis read as a teenager and which seems to have set him on a permanent course: intelligence is patterns, patterns can be discovered, discovery is the highest human activity, and the mind is therefore in principle buildable. After a stretch as a teenage programmer at Bullfrog Games, where he worked on Theme Park in Peter Molyneux’s studio, Hassabis went to Cambridge, founded a games studio called Elixir, and then turned to neuroscience for his PhD at University College London. He wanted to understand how brains stored memories, because he wanted to build machines that could too. In 2010 he founded DeepMind with Shane Legg and Mustafa Suleyman. The mission, depending on the audience: cure cancer, solve intelligence, find God’s algorithm. The pitch worked on Peter Thiel, who became the company’s first major investor — a detail the book treats with appropriate awkwardness, given Thiel’s later turn into MAGA fundraising.

Games as the proving ground

DeepMind’s early reputation was built on games. Atari first — a single neural network learning to play dozens of arcade titles from raw pixels, with no human-engineered features. Then chess and shogi and Go, with the AlphaZero family. The Go match against Lee Sedol in Seoul in 2016 is the book’s first big set piece. The global audience reportedly topped two hundred million. AlphaGo won four games to one. After his third loss Lee said, “I, Lee Se-dol, lost, but mankind did not.” The line has been quoted half to death since but it still lands.

Why Go mattered: chess fell to brute-force search in 1997 with Deep Blue. Go has too many legal positions for that approach to work. AlphaGo had to learn something closer to intuition, training itself through self-play with reinforcement learning until it understood the game on its own terms. The system’s now-famous Move 37 in game two is the moment Mallaby treats as a hinge. A machine had not just calculated faster than a master. It had played a move that masters initially thought was a mistake until they realised it wasn’t.

The chapter is good because it explains what reinforcement learning actually does, not just what it claims to do. If you’ve read previous explanations of AlphaGo and bounced off them, this version will probably stick.

The protein-folding moment

AlphaFold is the strongest argument the book makes for AI as civilisational good rather than venture-capital bubble. Predicting the three-dimensional shape of a protein from its amino-acid sequence had been an open problem in biology for half a century. AlphaFold2 didn’t solve it perfectly — its predictions still aren’t always accurate enough for direct practical use, as some careful reviewers have noted — but it collapsed the cost of getting a usable structure from years of crystallography to minutes of inference. Hassabis and his colleague John Jumper shared the 2024 Nobel Prize in Chemistry for the work.

Mallaby’s narration of AlphaFold’s development is the book at its best. It shows real research as it actually unfolds: false starts, abandoned approaches, fights about whether to keep using reinforcement learning or pivot to attention-based architectures, the slow realisation that the problem might actually be tractable. This is the kind of work — quiet, technical, undeniably useful — that makes the case for AI better than any podcast appearance ever will. It also makes the strongest implicit argument for DeepMind’s culture, which prioritised scientific publication over product velocity for a long time. The contrast with OpenAI’s ship-first ethos is sharp, and Mallaby uses it well.

When the science is this good, awe is appropriate. The AlphaFold chapters are the part of the book where Mallaby’s reverence reads as proportionate rather than naïve.

When the ground shifted

DeepMind missed the transformer. That’s the uncomfortable centre of the book’s second half. The 2017 paper that introduced the architecture came out of Google Brain, not DeepMind. The team that turned transformers into a product wasn’t even at Google — it was OpenAI, then a small lab Hassabis viewed as more enthusiast than rival.

When ChatGPT shipped in November 2022, DeepMind was caught flat-footed. Hassabis had bet on reinforcement learning as the path to general intelligence. Sam Altman bet on scale. Altman won the press cycle, then the user numbers, then the funding rounds. Mallaby is honest about the panic that followed inside Google: the merger of Google Brain into DeepMind, the rushed Bard launch, the awkward rebrand to Gemini, the reasoning-model arms race that followed, the realisation that LLMs had eaten the field.

He is also good on the governance fights nobody outside Google saw. The internal negotiation that DeepMind insiders called Project Mario — an attempt to get Google to ringfence DeepMind’s autonomy and route any AGI-grade outcome through some independent oversight — collapsed under the weight of competitive pressure once OpenAI’s lead became visible. The lesson, which the book lets the reader draw without quite stating it, is bleak: charters and oversight boards are durable until the moment a competitor ships, at which point they become an obstacle to be renegotiated.

What Mallaby is less honest about is whether Hassabis was actually right on the technical bet. The book gives him the last word: maybe LLMs are a detour, maybe true general intelligence still needs the kind of model-based reinforcement learning that built AlphaGo. He may be right. He may also be a man explaining away the moment he got beaten. Mallaby can’t seem to decide, and so leans toward the more flattering reading.

The Sam Altman portrait is unflattering and probably accurate. He appears as a charming opportunist who ships before competitors are ready and uses safety rhetoric as marketing surface. If you’ve read Karen Hao’s Empire of AI on OpenAI, the picture here will rhyme. Altman as a careful student of Robert Caro’s biographies of Lyndon Johnson — a detail Mallaby drops in passing — tells you most of what you need to know.

The Suleyman problem

The book’s most uncomfortable mid-section is the slow exit of Mustafa Suleyman, DeepMind’s third co-founder. Suleyman ran the applied projects, including a partnership with the UK’s National Health Service that he hoped would prove AI could improve real institutions in real time rather than waiting for AGI. The intent was good. The execution was chaotic — patient-data handling that drew regulatory scrutiny, internal complaints about Suleyman’s management style, eventual reports of bullying that led to his sidelining.

Mallaby treats this with restraint, which is to say less critically than it deserved. Suleyman moved on to Inflection, then to Microsoft. The pattern — sincere about social transformation, destructive in the rooms where the transformation is supposed to happen — is one the AI industry now has many examples of. The Suleyman section is the closest the book gets to acknowledging that “build it for humanity” can mask the same dysfunctions as any other Silicon Valley playbook.

Where the book is weakest

Three problems.

First, Mallaby cannot stop himself from writing in the voice of a fan. People do not say things in this book. They confess, fret, vow, declare, reflect. The Guardian’s reviewer counted enough of these dialogue tags to suggest Mallaby was trying to dress up otherwise dull conversations. He’s right. Tech executives in their own words are often less interesting than they think, and a stylist this aggressive about adverbs is usually compensating.

Second, the book accepts the AGI frame on the AGI builders’ terms. When Hassabis says, “Doing science is, sort of, like reading the mind of God,” the appropriate journalistic move is to note the hedges — sort of, kind of — and ask what the speaker actually means. Mallaby quotes the line approvingly. The hedges signal a man who knows the rhetoric outruns the reality and isn’t quite willing to commit. A more skeptical biographer would have made them count for something. The same pattern shows up around the book’s headline claim that AI represents the most consequential change in human cognition since the rise of abstract thought. It does not. Agriculture, language, and writing all have stronger cases. Mallaby states the claim as if it were uncontroversial.

Third, the harms get cursory treatment. Training data scraped without consent, the energy and water demands of the new datacentres, the labour displacement already happening in translation and customer support and copywriting — these come up briefly and get waved away. The book is interested in the boardroom drama of who runs the AGI charter, not in the question of whether anyone outside the boardroom should have a say.

The Bill Gates aside the Guardian flagged is a small but telling tell. Mallaby writes, in defence of tech CEOs, that for every Trump-aligned figure in the industry there is also a Gates. The line was probably written before the Epstein-correspondence release. It now reads as a reminder that the genre of tech-CEO hagiography can age in months.

What’s missing

For a book about the AI race, there is surprisingly little on Anthropic. Dario Amodei and Daniela Amodei split from OpenAI in 2021 to build the company that ships Claude, the model many developers now consider the most consistently useful for technical work. Anthropic is the third major frontier lab. Mallaby gives them paragraphs where they deserve chapters. Several reviewers have flagged this and they’re right.

There is also very little on China. DeepSeek’s late-2025 releases get a brief mention; the broader arc of Chinese AI labs, the CHIPS Act, and the geopolitics of compute access do not. For a book that styles itself as the story of the quest for superintelligence, the omission of roughly half the world’s AI capacity is strange.

Finally, the doomer side of the AGI conversation gets short shrift, which is odd given that Shane Legg, DeepMind’s third co-founder, has been talking publicly about extinction risk for over a decade. Mallaby acknowledges that Legg cares about safety and moves on. He does not seem to take the arguments seriously enough to lay them out fairly. He also does not seriously engage the counter-argument that LLMs are a dead end for AGI and the panic is overblown. The reader looking for an honest map of the disagreement won’t find it here.

Who should read it

If you are a developer or a technically curious reader who has been following the AI story through the OpenAI side of the bookshelf and wants the other half of the picture, The Infinity Machine is the right book. It will tell you why Hassabis matters, what AlphaFold actually did, why the transformer caught DeepMind off guard, and what the inside of the Google-DeepMind merger looked like. The technical explanations are clearer than most tradebook attempts. The corporate intrigue is well-reported.

If you want a critical history of AI’s costs — to workers, to artists whose work was scraped, to the climate, to the open web — this is not that book and was never going to be. Read Karen Hao on OpenAI, or any of the recent investigative work on training-data lawsuits, instead.

If you came hoping for a definitive judgment on whether Hassabis is the responsible adult in the AI room he’s often portrayed as, you’ll leave undecided. He comes across as decent, driven, less glib than his peers, and convinced that he can build something on the order of a new mind without becoming the kind of person who breaks the world to do it. Mallaby believes him. We’re less sure. But we’re glad we read the book anyway, because the work behind AlphaFold is real, the chess kid’s story is genuinely improbable, and even a reverent biographer can’t fully obscure how strange the last decade has been.