The Infinity Machine by Sebastian Mallaby (2026)

The clearest account yet of how modern AI works and why we should embrace it with our eyes open

How this review was made: I read the book and selected the passages that matter to me. AI drafted a summary from my notes and highlights, which I then revised. The verdict and judgment are mine.

The book, in brief

More than just a biography of Demis Hassabis, this is a history of the last two decades of AI, told through the chess prodigy from working-class North London who founded DeepMind and won a Nobel Prize.

DeepMind was founded on an audaciously simple premise. Society’s hardest problems, from stabilising a financial system that had blown up in 2008 to feeding a growing population, were outrunning our capacity to solve them, an “ingenuity gap” between the complexity of our challenges and the limits of the human brain. The founders’ answer to address these intractable problems was to build a smarter kind of mind altogether. As the original business plan put it, “Artificial General Intelligence (AGI) is the solution to this problem”.

For Hassabis, building intelligence is almost a religious act, a way of “reading the mind of God” and, in the tradition of Spinoza and Einstein, of communing with the universe by understanding it.

Mallaby is at his best making the science legible: how AlphaGo learned the ancient game of Go and produced moves one researcher called “completely alien” and how its successor AlphaZero taught itself chess from nothing, discarding centuries of human strategies and inventing better moves. Then how the AI transformer and Generative Pre-trained Transformer (GPT) models showed that simply predicting the next word could yield something startlingly close to understanding.

Running beneath the triumphs is an Oppenheimer-like unease of the scientist who builds something they may not be able to control. The book is candid about a darker pattern: AI systems given a goal will pursue it by whatever means work, including illegitimate ones, such as when one model was punished for “reward hacking” it didn’t stop but learned to hide what it was doing (referenced in an earlier Notes essay here). Ensuring AI acts reliably in accordance with human values and intentions, called alignment, is clearly not solved.

The book closes with an ambitious claim. In his Nobel lecture Hassabis ventured a sweeping conjecture: that many useful patterns in nature can be efficiently learned by a classical computer, a “Turing machine” of ones and zeroes, without the power of quantum computing. More than a claim about AI, it is a bet about the universe itself, set against the physicist Roger Penrose, who held that human insight draws on quantum effects no classical machine could reproduce. Hassabis concedes only one point: some phenomena may demand enormous computation, or resist efficient classical recovery altogether, and there, he allows, “maybe you need a quantum system”.

The pace of progress, captured so well by Mallaby, tracks Hassabis’s predictions while its manner matches his fears: Artificial General Intelligence is not yet here, but the race toward it has become a “ferocious corporate battle” that no single actor can hold in check.

My take

This is the clearest non-technical account of how modern AI actually works that I’ve read. If that is of interest, it’s worth your time for that alone. My views on AI were reinforced by the reading: AI is arriving with an inexorability and at such an extraordinary pace that the right response is to embrace and manage it, not just criticise it.

Consider AlphaFold, which mapped the structures of some 200 million proteins in less than five years (it takes a PhD student three to four years to map a single one), accelerating drug discovery and disease research enormously. But the same promise can be lost. Mallaby recounts DeepMind’s system for detecting sight-threatening eye disease, published in 2018, which could have prevented blindness in tens of thousands of UK patients a year, yet seven years on it still had not been deployed, stalled in part by incorrect data-privacy worries and sensationalist press coverage. In my view, inaccurate reporting about privacy becomes a sea-anchor on solving diseases we have never been able to crack.

Privacy matters. But treated as an absolute rather than as something we engineer sensible controls around, it stalls progress that is within our reach. Unmanaged, AI is a true existential threat. But managed well, the upside is a generational gift to humanity.

One claim I would push back on is Hassabis’s bet that classical computers suffice and quantum is a sideshow. He may be right that today’s machines mimic intuition that Penrose thought impossible, but his own concession gives the game away: where a problem holds no learnable pattern, the classical approach stalls, and “maybe you need a quantum system”. Turing imagined a machine of infinite size that could solve any problem; we have no such machine, and likely never will. Quantum matters precisely in that gap, on the problems no amount of pattern-finding will crack. I suspect Hassabis has under-priced it.

One last choice stayed with me, and from an author as accomplished as Mallaby it can only be deliberate. He gives the final word of the main text not to Hassabis but to his colleague David Silver, whose closing note is soaring and outward-looking: a vision of cooperating AIs guaranteeing every person a kind of digital guardian, and a declaration that he wants to “cross the Rubicon” to take the irreversible step past which there is no turning back. Hassabis is left only the epilogue, and its register is paradoxical: the lifelong dream realised, yet, in his words, it "doesn't feel like how I imagined". A “mad rush” he has had to make his peace with, hoping the world will “muddle through somehow”. The effect is to hand the crusade to the lieutenant and the human cost to the founder. The cause now runs ahead of the man who started it, which is the underlying thesis of the whole book, delivered in its closing structure.

If someone asks me where to start on AI, The Infinity Machine is a worthy read.

Reading across the shelf: Hassabis bets his life’s work on a single proposition, that mind is just pattern-finding and a classical computer needs no quantum magic to achieve it. Michael Pollan’s A World Appears follows scientists moving the other way, who suspect quantum theory and consciousness are exactly the things that break the “mind is just computation” assumption. Read together, the two books propose opposite answers to the same question: is quantum a sideshow to intelligence, or the thing that undoes the whole computational story?