A dispatch from the last sane era of AI

In November 2022, Sebastian Mallaby negotiated deep access to Demis Hassabis. A week later, ChatGPT launched. Mallaby began his interviews just as the world he was proposing to describe was being demolished around him. In a way, this painstakingly crafted book that came three years and hundreds of interviews later is a perfect chronicle of the pre-2022 moment in AI, when the field still had protagonists. Now, when we think of AI, we think of faceless monoliths with giant data centres grinding up capital and compute into an equilibrium presented to us as a chat box on a screen.

Janet Malcolm’s controversial book The Journalist and the Murderer opens with what is widely considered one of the most incendiary first sentences in modern literature: “Every journalist who is not too stupid or too full of himself to notice what is going on knows that what he does is morally indefensible.” Malcolm argues that the fundamental relationship between a reporter and a biographical subject is a con game. To get a compelling story, journalists cultivate intimacy and feign sympathy, only to inevitably betray the subject to serve the narrative.

The Infinity Machine is not such a book, primarily because Sebastian Mallaby is not such a journalist.

Each book Mallaby writes takes a class of financial actor the public suspects and pens a sympathetic profile to launder them, making them appear more socially useful than you might think. More Money Than God (2010) argued hedge funds stabilise markets; The Power Law (2022) argued venture capital’s hit-driven returns justify the model. His books tend to be rich inner accounts of people in positions of financial or technological power, completely insulated from the consequences of their actions. The Power Law had almost nothing on what VC-subsidised blitzscaling did to incumbent workers, and his book on Alan Greenspan barely addressed the 2008 wreckage whose preconditions started on his watch.

Yet, his tendency to write rich inner accounts devoid of external wreckage makes him the perfect, if unwitting, biographer for a man who set out to cure disease but ended up commanding what might be the defining corporate war of the 2020s.

Early AI

Demis Hassabis grew up poor in North London, made chess master at 13, shipped a bestselling video game at 17, and took a neuroscience PhD as a stealth route to artificial intelligence. At its core, early AI was “narrow.” You wrote a program to do one specific thing with a deterministic output (the same set of inputs would always provide the same set of outputs).

Hassabis’s founding thesis for DeepMind was entirely different: true intelligence is a flexible, general-purpose learning algorithm. Reverse-engineer how the human brain learns from scratch, and you can build Artificial General Intelligence. Once you have a machine that can learn universally, you point it at humanity’s biggest bottlenecks.

And to a staggering extent, he did. DeepMind tackled the 50-year-old grand challenge of biology: protein folding. Rather than trying to simulate the raw physical and chemical forces of how a 1D string of amino acids crumples into a 3D shape, Hassabis and lead researcher John Jumper treated it as a data problem. By leveraging evolutionary patterns and the same “attention networks” that underlie modern chatbots, AlphaFold successfully predicted the 3D structures of almost all 200 million known proteins on Earth. They released it as a free “Google Maps for biology,” instantly accelerating global research into new medicines and earning Hassabis and Jumper the 2024 Nobel Prize in Chemistry.

But AlphaFold required massive capital to scale, a reality Mallaby captures when documenting Hassabis steering DeepMind into Google’s arms. In today’s AI stack, a raw model is nothing without its harness. A model by itself is a collection of weights. The harness is everything wrapped around it — the tools, capital, and post-training that turn capability into something useful outside a chat box.

Mallaby writes Hassabis as the model: singular, irreplaceable, the source. But his own evidence reveals the harness: a builder of teams, a shepherd of compute, and above all, a narrator whose persuasion colleagues referred to as a “Jedi” mind trick.

Narrative control

And here the harness loops back to Malcolm: her con presumes the journalist controls the narrative. In this book, the subject does. Hassabis is, among his many and varied talents, one of the best narrative spinners in tech. And in The Infinity Machine, he met a biographer whose defining tendency is to believe rich inner accounts.

The book’s silences follow Mallaby’s pattern. There is almost nothing on the compute-industrial base whose costs bent DeepMind toward Google in the first place, and one sentence on Google discarding its acquisition-era promise to keep the technology away from weapons and surveillance.

The book’s tragedy, which Mallaby documents in the whiplash of his late chapters is that its subject spent a decade building AI as a scientific instrument, only to spend the years since ChatGPT conscripted into a gold rush, commanding Google’s efforts to catch up to OpenAI and then Anthropic, with its first doomed Bard family of models, followed by the now ubiquitous Gemini.

Most striking is the fate of the book’s own closing doctrine. Hassabis, convinced of the danger unregulated AI would pose to the world, spent a large portion of the narrative attempting safeguard after safeguard that kept failing.. Like the ethics board that never met, or the spin-out under an independent charter that Google smothered. Mallaby lets his subject conclude that AI safety rests not on governance structures but on trust between the people building the technology, and half-blesses this as maturity. This doctrine did not survive to the paperback.

If Malcolm argued that biography is a con where the writer betrays the subject, The Infinity Machine offers a tragic inversion: Mallaby didn’t betray Hassabis; the field itself, moving faster than any book could, betrayed them both.

Within months of publication, the US Department of War was blacklisting Google’s rival Anthropic for refusing surveillance work, and its sister Department of Commerce was switching off Anthropic’s frontier model Fable worldwide by an export-control letter.

Meanwhile, Chinese authorities have held meetings with their own top tech firms during the same period, about potentially restricting overseas access to China’s most advanced AI models. Hassabis’ idealistic vision of a self regulating class of enlightened technocrats responsibly steering humanity towards superintelligence effectively rolled over when it met the actual institutions that make the decisions about how the world works.

Commodification

And what’s even more ironic is that the technology itself, despite the aforementioned governments’ frantic efforts to close it off, might be rapidly turning out to be a commodity, akin to electricity or telecommunications, rather than an enduring strategic advantage.

Multiple Chinese labs under US sanctions, training on Huawei chips with no Nvidia hardware at all, now score within a few points of the best proprietary models on all the major benchmarks. And these are open models which means anyone on the planet, in theory, with sufficient resources can deploy and run them.

The speed of advancement of the field makes the book feel dated immediately in print. I’d urge curious readers to read it anyway. It is the definitive portrait of the most interesting mind in the field, and the last major document of the era when AI had protagonists rather than infrastructure.

AlphaFold, however, stands as Hassabis’ singular achievement and required no arms race. It required taste, conviction, and perhaps twenty stubborn people. It remains the strongest evidence that Hassabis’s vision of AI as a scientific instrument was never the same project as the commercial frenzy that eventually came to eclipse it.

Check it out.

(The reviewer is a product designer and researcher working in the tech sector. He lives and works in Bengaluru.)

About the book Book title: The Infinity Machine: Demis Hassabis, DeepMind and the Quest for Superintelligence Author: Sebastian Mallaby Publisher: Penguin Press No. of pages: 480 pages Cost: ₹1,299

Published on August 2, 2026