Book Review: The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence by Sebastian Mallaby

Having previously read The Power Law, which I consider the best book written on venture capital, I approached Sebastian Mallaby’s latest book with unusually high expectations. The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence did not disappoint.

Like The Power Law, the book succeeds because Mallaby understands that transformative companies cannot be explained solely through technology or financial results. They must be understood through the ambitions, personalities, rivalries, and deeply held beliefs of the people building them.

The Infinity Machine is partly the story of an entrepreneur creating one of the most important artificial intelligence companies in history. More importantly, however, it is the story of a scientist who happens to use entrepreneurship as a vehicle for pursuing scientific questions.

Demis Hassabis does not come across as someone who founded DeepMind primarily to become wealthy, famous, or celebrated as a technology founder. He comes across as someone who identified intelligence as the most consequential problem he could work on and then constructed his life around solving it.

That distinction is what makes the book so compelling.

A Scientist First and an Entrepreneur Second

Mallaby traces Hassabis’s path from child chess prodigy and video game designer to neuroscientist, AI researcher, and eventually the co-founder of DeepMind. The book draws on more than 30 hours of conversations with Hassabis, alongside over 100 interviews with colleagues, competitors, critics, and former collaborators.

This access allows Mallaby to present Hassabis not simply as the public face of Google DeepMind, but as a person whose interests have remained remarkably consistent.

Chess taught him to think several moves ahead. Designing simulation games taught him how complex worlds could emerge from relatively simple rules. Neuroscience encouraged him to ask how memory, imagination, planning, and intelligence work inside the human brain.

These pursuits were not disconnected stages of his career. They were different approaches to the same underlying question: Can intelligence be understood well enough to be recreated?

What stands out throughout the book is Hassabis’s tendency to reduce problems to their foundations. Rather than beginning with what existing technology could accomplish, he repeatedly began with the outcome he believed should be possible and worked backward.

This first-principles approach also explains his unusually long time horizon. Hassabis was willing to hold a problem in his mind for years—or even decades—until science and computing had advanced enough for him to attack it properly.

The clearest example is protein structure prediction.

The Conversation That Planted the Protein-Folding Seed

While studying computer science at Cambridge in the 1990s, Hassabis became friends with biology students. One of them was particularly fascinated by the protein-folding problem and explained that cracking it could transform biology.

Hassabis immediately recognized it as the type of enormous search problem that might someday be addressed with artificial intelligence. He reportedly maintained notes on scientific problems that could eventually become suitable for the algorithms he hoped to build. Protein folding remained among them for nearly two decades.

This is one of the most enthralling parts of the book.

A seemingly ordinary conversation introduced Hassabis to a problem that would remain in the back of his mind through university, neuroscience research, entrepreneurship, the creation of DeepMind, and the development of increasingly capable learning systems.

To understand the significance of what followed, it is necessary to appreciate the scale of the challenge.

Proteins begin as chains of amino acids. Those chains fold into intricate three-dimensional structures, and the resulting shape largely determines what a protein can do inside a living organism. Understanding that structure is therefore essential for studying disease, designing medicines, and comprehending many of the mechanisms on which life depends.

The difficulty is that even a relatively small protein can theoretically assume an astronomical number of possible configurations. Experimentally determining a structure through methods such as X-ray crystallography or cryogenic electron microscopy can require substantial time, expertise, and expense.

For roughly 50 years, scientists struggled to reliably predict a protein’s three-dimensional structure from its one-dimensional amino-acid sequence. The challenge became one of the great unsolved problems in computational biology.

It is often said that AlphaFold “solved protein folding.” Technically, that description is too broad. AlphaFold transformed protein structure prediction; it does not explain every stage of the physical folding process, protein motion, misfolding, or molecular interaction.

That qualification does not diminish the achievement. Predicting structures with near-experimental accuracy was itself a breakthrough that changed an entire scientific field.

Games Were the Training Ground, Not the Destination

DeepMind did not begin with biology. Its early breakthroughs came through games.

Games offered something invaluable to an AI laboratory: controlled environments with clear rules, measurable outcomes, and enormous numbers of possible decisions. An agent could experiment, fail, receive feedback, and improve without the ambiguity and physical risk of operating in the real world.

DeepMind’s deep Q-network demonstrated that a single learning system could master a wide selection of Atari games from screen pixels and reward signals rather than game-specific instructions. The work helped establish deep reinforcement learning as one of the most promising approaches in modern AI.

The next major test was Go.

Go had resisted traditional computing approaches because its number of possible board positions made exhaustive search impossible. Success required pattern recognition, strategic planning, and the ability to identify promising moves without calculating every potential outcome.

In March 2016, AlphaGo defeated legendary Go player Lee Sedol four games to one. More than 200 million people watched a machine display moves that even elite players initially struggled to understand.

AlphaGo Zero then removed the dependence on human game records. Starting only with the rules, it learned by playing against itself and eventually surpassed the earlier version of AlphaGo. AlphaZero generalized the approach further, mastering chess, shogi, and Go through self-play without relying on handcrafted strategies or human examples.

These systems were not direct prototypes of AlphaFold. The connection is more philosophical and organizational than architectural.

Games taught DeepMind how to build systems capable of navigating vast possibility spaces. They demonstrated that neural networks, reinforcement learning, search, and enormous amounts of computation could uncover solutions that humans had not explicitly programmed.

More importantly, they gave Hassabis confidence that AI could move beyond classification and pattern matching to generate genuinely useful insights.

AlphaGo was a historic achievement, but it was never the final destination. Games were laboratories in which DeepMind could develop the ideas, people, infrastructure, and confidence needed to confront real scientific problems.

Protein structure prediction was where the original mission would finally be tested.

AlphaFold and the Difference Between Winning and Solving

DeepMind formally began working on protein structure prediction in 2016. Its first AlphaFold system entered the CASP13 protein-structure competition in 2018 and achieved the highest accuracy among the participants.

For many organizations, winning the competition would have been enough. It would have generated headlines, academic recognition, and proof that the project had succeeded.

Hassabis wanted more.

A system that topped a benchmark but remained insufficiently dependable for everyday scientific work had not truly solved the problem. DeepMind expanded the team, placed John Jumper in a central research role, and substantially redesigned the system rather than simply refining its first approach.

The result was AlphaFold2.

At CASP14 in 2020, AlphaFold2 achieved accuracy that organizers and researchers regarded as comparable to experimental methods across many protein targets. The associated research demonstrated that computational prediction could regularly approach atomic accuracy, including in cases where no closely related known structure was available.

DeepMind and the European Bioinformatics Institute later released predictions covering more than 200 million protein structures—nearly every protein catalogued by science. The database has since been accessed by more than three million researchers across over 190 countries.

In 2024, Hassabis and Jumper shared half of the Nobel Prize in Chemistry for protein structure prediction, with the other half awarded to David Baker for computational protein design.

The sequence of events captures what I found most impressive about Hassabis.

He does not appear satisfied by winning the accepted measurement of a problem. He continually asks whether the underlying problem has actually been solved.

That is first-principles thinking in its purest form. A benchmark is only a proxy. The objective is not to rank first. The objective is to create something that changes what scientists are capable of doing.

The LLM Blind Spot That Helped OpenAI Pull Ahead

The book also clarified one of the most confusing episodes in recent AI history.

Google researchers introduced the transformer architecture in the landmark 2017 paper Attention Is All You Need. Transformers became the foundation on which modern large language models were built.

Yet Google did not convert that head start into the product that defined the generative AI era. OpenAI released ChatGPT and established the interface through which hundreds of millions of people would first encounter advanced AI. Anthropic subsequently emerged as another leader, particularly among developers and enterprise users.

The common version of this story is that Google invented the transformer and then simply failed to build language models. That is not quite accurate. Google and DeepMind published substantial language-model research, including DeepMind’s 280-billion-parameter Gopher model in 2021.

The failure was not an absence of research. It was a failure of strategic conviction and product execution.

Hassabis initially doubted that language alone could produce genuine intelligence. He believed an intelligent system needed to be grounded in the world through perception, action, robotics, or simulated environments.

His concern was reasonable. A machine could store a definition of weight, but would it truly understand weight without ever lifting something? It could process descriptions of gravity, but would it understand that a glass would shatter when dropped?

DeepMind’s research agenda consequently emphasized agents that acted inside games and simulated worlds. Its researchers even built systems specifically intended to connect language with perception and action in three-dimensional environments.

What Hassabis underestimated was how much information about the physical and social world had already been encoded in human language. Large models could also inherit a form of grounding through feedback from people who had directly experienced the world.

Hassabis has since acknowledged that this was something he misjudged, describing language models as “unreasonably effective.”

This was one of the most valuable sections of the book because it makes Google’s behavior far more understandable.

From the outside, it seemed inexplicable that the organization responsible for the transformer would allow OpenAI to define the LLM era. From inside Hassabis’s intellectual framework, it makes more sense. He was searching for a deeper form of intelligence and initially regarded language prediction as an incomplete path toward it.

That judgment may eventually prove correct at the level of AGI. Language models alone may not be sufficient. But as a product and platform decision, it gave competitors an extraordinary opening.

In my own observation of the market, Gemini has become a formidable system and may outperform competing models on individual evaluations. Yet it still feels as though Google is trying to redefine a category whose expectations were established by OpenAI and Anthropic.

The book helps explain how one of the organizations with the greatest concentration of AI talent and infrastructure found itself in that position.

Why Selling DeepMind to Google Was Consistent with the Mission

The other episode that changed my understanding of Hassabis was DeepMind’s sale to Google.

Founder mythology tends to celebrate independence. The ideal entrepreneur is supposed to retain control, resist acquisition, build an empire, and receive personal recognition for creating a dominant company.

Hassabis appears to have evaluated the decision differently.

During the 2013 competition to acquire DeepMind, Larry Page made an argument that went directly to Hassabis’s priorities. If his actual objective was to create AGI, why spend years rebuilding infrastructure, raising capital, and constructing a company comparable to Google when Google’s resources already existed?

The conversations unfolded amid an almost surreal series of meetings involving Page, Mark Zuckerberg, Elon Musk, and other technology figures, including a gathering at a rented castle in New York for Musk’s birthday.

Google could offer computing infrastructure, capital, research talent, and patience on a scale that an independent DeepMind would have struggled to reproduce. It was also prepared to accept ethical conditions that DeepMind considered important. Hassabis ultimately chose Google despite reportedly receiving a larger offer from Facebook.

Seen through a conventional entrepreneurial lens, selling DeepMind so early could appear to be surrendering independence.

Seen through Hassabis’s lens, remaining independent may have been the greater distraction.

His objective was not to become known as the founder of the next Google. It was to use the fastest credible route toward building AGI and applying advanced intelligence to science. Google shortened that route.

The acquisition later produced real tensions over independence, governance, commercialization, and the ethical limits of AI. The book does not suggest that placing DeepMind inside one of the world’s largest corporations resolved those questions. In some ways, it made them more difficult.

Nevertheless, the decision reveals something important about Hassabis. He appears less attached to the conventional status of entrepreneurship than to the scientific mission entrepreneurship made possible.

One of the Good Guys—But Not an Infallible One

Hassabis ultimately comes across as one of the good guys in the AI race.

That does not mean every decision has been correct. His hesitation over language models was consequential. DeepMind’s relationship with Google has involved compromises. The concentration of powerful AI systems inside a small number of corporations creates questions that good intentions alone cannot answer.

But his motivations appear unusually consistent.

He wants to understand intelligence. He believes advanced AI can accelerate scientific discovery. He has repeatedly emphasized safety and the need to consider the long-term consequences of increasingly capable systems. Most significantly, AlphaFold provides concrete evidence that his vision of AI as an instrument for science is more than an aspirational talking point.

There is a difference between promising that AI will benefit humanity and releasing a tool used by millions of researchers to better understand the machinery of life.

AlphaFold gives Hassabis credibility that few other leaders in the AGI race possess.

Final Thoughts

The Infinity Machine succeeds as a biography, a history of DeepMind, and an accessible account of several of the most important breakthroughs in modern artificial intelligence. It also makes events we witnessed in real time—from AlphaGo and AlphaFold to Google’s delayed response to ChatGPT—feel considerably more coherent.

The most powerful lesson is not simply that Hassabis is exceptionally intelligent, although he clearly is. It is that he has been unusually deliberate about deciding which problems deserve that intelligence.

A conversation with biology students introduced him to protein folding decades before AI was capable of addressing it. He kept the problem in mind, built the organization and systems needed to confront it, and returned to it once the technology had advanced far enough.

Most entrepreneurs begin with an available technology and search for a market. Hassabis began with questions he believed could change humanity and worked toward creating technology powerful enough to answer them.

That is what makes The Infinity Machine such a memorable book and Demis Hassabis such an unusual figure in the history of artificial intelligence.