A scientist with a vision, not another tech CEO

Sebastian Mallaby, author of the acclaimed book The Power Law on venture capital, spent over 30 hours of interviews with Hassabis and conducted more than 100 additional interviews with his colleagues, competitors, and critics. The result is a 480-page book that holds a 4.45 out of 5 rating on Goodreads from nearly two thousand reviews.

What emerges from the book above all: Hassabis is not another Sam Altman or Elon Musk. He didn't go into AI for money or fame. He is a scientist who used business as a tool to answer the deepest scientific questions. His entire life he has been asking a single thing: can intelligence be understood well enough to be recreated?

Mallaby shows how Hassabis's obsession with intelligence stretches back to childhood. At age five he was beating adults at chess, at twelve he became the second-best player in the world in his age category. But instead of a career as a chess grandmaster, he chose something bigger — to understand the very mechanism of thought.

Chess taught him to plan several moves ahead. Programming computer games (he worked on the hit Theme Park by Bullfrog Productions) showed him how complex worlds emerge from simple rules. A PhD in neuroscience at University College London then gave him the key to how memory, imagination, and planning work in the human brain.

Games as a laboratory for superintelligence

When Hassabis founded DeepMind in 2010 together with Shane Legg and Mustafa Suleyman, he had no idea that the first major triumph would involve a board game thousands of years old.

DeepMind started systematically — with games. They offer an ideal testing environment: clear rules, measurable results, and an astronomical number of possible decisions. First, their system beat human players in classic Atari games. Then came a challenge that had resisted computers for decades: Go.

In March 2016, AlphaGo defeated the legendary South Korean player Lee Sedol 4:1. Over 200 million people watched. The computer made moves that even elite players initially considered mistakes — and only later understood their brilliance. This was followed by AlphaGo Zero, which learned solely by playing against itself (without human game records), and AlphaZero, which used the same approach to master chess, shogi, and Go.

But games were never the goal. They were a testing ground. Mallaby describes in detail in the book how these systems taught the DeepMind team to navigate vast spaces of possibilities — a skill that would later pay off when solving real scientific problems.

The Cambridge conversation that changed biology

One of the most powerful passages in the book concerns a seemingly ordinary conversation from the 1990s. Hassabis was studying computer science at Cambridge and was friends with biology students. One of them told him about the protein folding problem — how a chain of amino acids can fold into a precise 3D shape that determines the protein's function. And that no one had reliably predicted this process for fifty years.

Hassabis noted protein folding on his list of problems that a sufficiently advanced AI might one day solve. And he waited. Twenty years.

In 2016, DeepMind officially launched the AlphaFold project. The first version won the CASP13 competition in 2018. For most organizations, that would have been the end — they won, time to celebrate. But not for Hassabis. He wanted a system that would be genuinely usable for scientists, not just a benchmark winner.

He brought John Jumper onto the team, fundamentally reworked the architecture, and in 2020 at the CASP14 conference presented AlphaFold2. Its accuracy approached experimental methods like X-ray crystallography, but at a fraction of the time and cost. DeepMind subsequently published, together with the European Bioinformatics Institute, the structures of more than 200 million proteins — virtually all known to science. The database has been used by over three million researchers from 190 countries.

In October 2024, Hassabis and Jumper received half of the Nobel Prize in Chemistry for this (the other half went to David Baker for computational protein design).

The blind spot: How DeepMind missed the language model revolution

The book also explains clearly for the first time one of the greatest mysteries of recent AI history: why Google — which in 2017 introduced the transformer architecture in the paper Attention Is All You Need, the foundation of all modern large language models — allowed OpenAI to define an entire era with ChatGPT.

The answer, surprisingly, makes sense. Hassabis simply didn't believe that language alone was enough to create true intelligence. He was convinced that an intelligent system must be grounded in the physical world — through perception, action, robotics, or simulated environments. A computer may have the definition of gravity, but does it understand it the same way as a person who has actually lifted something?

That's why DeepMind spent years investing in agents that learned in games and simulations. And underestimated how much information about the world is already encoded in language itself — and that large models can acquire a certain form of "grounding" through human feedback. Mallaby quotes Hassabis, who later admitted he considered language models "unreasonably effective" and that he underestimated this area.

This intellectual disposition explains why Google, with transformers up its sleeve, let OpenAI run away. Hassabis was searching for deeper intelligence; language seemed like an insufficient shortcut to him. As a product decision, however, it was fatal.

Why the sale to Google wasn't a capitulation

Another part of the book that upends conventional wisdom concerns the sale of DeepMind to Google in 2014. In startup mythology, independence is valued above all — a proper founder doesn't sell the company but builds an empire.

Hassabis saw it differently. When Larry Page (Google), Mark Zuckerberg (Facebook), and Elon Musk were all vying for DeepMind in 2013 (Musk's birthday party in a rented New York castle, where everyone dressed up as samurai, is one of the most bizarre passages in the book), Hassabis evaluated only one thing: what would get him closer to AGI faster.

Larry Page made an argument that worked: "If your real goal is AGI, why waste years building infrastructure when it already exists?" Google offered computing power, capital, talent, and patience on a scale that an independent DeepMind could hardly replicate. Hassabis therefore reportedly turned down a higher offer from Facebook and sold to Google.

The book doesn't hide that this move later created tensions around independence, commercialization, and ethical limits of AI. But it shows Hassabis as someone for whom mission always outweighed ego.

What early readers say about the book

The reactions are not unanimous. While many reviewers praise the unprecedented access to the behind-the-scenes of the AI race, others criticize Mallaby for being too lenient toward Hassabis. "The author doesn't even try to hide his admiration," wrote reviewer Kyle C. on Goodreads. Other readers would welcome more technical depth. Professor Jason Furman of Harvard, former economic advisor to Barack Obama, on the other hand called the book "phenomenal" and praised the portrait of Hassabis as a person driven by knowledge, not money.

One thing is certain: The Infinity Machine is so far the most comprehensive look at the personality who fundamentally influenced the direction of artificial intelligence — whether through AlphaFold, which genuinely helps scientists in drug development and disease research, or through the philosophy of mind that Hassabis brings to the AGI debate.

What to take away from the book — even without reading it

Mallaby's book (published March 31, 2026, currently only in English, no Czech translation announced) offers several universal lessons. The main one is that Hassabis doesn't start with technology and look for a market — he starts with a question and looks for the technology to answer it. That is a fundamental difference from most tech founders. And it is precisely this obsession with questions, not answers, that makes him perhaps the most interesting figure on the contemporary AI scene.

For Czech readers, the book is relevant in a broader context as well: it shows that the European technology environment (DeepMind was founded and is headquartered in London, outside traditional Silicon Valley) can compete with American giants — if it has a clear vision and enough patience.