Mallaby, Sebastian. 2026. The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for SuperIntelligence. Penguin Press.
Sebastian Mallaby reports on the life of Demis Hassabis, who won the Nobel prize as one of the researchers that created AlphaFold, a program that, given a sequence of amino acids, predicts how it will fold to form a protein.
At a very early age, Demis Hassabis was recognized as a chess prodigy. At one point, he had the second highest chess ranking in the world for a child. But as a teenager he abandoned chess and started to develop his life-long interest in machine intelligence.
After getting an undergraduate degree in computer science, he worked for a decade in the computer gaming industry. Then he left that to get a Ph.D. in neuroscience. Soon after he founded DeepMind in 2010, where he formulated the plan to develop Artificial General Intelligence.
From that point on, the accomplishments started rolling in. First, the DeepMind team built a program (Deep Q-Network, or DQN for short) to learn how to play Atari and several other computer games. A signature of these early efforts was the use of “reinforcement learning”. The program would play Atari and lose, but then learn from its mistakes. Next, DeepMind conquered the board game Go, with a program called AlphaGo. They built a version so powerful it defeated the world champion in four out of five games, an accomplishment that astounded the world. Lastly, the DeepMind team took on the challenge of protein folding, which earned Hassabis a 2024 Nobel prize in Chemistry (along with John Jumper).
The book details the struggles that Hassabis and his collaborators had in selling their vision of AGI to investors and finally how DeepMind became part of Google in 2014.
I was struck by how little relationship there is between the AI community and the cognitive science community. In the whole book, I could only see two points of contact. First, Hassabis got his Ph.D. in neuroscience (where he worked on the role of the hippocampus in memory), and was interested in importing ideas from neuroscience into AI. Second, the concept of reinforcement learning had been investigated in psychology. But other than these two points of contact, it seems to me that there was a complete lack of communication between DeepMind teams and what was known about the human mind from cognitive science.
It was also fascinating to me to see how blindsided Hassabis and other members of his team were by the advent of LLMs. They more or less ignored developments at OpenAI, until ChatGPT was launched (November 30, 2022), and then there was no way they could ignore it. At that point, Google put all of its resources (including the efforts of Hassabis) into developing Gemini. It took them a full year to catch up.
