Long read: Review of Sebastian Mallaby’s book “The Infinity Machine: Demis Hassabis, DeepMind and the Quest for Superintelligence”

Inspite of the current hype around AI, most people we meet seem to misunderstand AI (often viewing LLMs and AI or machine learning and AI interchangeably). Sebastian Mallaby’s book (coming in the wake of his outstanding books on Alan Greenspan and VC investing) not only gives the reader a deep dive into the remarkable mind of Nobel laureate Demis Hassabis but also helps non-techies understand what exactly AI is and why it can transform the world on a scale comparable to the Industrial Revolution.

We found the book to be an interesting and insightful read for three reasons. Firstly, we get a detailed account of Hassabis’ remarkable upbringing and how we ended up creating DeepMind (which ended up creating much of the AI we currently associate with Google). Tim Clare writes for The Guardian: “The Infinity Machine is a detailed account of Hassabis’s journey from chess prodigy to the driving force behind a global leader in AI technology. Parts of this journey are rather remarkable. His Chinese Singaporean mother, we are told, “grew up in absolute poverty”, spending part of her childhood as an orphan on the streets of Singapore, while his Greek Cypriot father had dreams of music stardom and “sold toys out of the back of a beaten-up red Volkswagen van”.

Hassabis was unusually bright from an early age. He started playing chess – and beating adults – at four. By five, he was competing in tournaments, sitting on a phone book on top of two stacked chairs so he could see the table. By nine, he was captain of the England under-11 team. By 13, he had achieved the rank of chess master and was the second-strongest player in his age group globally.

It was an intensely pressured environment – wooden boards were placed under the tables “to prevent players from kicking each other”. If Hassabis lost, his dad frequently “went mental”, screaming at him, and when he counselled him to “do his best”, Hassabis took it literally; the only way he’d know he’d succeeded was if he exerted himself to near-collapse – “basically if I pushed myself to the point just before death”.

After a stint at Bullfrog games, working on the wildly successful Theme Park under designer Peter Molyneux, Hassabis studied at Cambridge, founded his own game studio, then returned to academia to complete a PhD in neuroscience. In 2010, along with his friend Mustafa Suleyman and Shane Legg, whom he met as a postdoc, he founded DeepMind”

Secondly, the book helps us understand the 15 years of slog & investing which has gone into creation of AI. Apart from Hassabis, we learn from Mallaby’s book about Larry Page, Elon Musk, Peter Thiel and others’ decades of investing time and big money in AI. By the time Chat GPT burst on to the scene in late 2022, Silicon Valley, British and American angels and VCs – often in competition with each other – had pumped tens of billions of dollars into AI. In the LA Review of Books, M Sriram writes:

“Despite an aversion to corporate hierarchy and Silicon Valley insularity, Hassabis sold DeepMind to Google in 2014 for $650 million because he saw this as his best shot at developing AI at his pace, using their resources…

Hassabis’s belief differed from that of billionaire investor and LinkedIn cofounder Reid Hoffman, who had first bet on an upstart called OpenAI in 2015 but later committed a billion-dollar fortune to DeepMind. Hoffman felt that multiple AI labs could separately gun for success, similar to a multiparty democracy in which, Mallaby writes, “this pluralism would be balanced by a shared commitment to bedrock values.” In the years since, Hoffman’s hunch has largely prevailed—but the “bedrock values” have frayed, not just in tech but also in the very democratic systems that inspired the analogy. That tension—between optimism and unease….

A series of big breakthroughs came between 2015 and 2017, when DeepMind mastered the 4,000-year-old board game Go, which is far more complicated than chess. DeepMind’s program AlphaGo beat successive champions, including one in Seoul in a match Google’s former CEO Eric Schmidt and cofounder Sergey Brin flew in to watch. The match drew 200 million viewers—more than the Super Bowl and more than double the audience that, two decades prior, had seen chess titan Garry Kasparov lose to a machine, a seismic moment in tech history.”

The third, in our view, the most important aspect of the book is that it helps us understand that AI is much more that LLMs. To be specific, and this is Hassabis’s biggest insight – AI will work like the human brain and hence AI will gradually become Artificial Generalised Intelligence i.e. tech agents will do much of the work we do today. Mallaby helps us understand that most LLMs currently mimic one of the two key dimensions on which the human brain excels i.e. deep learning (DL) which is akin to how neural networks function in the human brain. This is the “training” data set which most AI models run on. Hassabis realised 15 years ago that DL won’t be enough; for AI to become AGI, it would need to be fed RL (i.e. reinforcement learning as well). This is the aspect of our brains which makes mistakes, learns from the mistake and then changes the way it makes decisions in the future. How certain AI models learn to toggle between DL and RL makes for fascinating reading in some of the most enthralling chapters of the book as Demis Hassabis and his colleagues move towards building tech that this planet has never seen before.