No, the people building AI don’t know what they’re doing

There are early scenes of Sebastian Mallaby new book The Infinite Machine: Demis Hassabis, Deep Mind, and the Quest for Superintelligence If you’re comfortable with the idea that the people building AI (artificial intelligence) know what they’re doing, this may be worrying.

Hailed as the “godfather of deep learning” and considered one of the most respected figures in modern science, Nobel Prize winner Geoffrey Hinton is a member of the Royal Society in London. In a conversation with philosopher Nick Bostrom, Hinton confessed that he believes AI systems will be weaponized against civilians. When Bostrom asked why he was building it, Hinton’s answer was honestly disturbing. The prospects for discovery are “simply too optimistic,” he says.

This exchange earned Malaby his title and an animated question. Is the race to build artificial general intelligence a scientific heroic tale or a tragedy cloaked in the guise of progress?

of book This question does not and may never be resolved, but it is a formidable attempt to maintain an unflinching answer.

The main character of this book is Demis Hassabis, co-founder of DeepMind and winner of the 2024 Nobel Prize in Chemistry. His company was acquired by Google and is now one of the most scientifically accredited AI labs anywhere.

Hassabis is an extraordinary person, to say the least. A child prodigy and neuroscientist, his doctoral research on memory and imagination became the basis for machine learning algorithms.

Mr. Mallaby, who has already written some of the best books on hedge funds, the World Bank, and venture capital, and ranks as one of the great storytellers about the complexities of capitalism, spent three years and more than 30 hours working directly with Mr. Hassabis to complete this book.

The effort is evident from the beginning. endless machine It flows with the pace and rigor of a thriller. Mallaby spoke to more than 100 people inside and outside of DeepMind, from marginalized co-founder Mustafa Suleiman to AlphaGo architect David Silver.

What makes Mallaby’s work so special is the way he weaves three separate stories into one narrative tapestry.

First, there is Hassabis’ own story. Mallaby is acutely aware of the paradoxical nature of his subject matter. “For me, power itself is not of interest,” Mr. Hassabis once told Mr. Mallaby, but he managed to sneak Mr. Suleiman out of the company they founded under the cover of an investigation into misconduct that would have amounted to a simple discussion.

Also read: I was never good at exams. Was I a bad student?

The second story traces the evolution of DeepMind, which was founded in a small attic in London in 2010. From back then, when venture capitalists laughed at its ambitions to build something approaching human-level intelligence, to the company’s current position as the powerhouse behind Google’s ambitions in the AI ​​space. Mallaby does an excellent job of documenting the atmosphere within the company and its complex and often difficult integration into Google.

The third story is science, and this is where Mallaby does something really useful. In DeepMind, he explains with never-before-condescending clarity how reinforcement learning and deep learning have been combined to create a system that can learn by doing rather than being told, navigating the nearly “infinite” space of possible moves to master “Go,” producing a system that folds “proteins” without being given the rules of chemistry, and beginning to function as an “infinite machine,” in the words of the book’s title.

AlphaFold, a system that predicts the three-dimensional structure of nearly all proteins, is particularly well-illustrated, a problem that has puzzled structural biologists for 50 years. The speed with which Hassabis goes from celebrating AlphaGo’s victory to proclaiming on a midnight street in Seoul that protein folding is now within reach is one of several moments in which this book captures the unique restlessness of highly talented people.

What’s the difference? endless machine What stands out from the crowded shelves of AI books is Mallaby’s refusal to let safety issues slide into the background. He returns to it again and again, becoming more and more urgent as the book draws to a close.

The failures of every governance experiment, from DeepMind Health’s independent review board to OpenAI’s nonprofit board, from the Bletchley Park summit to Hassabis’ own dreams of a CERN-like international body overseeing the final steps to AGI, are documented with the exhausting precision of someone who has seen idealism meet institutional reality and come second best.

By 2025, the emergence of DeepSeek signaled that the AI ​​race had become truly multipolar and therefore essentially ungovernable, but Hassabis was uncharacteristically blunt at Davos. “The age of agents that we’re about to enter is an entry point where the risks in the system are much higher.”

Hassabis himself acknowledges that the safety agenda, once championed through formal governance structures, has become more personal. In other words, Mr. Hassabis will try to be present where decisions are made. It’s the most comfortable to wear. There aren’t that many.

Indian readers will find a reason to read this book.

The questions of who governs innovative technologies and whose values ​​are encoded into the systems that determine how governments process information, allocate economic resources, and direct scientific research are not abstract.

In a world where a handful of labs in London, San Francisco, and Shenzhen compete to build near-omniscient pattern recognition systems, the rest of the world is largely a bystander. Mallaby does not elaborate on this asymmetry, but it is present in the book’s silence.

The Global South does not appear as an actor in this story. At best, it is the future beneficiary of protein databases and medical diagnostics, and at worst, it is the result of a civilized gamble being taken without consent.

There are some parts of the book that are distorted. Mallaby admires Hassabis, and that admiration sometimes leads to advocacy work. Altman’s portraits are not as subtle as those of his subjects. Mr. Suleiman’s treatment has the feel of a case that has already been decided, as his own explanation of his withdrawal from DeepMind is greatly at odds with Mr. Malaby’s.

Also read: Will AI be the end of literature?

The book also avoids certain uncomfortable questions with extraordinary agility. The intellectual property rights of the writers, artists, and scientists who trained these systems are briefly discussed. There are almost no labor practices to maintain data centers or human reviewers to shape model behavior. These are not peripheral issues. They are part of the same moral structure that Mallaby wants to examine.

But these are all arguments against a book that has earned the right to be taken seriously. endless machine It succeeds in what good narrative nonfiction books do: make us feel the gravity of events that happen too quickly and are too technical to comprehend in real time.

Phrases from oppenheimerThe phrase “If I find something technically sweet, I’m willing to do it” appears in Mallaby’s work like a reproach. Hassabis knows the danger. As he sat at his desk at 2 a.m., Mallaby said, “I felt like reality was staring at me, screaming at me… trying to tell me something, if I only listened hard enough.”

He built the machine anyway.

The combination of clarity and compulsion, knowing and moving on despite, is the state that defines our moments. Mallaby wrote about this with intelligence, honesty, and considerable grace. Whether the machines his subjects built end up being kind to the species that created them is a question this book wisely refuses to answer. I made a great discovery. Discussions about what to do now are just beginning.

Deepanshu Mohan is Head and Professor of Economics at OP Jindal Global University. He is a Visiting Fellow at the School of International Development at the University of Oxford and a Visiting Professor at the London School of Economics (LSE).

Aman Chain is a senior research assistant at the Center for New Economic Studies (CNES) at OP Jindal Global University and is studying law at the university.

This article was published on May 29, 2:26 a.m., 4:24 p.m.

The Wire is now on WhatsApp. Follow our channel for incisive analysis and opinions on the latest developments.