Sebastian Mallaby on 'The Infinity Machine': Demis Hassabis secretly built a hedge fund inside DeepMind to beat Jim Simons

Summary

Sebastian Mallaby, author of The Infinity Machine, offers one of the more grounded portraits of Demis Hassabis to emerge from the current AI boom — a researcher who co-founded DeepMind with safety as a founding condition, quietly fantasizes about retiring to the Institute for Advanced Study, and yet is too competitive to walk away from the fight.

Mallaby's access was unusual. He met Hassabis at European tech conferences before writing the book, and their first substantive exchange centered on Renaissance Technologies — specifically that Peter Brown had done a PhD with Geoffrey Hinton on AI and applied it to markets. That hook got Hassabis's attention. By the time Mallaby formally pitched the project, ChatGPT had gone viral that same week, and Hassabis's response was unambiguous: "This is war. They've parked the tanks in my front yard."

The Google DeepMind merger

The merger between Google Brain in Mountain View and DeepMind in London — eight time zones apart, executed mid-race against OpenAI — was widely expected to fail. Mallaby's read is that it succeeded largely because Hassabis imported a management structure DeepMind had used for years: the strike team. Borrowed from video game development, which Hassabis practiced before founding DeepMind, the model works by suspending bottom-up blue-sky research at a chosen moment and redirecting everyone onto a single shared codebase toward a defined outcome. DeepMind used it for AlphaGo and AlphaFold. Google Brain had no equivalent. Within two and a half years of the merger, Gemini was outranking OpenAI's models on the benchmarks.

The person who made that possible, in Mallaby's framing, is Sundar Pichai. Hassabis describes the Pichai relationship as the most important business relationship in AI today — Pichai absorbs the capital commitments, tens to potentially hundreds of billions in compute spend, and gives Hassabis the oxygen to stay focused on the science. Hassabis has told Mallaby several times that building Gemini is no longer a distraction from AGI research; at this stage of maturity, shipping the next LLM, building toward reasoning and agentic models, is the scientific progress.

Google's reputational drag

Being inside Google also slowed things down. Google had working chatbot prototypes in 2022 but held them back, worried about hallucinations damaging the company's core identity as a reliable information source. OpenAI launched first, forced Google's hand, and Hassabis's candid admission to Mallaby was that the public turned out to be far more tolerant of hallucinating tools than Google had assumed. That caution cost them the first-mover moment.

Safety and the race dynamic

Hassabis met his co-founder Shane Legg at an AI safety lecture, and safety was a formal condition of the 2014 Google acquisition — no weapons applications, and an independent oversight committee for AGI deployment. Mallaby acknowledges Hassabis has slipped on the military commitment. More broadly, Hassabis told him it is a paradoxical moment: he is succeeding as an AI inventor and failing as an AI steward. The race dynamic, which now includes China, makes it structurally difficult to slow down unilaterally.

Mallaby spent two hours in Geoffrey Hinton's Toronto kitchen debating machine motivation. His own prior assumption was that machines, unlike humans, have no evolutionary survival drive and therefore no reason to harm people. Hinton's counter: the moment you instruct an AI to defend itself against cyberattack, you have programmed it with a survival instinct, and that instinct can compound in ways that don't require biological evolution. Mallaby finds that argument hard to dismiss.

OpenAI's financing problem

On the bubble question, Mallaby draws a sharp line. The technology is not a bubble — three and a half years of progress from hallucinating chatbot to multimodal reasoning to agentic models is real acceleration, with physical AI and robotics likely the next major wave. But OpenAI's capital structure is a different matter. The company raised $41 billion in 2024, a record for any private fundraising, larger than any IPO that year. Yet most users still don't pay, and the path to break-even is currently targeted at 2030. The $100 billion figure recently cited is, in Mallaby's view, largely contingent — tied to a future IPO, future compute commitments in-kind, and forward tranches. The actual cash raised is materially smaller. DeepMind's embedded position inside Alphabet is, by contrast, a structural advantage that OpenAI cannot replicate.

The Hassabis paradox

Asked whether Hassabis could be CEO of Google, Mallaby lands on genuine uncertainty. Hassabis has told him he fantasizes about retreating to pure research at a Princeton-style institute. He also clearly enjoys the competitive fight. Both things are true simultaneously, and Mallaby suspects Hassabis himself doesn't know which would win if the choice were forced.