Google DeepMind C.E.O. Demis Hassabis on Living in an A.I. Future | EP 137
Summary
The hosts read Google’s I/O pitch as shifting from “let Google do the Googling for you” to an AI workbench that users actively operate—and pay $250 a month to access at the frontier. AI Mode’s cleaner interface can fan one query into dozens of searches—72 websites for a Costco-membership question—but costs more to serve and has no ads yet. The investor hinge is whether shopping, subscriptions, and other transactions can replace economics built around blue links.
Google presented evidence of credible distribution and competitive momentum: Gemini reached 400 million monthly users while its token output rose 50-fold in a year. Unlike the 1.5 billion people passively shown AI Overviews, Gemini users must deliberately open the app or site. Kevin Roose described Google as a team that “knows it’s going to be in the playoffs at least,” though Wall Street’s muted reaction reflected difficulty connecting a two-hour product fire hose to incremental value.
Demis Hassabis puts most of his AGI probability mass five to 10 years out, with Google DeepMind now the “engine room of Google”; he said he thinks the AGI effort is already “past the middle game.” His definition is deliberately demanding: AGI must theoretically perform “all of the things that the human brain can do,” originate important conjectures rather than merely solve them, and become so consistent that even top experts struggle to expose trivial flaws.
The research strategy is explicitly additive: scale the general models, distill them into efficient workhorses, specialize them for domains such as protein structure, and pursue breakthroughs beyond the standard stack. “I’m an and,” Hassabis said. Roughly 90% of Google’s productivity and science work rests on core models—especially Gemini 2.5—with domain data and experts supplying the remaining specialization and feeding discoveries back into the general model.
AlphaEvolve is an early but commercially relevant specimen of automated AI research, using Gemini models and evolutionary selection to improve code, chip design, data-center scheduling, and matrix multiplication. It remains narrow, human-supervised, and limited largely to problems with provably correct evaluation; Hassabis said it is shaving useful percentage points, not producing recursive-intelligence “step changes.” Its creative lesson is that hallucination can become “imagination” when wild proposals are filtered by rigorous evaluation.
Hassabis believes the safety stakes rise sharply as capable agents emerge, possibly in two or three years, requiring a “step change” in controllability research and international benchmarks. He has changed his mind toward limited real-world deployment because tens of millions of users reveal edge cases no thousand-person test team can find, but still argues for rigorous internal testing first. Export controls remain an unresolved trade-off between limiting frontier-model proliferation and ensuring Western technology is adopted globally.
For the next five years, Hassabis expects AI mainly to augment workers, but he would not forecast confidently beyond that transition. He still recommends STEM and coding, combined with becoming a “ninja” at current tools and cultivating creativity, adaptability, resilience, and “learning to learn.” Longer term, he imagines smaller AI-leveraged teams, personal fleets of agents, “radical abundance,” and likely “universal high income”—while insisting that human connection and artistic struggle may retain value even when machines outperform technically.
Deep dive
Not yet available upstream; scheduled sync will retry.