In an Intelligence Squared conversation moderated by Tom Mackenzie, Demis Hassabis explains why Google and DeepMind did not immediately release their early language models. He describes concerns about hallucinations and reputational risk, and says ChatGPT demonstrated that users could find imperfect systems useful. Sebastian Mallaby discusses the later merger of Google Brain and DeepMind, presenting his account of how an integrated research effort helped Google respond.
Demis Hassabis describes a recurring research pattern: train a model to understand a domain, then use that model to guide a search toward a clear objective. He connects game-playing work to AlphaFold and argues that learned guidance makes otherwise impractical searches manageable. His discussion distinguishes this approach from enumerating every possibility by brute force.
Demis Hassabis outlines hoped-for applications in medicine, energy, weather and materials, alongside longer-term scientific questions. These are presented as his ambitions rather than guaranteed outcomes. He also discusses dual-use risks and the tension between commercial competition and international coordination on advanced AI safety. Book context is retained where relevant; no paid sponsor segment appears in the reviewed excerpt.
Watch on YouTube
