Demis Hassabis on Gemini 3, World Models, and the Path to AGI
In this season finale of Google DeepMind: The Podcast, host Hannah Fry sits down with Demis Hassabis to reflect on a landmark year for AI. Hassabis frames the past year as "10 years in one," highlighting the release of Gemini 3 with its multimodal capabilities and major advances in world models such as Genie and Veo. He discusses the "root node" problems DeepMind is tackling—from materials science and room-temperature superconductors to a newly deepened partnership with Commonwealth Fusion on plasma containment and a collaboration with Google's quantum AI team on error correction codes, all aimed at unlocking abundant clean energy. Hassabis explains why AI still falls short of AGI, pointing to the "jagged intelligence" paradox where models can win IMO gold medals yet fail at high-school-level logic problems. He attributes the gap to missing pieces: consistency, reliable use of inference-time reasoning, and the inability to learn continuously in the world. On scaling, he pushes back against claims of a wall, arguing DeepMind sits in a middle regime of diminishing-but-meaningful returns and that synthetic data plus research breakthroughs keep progress moving, aided by Google's TPU stack. Looking ahead, Hassabis sees agentic systems as the next major inflection point—more capable but also riskier, prompting work on cyber defense for a world of millions of autonomous agents. He stresses that commercialization has pulled resources and attention but worries it has created a "crazy race condition" that complicates rigorous safety research. On societal impact, he draws lessons from the industrial revolution—projected to be "10x bigger and 10x faster"—and suggests universal basic income or more radical democratic-economic models may be needed. He reiterates his 5–10 year AGI timeline, warns that existing institutions are too fragmented to handle what's coming, and philosophically bets that the universe is computationally tractable, with consciousness, creativity, and qualia potentially reducible to information processing—though he leaves room for quantum or non-computable surprises.
Original sources
Podcast Google DeepMind: The Podcast
Episode:The Future of Intelligence with Demis Hassabis (Co-founder and CEO of DeepMind) ↗
YouTube Google DeepMind: The Podcast
Video:The future of intelligence | Demis Hassabis (Co-founder and CEO of DeepMind) ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.