In this extensive dialogue, Demis Hassabis outlines his foundational worldview that information is the fundamental substrate of reality, bridging theoretical computer science, physics, and biology. Addressing his Nobel Prize lecture conjecture, Hassabis explains that systems shaped by physical and evolutionary selection ('survival of the stablest') inhabit compact low-dimensional manifolds. This structured nature allows classical deep learning algorithms to bypass brute-force combinatorial explosions and model complex phenomena such as protein folding, genomics, and fluid dynamics in polynomial time.
The conversation examines generative world models and intuitive physics through models like Veo. Hassabis argues that high-fidelity video generation demonstrates an internal representation of physics and material interactions, which naturally connects to the future of interactive simulations and video games. Discussing algorithmic evolution, Hassabis explains how combining large language models with evolutionary search algorithms (such as AlphaEvolve) and tree search mechanisms enables AI to discover entirely novel algorithms, such as faster matrix multiplication methods, moving beyond human-generated training distributions.
Hassabis outlines DeepMind's flagship biological ambition: creating a comprehensive 'Virtual Cell' within the next decade by composing hierarchical models of molecular interactions across vast temporal and spatial scales. Moving to the trajectory toward AGI, Hassabis estimates a 50% likelihood of achieving AGI by roughly 2030. He argues that true AGI requires uniform cognitive capabilities across diverse domains, specifically the capacity for high-level creative synthesis, hypothesis formulation, and formulating breakthrough conjectures akin to Einstein or von Neumann.
On engineering and deployment, Hassabis discusses managing Google DeepMind's rapid release cadence across the Gemini model family, the strategic management of compute constraints across pre-training and test-time reasoning, and long-term energy solutions like nuclear fusion. Addressing existential risks, Hassabis advocates for cautious optimism, arguing that while catastrophic misuse by malicious actors and loss-of-control scenarios are non-negligible, international scientific collaboration and rigorous empirical evaluation are essential to unlocking AI's transformative potential for global abundance.