Demis Hassabis on AI's Capacity to Model Nature, Simulate Reality, and Revolutionize Video Games

Demis Hassabis, leader of Google DeepMind and a Nobel Prize winner, elaborates on his provocative conjecture that "any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm." He posits that natural systems, from protein folding to geological formations and planetary orbits, possess inherent structure due to evolutionary processes, which he terms "survival of the stablest." This underlying structure, unlike truly random or abstract problems, makes them amenable to efficient computational modeling and prediction by neural networks running on classical computers, as exemplified by AlphaGo and AlphaFold's successes in combinatorially vast spaces.

The discussion delves into the P=NP question, with Hassabis suggesting a new complexity class, "Learnable Natural Systems" (LNS), for problems efficiently solvable by classical AI systems. He views the universe as fundamentally an informational system, implying that understanding its information processing could unlock solutions to deep physics questions. Hassabis argues that classical systems have far greater capabilities than previously imagined, challenging the notion that quantum computing is necessary for complex tasks like protein folding. He also touches on the boundaries of this paradigm, considering emergent phenomena like cellular automata as potentially modelable, while chaotic systems might remain difficult.

A significant portion of the conversation focuses on DeepMind's video generation model, V3, and its astonishing ability to accurately model complex physics, including fluid dynamics, materials, and lighting, purely from passive observation of YouTube videos. This capability challenges the long-held belief in neuroscience that embodied interaction is essential for developing an intuitive understanding of physics. V3's performance suggests it extracts a lower-dimensional manifold or underlying structure of reality, hinting at a fundamental organization of the universe that can be learned without direct physical interaction, moving towards the development of a comprehensive "world model" crucial for AGI.

Hassabis, drawing from his background in game development, envisions a future where AI transforms video games. He foresees AI systems capable of creating truly open-world, deeply personalized, and dynamically storytelling experiences that adapt to player choices, moving beyond the "illusion of choice" prevalent in current games. He dreams of AI-powered "vibe coding" to create games and considers game development, alongside advancing his physics theory, as potential post-AGI projects. He emphasizes the profound connection between simulating reality in games and understanding the fundamental nature of the universe, prompting a philosophical inquiry into the distinction between increasingly realistic simulations and physical reality, and the sources of meaning in a world where AI handles much of the "work."

"any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm."

"in nature, natural systems have structure because they were subject to evolutionary processes that that shape them."

"Anything that can be evolved can be efficiently modeled."

"I think information is primary. Information is the most sort of fundamental unit of the universe more fundamental than energy and matter."

"I think we haven't really uh even sort of scratched the surface yet of what uh classical systems socalled uh uh could do."

"I think it's telling us something quite fundamental about how the universe is structured in my opinion."

"I don't think you can generate that kind of video without understanding and then our own philosophical notion what it means to understand then is like brought to the surface."

"it seems like um you can understand it through passive observation which is pretty surprising to me and and again I think hints at something underlying about the nature of uh reality."

"I think the only way you can do that is to have uh generated systems, systems that uh will generate that on the fly."

"there's no other media uh entertainment media where you do that where you as the audience actually co-create the the story."