Nobel Prize Winner's AI Theory: Why Any Natural Pattern Can Be Learned Classically

Watch the Full Interview

Nobel Prize winner Demis Hassabis drops a bombshell theory that could reshape how we think about AI, physics, and computation. In this wide-ranging conversation with Lex Fridman, the DeepMind CEO goes way beyond discussing past achievements like AlphaFold. He reveals a profound conjecture about the nature of reality itself.

Key Insights

• Hassabis proposes a revolutionary conjecture: any pattern generated or found in nature can be efficiently discovered and modeled by classical learning algorithms like neural networks
• Natural systems contain underlying structure shaped by evolutionary processes that makes them computationally learnable, contrary to assumptions of randomness or chaos
• DeepMind’s video generation model Veo has learned accurate physics simulations purely from watching videos, without embodied interaction or explicit physics equations
• AI could revolutionize gaming by enabling truly open-world experiences with dynamically generated narratives that adapt uniquely to each player’s choices
• AlphaEvolve combines large language models with evolutionary algorithms, creating hybrid discovery engines that find solutions neither approach could identify alone
• DeepMind aims to build a complete AI model of a biological cell, representing the next frontier after AlphaFold’s protein folding breakthrough
• The success of neural networks at modeling natural phenomena hints at deep connections between computation, physics, and the fundamental structure of reality
• Games serve as laboratories for intelligence development, providing compressed versions of reality where AI systems can safely explore and learn
• The convergence of AI capabilities in language, video generation, and game mechanics could produce virtual worlds indistinguishable from human-crafted experiences

The Nobel Prize Conjecture

Hassabis presents a groundbreaking conjecture that any pattern found in nature can be efficiently discovered by classical learning algorithms. This theory bridges physics, computation, and AI by proposing that natural systems contain built-in computational shortcuts that evolution discovered and AI can rediscover.

The conjecture is grounded in concrete breakthroughs like AlphaFold solving protein folding, a problem previously thought computationally intractable. Hassabis argues that natural systems aren’t random chaos but have underlying structure shaped by billions of years of evolutionary optimization.

Evolution acts as a massive optimization process, carving efficient paths through problem spaces that neural networks are uniquely suited to discover. This creates what Hassabis suggests might be a new complexity class for “learnable natural systems” - problems that classical computers and neural networks can handle efficiently because they reflect patterns found in the physical world.

The theory connects to fundamental questions in computer science like P vs NP. Hassabis proposes that the universe provides computational shortcuts that make seemingly intractable problems tractable when approached through the lens of evolutionary structure.

Learning Physics from Videos

DeepMind’s video generation model Veo demonstrates a remarkable capability: learning accurate physics simulations purely from watching videos, without equations or embodied interaction with the physical world. The model accurately simulates liquids, lighting, and other physical phenomena through pattern recognition alone.

This challenges fundamental assumptions about how intelligence develops understanding of the physical world. Humans typically spend years interacting with objects, dropping things, splashing in water, and watching shadows move to build intuitive physics. Veo achieves similar understanding through pixel pattern analysis.

The implications suggest that embodiment may not be necessary for understanding physics. Massive-scale pattern recognition on video data contains sufficient implicit physical knowledge to extract underlying rules governing physical behavior.

YouTube and other video repositories effectively serve as physics textbooks, containing rich information about how objects behave under various conditions. AI systems can extract this knowledge without direct physical experience.

The Future of AI-Generated Worlds

Hassabis envisions a revolution in interactive entertainment where AI generates content dynamically, creating truly open-world experiences beyond current games’ “illusion of choice.” Current games offer pre-scripted branches that eventually converge, but AI could create genuinely unique experiences for each player.

The convergence of large language models for narrative generation, video generation for visuals, and reinforcement learning for game mechanics could produce virtual worlds that rival human-crafted experiences. AI agents wouldn’t just play games but actively create compelling game experiences.

This represents a shift from developers creating every quest, character, and storyline to AI generating them dynamically based on player actions. The result would be emergent narratives that adapt deeply to choices, creating stories that even developers never imagined.

DeepMind is actively experimenting with these concepts, exploring how AI systems can become creative partners in game development rather than just optimization tools.

AlphaEvolve and Hybrid Intelligence

AlphaEvolve represents a new paradigm combining large language models with evolutionary algorithms to create hybrid discovery engines. The system leverages LLMs for creative hypothesis generation while evolutionary algorithms test and refine ideas through massive parallel search.

This approach creates emergent capabilities where the combination discovers solutions that neither LLMs nor evolutionary algorithms could find independently. It’s analogous to having millions of creative scientists proposing ideas while natural selection identifies and improves the most promising concepts.

AlphaEvolve has already discovered novel solutions that surprised DeepMind researchers, demonstrating the power of hybrid architectures to explore genuinely new territories in solution space. The system doesn’t just optimize within known patterns but formulates new problems and hypotheses.

This hybrid approach may be key to achieving artificial general intelligence by combining the creative capabilities of language models with the optimization power of evolutionary search.

Building a Virtual Cell

DeepMind’s next major project aims to create a complete AI model of a biological cell, extending beyond AlphaFold’s protein structure success to encompass the full complexity of cellular biology. The project involves modeling proteins, RNA, DNA, metabolites, and their countless simultaneous interactions.

The team starts with simpler organisms like E. coli before tackling human cells, applying the same principles that made AlphaFold successful. Despite cellular complexity, cells follow rules shaped by evolution to be robust and efficient, creating learnable patterns in biochemical networks.

The virtual cell would enable scientists to run experiments impossible in real cells, such as testing gene knockouts or drug effects across all cellular pathways simultaneously. Researchers could perform billions of virtual tests before conducting real-world trials.

This project represents a shift from AI as automation tool to AI as genuine scientific discovery partner, capable of revealing new insights about fundamental biological processes.

P vs NP and the Nature of Computation

Hassabis connects AI success to fundamental questions in computer science and physics, suggesting the universe itself might be an information-processing system where computation and physics represent two sides of the same coin. He proposes that fundamental questions like P vs NP could be answered using AI itself.

The success of neural networks at modeling natural phenomena hints at deep truths about computation and reality. If nature efficiently solves seemingly intractable problems like protein folding, perhaps P does equal NP for a special class of “natural” problems shaped by evolutionary processes.

Hassabis envisions AI systems that don’t just solve problems but discover new mathematical structures and physical laws. Future AI might reveal hidden patterns in the space of all possible computations, similar to how AlphaFold revealed hidden patterns in protein space.

This perspective suggests that AI development is not just technological progress but a method for uncovering the computational principles underlying physical reality.

The Philosophy of Games and Reality

Hassabis views games as more than entertainment - they serve as laboratories for intelligence development. Games provide compressed versions of reality with clear rules and goals, making them perfect environments for developing and testing artificial intelligence.

Games have driven AI progress from the beginning, including chess, Go, StarCraft, and now open-world simulations. Each game presents a simplified but complete world that challenges different aspects of intelligence and strategic thinking.

As AI capabilities improve, the distinction between games and reality begins to blur. Virtual worlds are becoming indistinguishable from reality, filled with AI agents that approach human-level behavior and intelligence.

This raises profound questions about the nature of reality itself. If AI can create perfect simulations of physical worlds and human behavior, the boundaries between virtual and real experiences become increasingly meaningless.

Key Quotes