Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games | Lex Fridman Podcast #475

Demis Hassabis discusses the surprising ability of classical learning systems, such as Google DeepMind's Veo video generation model, to accurately model complex, nonlinear dynamical systems like fluid dynamics and materials. He highlights how these systems, by reverse engineering from vast amounts of data like YouTube videos, can extract underlying structures and behaviors, suggesting the existence of a lower-dimensional manifold that can be learned from natural systems.

Hassabis, in his Nobel Prize lecture, proposed that any pattern in nature can be efficiently discovered and modeled by a classical learning algorithm. He explains that this conjecture stems from processes like AlphaGo and AlphaFold, which build models of high-dimensional spaces to guide efficient search. He attributes this learnability to nature's inherent structure, shaped by evolutionary processes over vast timescales, from protein folding to cosmological phenomena. He also suggests the potential for a new complexity class for 'learnable natural systems' (LNS).

Hassabis views information as the fundamental unit of the universe, suggesting that the P versus NP problem becomes a physics question within this informational framework. He believes that classical learning systems, like neural networks, can solve a huge class of problems by modeling system dynamics to make search efficient, essentially achieving polynomial time solutions. He emphasizes that the capabilities of classical systems are far greater than previously imagined, citing AlphaFold and AlphaGo as examples that were once thought to require quantum computing.

Drawing from his background in game development, Hassabis envisions a future for video games where AI systems dynamically create interactive open worlds tailored to player imagination. He foresees "playable world models" that can adapt narratives and generate content on the fly, offering personalized and deeply engaging experiences akin to the ultimate 'choose your own adventure' games. This could lead to simulating mechanics and physics of a world, which is crucial for AGI development.

Hassabis discusses the promise of evolutionary techniques like AlphaEvolve, which combine large language models (LLMs) with evolutionary computing to propose solutions and explore novel areas of the search space. He sees these hybrid systems as crucial for scientific discovery and pushing the boundaries of knowledge. He also reflects on nature's evolutionary mechanism, a relatively simple algorithm that generated immense complexity over billions of years, suggesting that AI could help in understanding and simulating similar processes.

Hassabis shares his long-term dream of modeling a complete cell, starting with simpler organisms like yeast. He explains how projects like AlphaFold and AlphaFold 3 are foundational steps towards simulating the dynamics and interactions within a cell, aiming to accelerate biological research by enabling in silico experimentation. He also speculates on the possibility of simulating the origin of life itself, moving from chemical soup to the birth of a living organism, viewing it as a search process through a combinatorial space.

He highlights Google DeepMind's progress in weather prediction, with systems outperforming traditional fluid dynamics models by using neural networks to model complex weather dynamics more efficiently and accurately. He stresses the practical importance of these advancements for hurricane path predictions and other critical applications. This exemplifies AI's potential to solve global challenges and his overarching goal of building AGI to address fundamental questions about the universe and reality, and other large questions.

Hassabis provides his definition of AGI as matching the cognitive functions of the human brain, exhibiting consistency across all intelligence domains, and possessing true invention and creativity. He suggests testing AGI through a vast array of cognitive tasks and by having world experts try to find flaws. He also looks for 'lighthouse moments' like inventing a new physics conjecture or creating a game as deep and elegant as Go as definitive signs of AGI.

Hassabis discusses the continued scaling of compute, not just for pre-training but also for inference and "thinking systems" that improve with more processing time. He anticipates a future where AI systems are so pervasive that compute demand will only increase, driving innovation in hardware and energy solutions. He highlights the role of AI in optimizing energy usage, from data center cooling to fusion reactor design and new materials for solar technology, expressing optimism about fusion and solar as primary energy sources in the coming decades.

Hassabis reflects on leading Google DeepMind's intense efforts to advance AI, emphasizing the world-class talent and research culture that drove rapid progress in Gemini. He describes the challenges of managing bureaucracy in a large company while maintaining a startup's decisiveness and energy. He also draws on his game design experience in building AI-first products, focusing on simplifying interfaces and anticipating future technological capabilities, leading to the idea of AI-generated, personalized interfaces.

He addresses the 'war for talent' in AI, emphasizing that while competitive salaries are a factor, true innovators are driven by the mission to influence AI's development and ensure its safe and responsible deployment. He touches on the broader societal implications of AGI, including the potential for radical abundance and the need to rethink economic models like universal basic provision to ensure shared benefits.

Hassabis discusses the impact of AI on jobs, particularly in programming, noting that while AI can make some tasks easier, it will also create new roles and enhance human productivity. He anticipates a significant societal transformation, comparable to the Industrial Revolution but at a much faster pace, necessitating discussions on new governance structures and philosophical questions about human purpose and identity.

Hassabis discusses "The Maniac" by Benjamin Labatut and the legacy of John von Neumann, a polymath who foresaw the profound impact of computing. He believes von Neumann would be fascinated by today's AI advancements and emphasizes the need for thoughtful stewardship of this powerful technology. He hopes for a collaborative, CERN-like approach to AGI development, avoiding weaponization, and fostering international cooperation on safety and ethical standards.

Hassabis expresses his long-standing fascination with consciousness, viewing AI development as a means to understand what makes the human mind special. He leans towards the view that consciousness is likely a classical computational phenomenon, but acknowledges the 'hard problem' of qualia and the potential for unique experiences across different substrates (carbon vs. silicon). He advocates for a spiritual or humanist dimension in AI development, inspired by figures like Feynman and Da Vinci, who saw no boundaries between science, art, and philosophy, emphasizing the interconnectedness of technology and the human experience.