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

Demis Hassabis introduces the idea that classical learning systems, like Google DeepMind's Veo video generation model, can surprisingly model complex physical phenomena like fluid dynamics and material behavior. He conjectures that any pattern generated or found in nature can be efficiently discovered and modeled by classical learning algorithms due to underlying structures shaped by processes like evolution. He proposes a new complexity class, LNS (Learnable Natural Systems), for systems amenable to efficient modeling by classical AI.

Hassabis discusses the future of video games, envisioning AI systems that can dynamically create open-world experiences around a player's imagination, leading to deeply personalized narratives. He draws on his own experience in game development, highlighting how AI could enable games with true player choice and emergent behavior, moving beyond pre-programmed illusions of choice. He dreams of creating an incredibly realistic, open-world simulated game after achieving AGI, connecting it to the fundamental questions of the universe.

The discussion turns to AlphaEvolve, a Google DeepMind system that evolves algorithms, and the promising future of combining large language models with evolutionary computing and other search algorithms. Hassabis views this as a path to discovering novel solutions and pushing scientific boundaries, enabling systems to create and invent. He also shares his long-standing dream of modeling a living cell (a 'virtual cell'), starting with yeast, as a stepping stone to understanding life's origins, and how AlphaFold is a component of this endeavor.

Hassabis estimates a 50% chance of AGI by 2030, defining AGI as a system that can match the cognitive functions of the human brain across a wide range of tasks and exhibit true invention and creativity. He suggests testing for AGI through rigorous evaluation across thousands of cognitive tasks and by observing 'lighthouse moments' like discovering new physics conjectures or inventing games as profound as Go. He believes that the ability of AI to generate new and groundbreaking scientific ideas will be a key indicator of AGI.

The conversation addresses the critical role of compute scaling for AGI, encompassing training, inference, and new 'thinking systems.' Hassabis sees no slowdown in the demand for compute, fueled by AI's increasing capabilities and adoption across billions of users. He emphasizes DeepMind's focus on hardware innovations and using AI to optimize energy usage, from data center cooling to fusion reactor design and new material discovery for sustainable energy sources like solar. He predicts that within decades, fusion and efficient solar power could lead to an era of radical abundance, transforming society by alleviating resource scarcity.

Hassabis reflects on leading Google DeepMind's rapid progress with Gemini, emphasizing the importance of top talent, a strong research culture, and relentless shipping of products. He discusses the challenges of bureaucracy in a large company and the need to prioritize technical product development by anticipating future AI capabilities. He advocates for maintaining collaborative relationships with other AI labs, particularly on safety, and highlights the scientific endeavor as a crucial area for cooperation. He expresses concern about the 'war for talent' in AI but believes that a genuine mission and the opportunity to make a positive impact are more important motivators than just salary.

Hassabis acknowledges that AI will significantly transform the job market, making some tasks, like coding, more amenable to AI generation. He predicts that individuals who embrace AI technologies will become 'superhumanly productive,' but also that there will be a rapid shift in where human skills are most valued. He compares this societal shift to the industrial revolution, highlighting that AI's impact will be 10 times greater and 10 times faster, leading to profound changes that require economists and philosophers to proactively consider new governance structures and resource distribution models.

Hassabis discusses the foresight of figures like John von Neumann regarding computing's impact and the double-edged sword of scientific discovery. He hopes humanity learns from past experiences, like the Manhattan Project, to choose collaborative paths (like CERN) for AI development rather than weaponization. He touches on the 'p-doom' debate, emphasizing the extreme uncertainty and stakes involved, arguing for a strategy of 'cautious optimism' driven by scientific research into AI safety. He addresses the hard problem of consciousness, suggesting that building AI and comparing it to the human mind may be the best way to understand what makes human consciousness unique, even if it cannot be fully replicated in silicon.

Demis Hassabis expresses hope for the future of human civilization rooted in humanity's limitless ingenuity and extreme adaptability. He shares his belief that our ability to cope with complex modern life, despite having 'hunter-gatherer brains,' demonstrates our capacity to adapt to AI's transformative changes. He concludes by emphasizing the importance of human qualities like curiosity, compassion, and the ability to love, and the need for humanity to approach the development of AI with a 'spiritual or humanist dimension' to ensure it serves the greater good.