#475 – Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games

Demis Hassabis, leader of Google DeepMind, shares insights on intelligence, the mysteries of the universe, and the emotional challenges faced by founders of rapidly growing companies. He discusses the efficiency of classical learning algorithms in modeling patterns in nature across various fields, including biology, chemistry, physics, and cosmology. Hassabis emphasizes the impracticality of brute force solutions for complex problems like protein folding, advocating for models that guide the search for solutions.

He introduces the concept of "survival of the stable," suggesting that evolutionary processes shape both living organisms and geological formations, extending to cosmological phenomena. The potential for a new complexity class, "learnable natural systems" (LNS), is explored, linking it to the P vs. NP question and its implications for physics and computer science. Hassabis highlights the success of AI models like AlphaGo and AlphaFold in efficiently modeling complex systems.

The conversation shifts to artificial general intelligence (AGI) and the role of neural networks on classical computers, raising questions about their capabilities. Hassabis notes advancements in video generation and physics rendering, which provide insights into the universe's structure. He reflects on the evolution of gaming, emphasizing AI's potential to create immersive game worlds and the challenges of programming open-world games.

Hassabis recalls his work on "Black and White," which featured early reinforcement learning AI, and discusses the relationship between gaming and significant intellectual challenges, such as the P vs. NP problem. He expresses nostalgia for the golden era of gaming and the evolution of the medium, particularly with open-world and simulation games.

The discussion touches on the importance of balancing virtual experiences with real-world interactions and the need for a scientific understanding of their value. Hassabis reflects on his favorite games and the connection between gaming skills and leadership in AI. He emphasizes the historical role of gaming in driving advances in AI and hardware, introducing "Alpha Evolve," a system that evolves algorithms.

The conversation also addresses the limitations of traditional evolutionary computing methods compared to modern AI, with aspirations to model natural systems, particularly cells. Hassabis introduces the concept of "research taste," emphasizing its significance in guiding scientific inquiry. He argues that generating meaningful conjectures is often more challenging than solving problems, highlighting the creativity required in scientific discovery.

The potential of AI in weather modeling is acknowledged, with advancements in predicting complex dynamics using neural networks. Hassabis estimates a 50% chance of achieving AGI by 2030, defining it as matching human cognitive functions across various domains. He discusses the necessity for AI systems to exhibit general intelligence and the excitement surrounding AI's capability to make new scientific discoveries.

Concerns about vague instructions for AI are raised, advocating for specific directives to yield better results. The conversation highlights the importance of scaling compute resources for building general intelligence and the potential for innovations in hardware. Predictions for energy sources in the coming decades include nuclear fusion and solar energy, with implications for advancements in various fields.

Hassabis explores the potential of space travel and the concept of "radical abundance," emphasizing the need for equitable resource distribution. He discusses the collaborative culture at Google DeepMind and the importance of maintaining connections among researchers. The conversation reflects on the dual potential of technology to solve global issues or lead to catastrophic outcomes, advocating for cautious optimism and international agreements on technology standards.

The nature of consciousness is debated, with references to neuroscience and the challenges of determining AI consciousness. Hassabis expresses hope for human ingenuity in facing modern challenges, including the integration of AI technology. The conversation concludes with reflections on the societal effects of technological advancements and the need for new governance structures to manage transitions effectively.