Demis Hassabis highlights DeepMind’s balanced approach to advancing AI through both scaling and innovation, emphasizing the importance of simulated worlds for enhancing AI’s reasoning and learning capabilities beyond language understanding. He also discusses the societal challenges posed by AI’s transformative potential, advocating for responsible development, international cooperation, and new social models to ensure AI benefits all of humanity.
In this insightful discussion, Demis Hassabis, Co-founder and CEO of DeepMind, reflects on the rapid advancements in AI over the past year, highlighting the release of Gemini 3 and progress in world models. He emphasizes the dual approach DeepMind takes, balancing 50% of their effort on scaling AI systems and 50% on innovation, believing both are essential to achieving Artificial General Intelligence (AGI). Hassabis also discusses the importance of using AI to solve root node scientific problems, such as AlphaFold’s breakthrough in protein folding, and expresses excitement about ongoing projects in material science, fusion energy, and quantum computing collaborations.
Hassabis addresses the paradox in AI’s mathematical capabilities, where models can excel at complex problems like those in the International Math Olympiad but still make basic errors in simpler tasks. He attributes this inconsistency to current limitations in reasoning, consistency, and the ability to learn continuously in real-world environments. He draws parallels to the development of AlphaGo and AlphaZero, suggesting that future AI systems might similarly evolve to discover knowledge independently, but notes that current models still lack the ability to learn online and adapt continuously after training.
A significant portion of the conversation focuses on the potential of simulated worlds and world models, which Hassabis considers crucial for advancing AI beyond language understanding to spatial and physical reasoning. He explains how DeepMind’s projects like Genie and Simma integrate AI agents into dynamic, interactive virtual environments, enabling them to learn through experience and exploration. These simulations not only have applications in gaming and robotics but also hold promise for scientific research by modeling complex systems such as weather or biological processes, potentially accelerating discovery through controlled, repeatable experiments.
Hassabis also reflects on the societal implications of AI and AGI, drawing lessons from the Industrial Revolution about the transformative yet disruptive nature of technological progress. He stresses the need for new economic and social models to address the profound changes AI will bring, including ideas like universal basic income and more participatory democratic systems. He expresses concern about the current fragmented state of international cooperation on AI governance and the risks posed by competitive pressures and rogue actors, advocating for responsible development and collaboration to ensure AI benefits all of humanity.
Finally, Hassabis shares personal reflections on the emotional weight and excitement of leading AI research at the frontier of science. He acknowledges the challenges and ethical considerations involved, especially as AI systems become more autonomous and agent-based. Despite the immense responsibility, he remains passionate and optimistic about the future, viewing his mission as helping to safely steward AGI for the benefit of humanity. He envisions a future where AI systems integrate diverse capabilities—from language to vision to world modeling—towards a unified, general intelligence, while also recognizing the philosophical questions about consciousness and the limits of computation that remain open.