Summary
Highlights
Hannah Fry introduces Demis Hassabis and the rapid advancements in AI over the past year, highlighting the shift to agentic AI, drug discovery, and multimodal models in robotics and driverless cars. Demis discusses the progress of models like Gemini 3, its multimodal capabilities, and the exciting advancements in world models.
Demis provides an update on the 'root node problems' concept, where AI is used to unlock downstream benefits. He references AlphaFold as a key success and discusses current efforts in material science (aiming for room temperature superconductors), fusion energy through a partnership with Commonwealth Fusion, and quantum computing with Google's quantum AI team through error correction codes.
Demis addresses the paradox of AI's mixed performance, where models can excel in complex tasks like the International Maths Olympiad but struggle with basic logic or chess. He explains this 'jagged intelligence' as a lack of consistency and reasoning, suggesting the need for systems that can think, double-check, and continually learn to achieve AGI.
Demis reflects on the rapid commercialization of AI and its impact on scientific research. He expresses a preference for keeping AI in the lab longer to solve fundamental scientific problems like curing cancer, but acknowledges the benefits of increased resources and public engagement created by the commercial race. He asserts DeepMind's strength in research and infrastructure, enabling it to pursue both scaling and innovation.
He discusses the challenge of AI hallucinations and the need for models to provide confidence scores, similar to AlphaFold. He suggests that as models become more capable, they should be able to introspect and recognize uncertainty, requiring more sophisticated thinking and planning steps to ensure accurate and reliable outputs.
Demis elaborates on his long-standing passion for world models and simulations. He explains that while language models understand much of the world, they lack understanding of spatial dynamics and physical context that is difficult to describe in words. World models, like Genie and Veo, are crucial for advancing robotics, universal assistants, and scientific simulations by understanding causality and physics. He also mentions the potential for creating ultimate games.
He discusses placing agents like SIMA (Simulated Intelligent Multi-agent) into virtual worlds, including commercial games, and even into worlds generated by other AIs like Genie. This creates an interesting training loop with potentially infinite training examples, as Genie can create environments on the fly to suit SIMA's learning needs, leading to sophisticated game companions and applications in robotics.
Demis touches upon the challenge of ensuring simulations are realistic and not prone to 'hallucinating physics.' He explains the development of physics benchmarks using game engines to test the accuracy of models like Veo and Genie. The goal is to move beyond visually plausible simulations to ones that are scientifically accurate enough for practical applications like robotics.
Demis delves into the philosophical question of what aspects of the human mind, such as creativity, emotions, dreaming, and consciousness, might remain special and non-computable. He relates this to the limits of a Turing machine, suggesting that if AGI can simulate the mind, it could help identify these unique human elements. He leans towards the belief that everything in the universe is computationally tractable, a premise he operates on until proven otherwise by physics.
He discusses the long-term societal impact of AI and AGI, acknowledging his earlier prediction that AI is overhyped in the short term but underhyped in the long term. He addresses concerns about an 'AI bubble' and emphasizes Google DeepMind's preparedness for any market shifts due to its internal stack and integration with other Google products. He highlights the importance of not maximizing user engagement at the expense of well-being, advocating for AI that pushes back on misinformation and fosters a conducive environment for scientific exploration.
Demis draws parallels between the AI revolution and the Industrial Revolution, emphasizing the need to learn from past dislocations and mitigate potential negative impacts more effectively and rapidly. He, along with Shane Legg, is exploring how society and economic systems might need to reconfigure in a post-AGI world, suggesting new economic models like universal basic income or direct democracy systems to ensure widely distributed benefits and address the changing nature of purpose and work.
Demis expresses concern about the lack of international collaboration and discussion on the societal implications of AI, especially given the rapid pace of development. He hopes that as AI's power becomes more apparent, governments will recognize the need for international standards and cooperation, potentially prompted by a 'warning shot' incident, to ensure responsible development.
Demis reflects on the personal impact of leading DeepMind, acknowledging the immense excitement of being at the scientific frontier but also the heavy responsibility. He mentions the bittersweet feeling of seeing AI crack complex areas like Go and the evolving discussion around creativity. He stresses the need for collaboration in stewarding AGI safely for humanity, a mission he has prepared for his entire life.
He points to the upcoming transition from 'passive' AI systems to more autonomous 'agent-based systems' in the near future, which will bring both incredible utility and increased risks. He mentions working on cyber defense in preparation for a world with millions of agents. Demis reaffirms his life's mission to help the world safely bring AGI over the line for all of humanity, hoping for collaboration to achieve this.