This video features a conversation between David Nermberg, Director of the Institute for Advanced Study, and Sir Demis Hassabis, CEO and co-founder of Google DeepMind. The discussion centers on artificial intelligence (AI), its capabilities, potential risks, and its transformative impact on human knowledge and humanity. Nermberg begins by highlighting the Institute's long history of considering these questions, setting the stage for Hassabis to discuss his journey into AI and his vision for its future.
AI's Development Through Games: Sir Demis Hassabis's journey into AI began with his childhood fascination with games, particularly chess. He viewed games as microcosms of human thought, providing valuable testbeds for AI development. The clear metrics and data generation capabilities of games made them ideal for early AI progress.
Generalizable AI & Scientific Application: Hassabis emphasizes the importance of developing generalizable AI algorithms applicable to real-world problems, particularly in science and medicine. AlphaGo's success demonstrated the potential of AI to tackle complex problems beyond games, leading to the development of AlphaFold for protein structure prediction.
The Importance of Data & Clear Metrics: Successful AI applications require large datasets, clear metrics for optimization, and the ability to handle massive combinatorial spaces. AlphaFold's success relied on existing protein structure data, highlighting the need for continued public investment in fundamental scientific research.
AGI & Multimodal AI: DeepMind's ultimate goal is Artificial General Intelligence (AGI), a system capable of exhibiting all human cognitive capabilities. This necessitates the development of multimodal AI systems that can understand and interact with the world in a comprehensive way, including through language, images, and intuitive physics.
Risks & Responsible Development: Hassabis acknowledges the potential risks of AI, including misuse by bad actors and inherent risks within increasingly autonomous systems. He advocates for international cooperation and the development of new institutions to govern and regulate AI development responsibly.
According to Sir Demis Hassabis, the protein folding problem was suitable for AlphaFold because it had a relatively clean dataset (around 150,000 known structures), a clear metric (minimizing free energy), and represented a massive combinatorial space—making it ideal for the model-building and intelligent search techniques employed by AlphaFold. The problem's foundational importance and potential impact on drug discovery also contributed to its selection.
Sir Demis Hassabis identifies three key characteristics for selecting problems suitable for AI research: (1) availability of (real or simulated) data, (2) clear metrics for optimization, and (3) a massive combinatorial space that necessitates an intelligent search process rather than brute-force methods.
Sir Demis Hassabis's journey into AI began with his childhood fascination with games, particularly chess. This early exposure provided a foundation for understanding the intricacies of human thought processes. His neuroscience background contributed to his understanding of the brain's workings as a potential model for AI, emphasizing that both are fundamentally classical systems capable of achieving surprising levels of generality. His focus on building general-purpose AI systems directly stems from this background and the desire to apply AI to complex scientific and medical problems.
Sir Demis Hassabis highlights two major risks associated with AI advancement: (1) the potential for misuse by bad actors (individuals or nations) repurposing the technology for harmful purposes, and (2) inherent risks within increasingly autonomous and agentic AI systems themselves, requiring careful control, guardrails, and understanding to prevent unintended consequences.
Here are some additional topics discussed in the conversation with Sir Demis Hassabis, presented as questions and answers:
Q: What is the P versus NP problem, and why is it of interest to you?
A: The P versus NP problem is a fundamental question in computer science concerning the difference between problems solvable in polynomial time (P) and problems whose solutions can be verified in polynomial time but not necessarily found in polynomial time (NP). It gets to the heart of what's computationally possible on classical machines. Hassabis finds it fascinating because it's a foundational question about computation and its limits, particularly in light of recent AI advancements. He views their work at DeepMind as contributing to understanding the capabilities of classical computation.
Q: What prompted you to formulate the conjecture that any pattern in nature can be efficiently modeled by a classical learning algorithm?
A: This conjecture stems from Hassabis' observation that many interesting natural systems have evolved to a state of stability. This stability implies underlying structure, not randomness, that a sufficiently powerful learning algorithm should be able to model given enough data. He uses the examples of protein folding, the search for room-temperature superconductors, and drug discovery as potential applications of this idea. The conjecture is still under development and may change.
Q: How do you account for the surprising success of neural networks in solving computationally difficult problems?
A: Hassabis's neuroscience background informs his perspective. He points out that as far as we know, the human brain is a classical system. Yet, it achieves remarkable generality and problem-solving abilities. The success of neural networks, therefore, isn't entirely surprising since they are also classical systems, mirroring certain aspects of brain architecture. He finds the extent of their capabilities quite astounding, emphasizing that the brain's ability to create complex structures and solve problems far exceeds initial expectations.
Q: What are the differences and challenges between model-specific AI (like AlphaFold) and more general multimodal AI systems aiming for AGI?
A: DeepMind's ultimate aim is AGI, a system exhibiting all human cognitive capabilities. Model-specific AI, like AlphaFold, focuses on mastering particular tasks, whereas multimodal AI aims at building a comprehensive world model capable of understanding and interacting with the world through various modalities (language, vision, etc.). Building a world model is crucial for achieving AGI, as it requires the AI system to possess intuitive physics, spatial understanding, and other capabilities humans possess effortlessly.
Q: What non-technological steps should societies take to navigate the risks of developing AI?
A: Hassabis stresses the need for new institutions, possibly modeled on CERN or the IAEA, to promote international cooperation, monitor potentially dangerous projects, and establish responsible governance for AI development. He highlights the dual-use nature of AI and emphasizes the need to balance the benefits with the risks. He also advocates for a wise council or a technical UN to guide these efforts. The critical need for international collaboration is underscored by the challenges of restricting access to powerful AI systems by bad actors in today's geopolitical landscape.
