Google DeepMind CEO Demis Hassabis: AI's Next Breakthroughs, AGI Timeline, Google's AI Glasses Bet
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this wide-ranging conversation from Davos, Demis Hassabis, CEO of Google DeepMind, addresses skepticism about AI progress and outlines his vision for achieving AGI within 5-10 years. (00:50) He discusses the current limitations of large language models, particularly their inability to continually learn, while sharing insights on breakthrough technologies like AI glasses and video generation models. (10:16) Hassabis provides a clear definition of AGI as systems that can exhibit all human cognitive capabilities, including the highest levels of creativity and scientific breakthrough - not just problem-solving but theory creation. (07:07) The conversation covers Google's product strategy, from smart glasses partnerships to advertising concerns, while exploring the transformative potential of AI across scientific discovery and knowledge work.
- Core themes include the path to AGI, the importance of continual learning and world models, Google's AI product roadmap, and the philosophical implications of artificial intelligence reaching human-level capabilities
Speakers
Demis Hassabis
Demis Hassabis is the CEO of Google DeepMind, formed through the merger of Google AI and DeepMind. He's a renowned AI researcher who co-founded DeepMind in 2010 before its acquisition by Google in 2014. Hassabis has led groundbreaking AI projects including AlphaGo, AlphaFold, and the development of the Gemini large language models, establishing himself as one of the world's leading voices in artificial intelligence research and development.
Key Takeaways
AGI Requires True Creative Breakthrough, Not Just Problem-Solving
Hassabis provides a rigorous definition of AGI that goes far beyond current capabilities. (07:07) He emphasizes that AGI must exhibit all human cognitive capabilities, including the highest levels of creativity - not just solving existing problems but creating entirely new theories like Einstein's relativity or revolutionary art forms like Picasso's innovations. Current AI systems, despite their impressive performance on specific tasks, remain "nowhere near" this level of true creative breakthrough. This perspective challenges the rush to declare AGI achievement based on benchmark performance and establishes a meaningful scientific standard for measuring genuine artificial general intelligence.
Continual Learning Remains the Critical Missing Piece
The inability of current AI systems to learn and retain information beyond individual sessions represents a fundamental limitation. (04:48) Hassabis acknowledges this "goldfish brain" problem where models can search and process information but cannot permanently update their knowledge base. While Google DeepMind has achieved continual learning in narrow domains like games through systems like AlphaZero, scaling this to real-world applications remains unsolved. (05:06) The breakthrough in continual learning would enable true personalization and allow AI assistants to develop deeper understanding of users over time, moving beyond simple context window storage to genuine model evolution.
World Models Are Essential for Long-Term Planning
Video generation models like Google's Veo represent more than creative tools - they're stepping stones toward "world models" that understand physics and causality. (10:16) These systems demonstrate intuitive physics understanding by generating realistic scenes where liquids and objects behave correctly. Hassabis explains that such world models are essential for AGI because they enable long-term planning in the real world, from simple robotic tasks to complex human-like planning spanning years. (11:12) Without this capability to model consequences over extended time horizons, AI systems remain limited to short-term problem-solving rather than the strategic thinking that defines human intelligence.
AI Glasses Represent the Next Major Computing Platform
The smartphone form factor is fundamentally inadequate for seamless AI interaction in daily life. (12:29) Hassabis reveals that Google's internal testing confirmed the obvious limitation of holding up phones to interact with the real world. AI glasses solve this through hands-free operation, particularly valuable for cooking, navigation, and accessibility applications. (13:30) Unlike Google Glass's previous failure due to lacking a compelling use case and clunky hardware, current AI capabilities provide the "killer app" of a universal digital assistant. Partnerships with Warby Parker, Gentle Monster, and Samsung signal serious commercial intent, with consumer availability expected by summer 2024.
Information as the Fundamental Building Block of Reality
Hassabis proposes a profound philosophical framework viewing information, rather than energy or matter, as the universe's most fundamental unit. (27:52) Living systems resist entropy by maintaining their informational structure, and this principle extends beyond biology to geological formations and planetary systems. (28:33) This perspective directly informs AI development strategies, particularly in projects like AlphaFold, where understanding the "information landscape" of stable protein structures makes seemingly intractable problems solvable. This framework suggests that AI's ultimate power lies in navigating complex information topologies to discover solutions in drug development, materials science, and fundamental physics.
Statistics & Facts
- Google DeepMind's AlphaFold has been used by 3 million researchers worldwide, with Hassabis predicting that almost every drug discovered from now on will likely use AlphaFold at some point in the development process. (31:00)
- The best Go player in the world is currently a South Korean player who was about 15 when the AlphaGo match occurred and is now in his mid-twenties, representing the first generation to learn natively with AlphaGo knowledge incorporated. (25:16)
- According to Hassabis, Gemini Flash models can run AI computing as cheaply as traditional search operations, potentially disrupting current infrastructure investment models in the AI industry. (20:32)
