Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games
Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games
Summary
This is the second time Demis Hassabis appears on the Lex Fridman podcast, now as a Nobel Prize winner for his groundbreaking work in protein structure prediction using AI. Demis is the leader of Google DeepMind and one of the most brilliant minds working on understanding and building intelligence.
The conversation explores learnable patterns in nature, computation and P vs NP, how video generation models like Veo 3 are learning to understand reality, the path to AGI, scaling laws, the future of energy to power AI, and Google’s position in the race to AGI. They also discuss the origin of life, consciousness, quantum computation, and what it means to simulate a biological organism.
Highlights
”Video models are reverse engineering physics from YouTube”
Clip command
yt-dlp --download-sections "*14:34-15:34" "https://www.youtube.com/watch?v=-HzgcbRXUK8" --force-keyframes-at-cuts --merge-output-format mp4 -o "-HzgcbRXUK8-14m34s.mp4"“If you look at Veo, our video generation model, it can model liquids surprisingly well - materials, specular lighting. I used to write physics engines and graphics engines in gaming. It’s painstakingly hard to build programs that do that. Yet these systems are reverse engineering from just watching YouTube videos.” — Demis Hassabis, 14:26
”Perhaps most of reality has a lower dimensional manifold”
Clip command
yt-dlp --download-sections "*14:34-15:34" "https://www.youtube.com/watch?v=-HzgcbRXUK8" --force-keyframes-at-cuts --merge-output-format mp4 -o "-HzgcbRXUK8-14m34s.mp4"“Perhaps there is some kind of lower dimensional manifold that can be learned if we actually fully understood what’s going on under the hood. That’s maybe true of most of reality.” — Demis Hassabis, 14:26
”P(doom) and human resilience”
Clip command
yt-dlp --download-sections "*122:50-123:50" "https://www.youtube.com/watch?v=-HzgcbRXUK8" --force-keyframes-at-cuts --merge-output-format mp4 -o "-HzgcbRXUK8-122m50s.mp4"“If p(doom) is actually high at some point, all of humanity will be aligned in making sure that’s not the case. There’s a self-modulating aspect. I have a lot of faith in humanity rising up to meet that moment.” — Demis Hassabis, 2:02:50
Key Points
- Learnable Patterns in Nature (2:06) - What AI can and cannot learn
- Computation and P vs NP (5:48) - Fundamental limits of computation
- Veo 3 and Understanding Reality (14:26) - How video models learn physics
- Video Games (18:50) - Demis’s early career in gaming
- AlphaEvolve (30:52) - Evolution of AI systems
- AI Research (36:53) - Current state of the field
- Simulating Biological Organisms (41:17) - Could we simulate a cell?
- Origin of Life (46:00) - How life began
- Path to AGI (52:15) - What’s needed to achieve AGI
- Scaling Laws (1:03:01) - Are they still holding?
- Compute (1:06:17) - The importance of computational resources
- Future of Energy (1:09:04) - Powering AI systems
- Google and the Race to AGI (1:17:54) - Competition among labs
- Competition and AI Talent (1:35:53) - Attracting the best researchers
- Future of Programming (1:42:27) - How AI will change coding
- John von Neumann (1:48:53) - The legendary polymath
- P(doom) (1:58:07) - Probability of AI catastrophe
- Consciousness and Quantum Computation (2:05:56) - Deep questions about mind
- David Foster Wallace (2:12:06) - Literature and meaning
- Education and Research (2:19:20) - Advice for aspiring researchers
Mentions
Companies
- Google DeepMind - Where Demis leads AI research
- Isomorphic Labs - Demis’s drug discovery company
Technologies
- AlphaFold - Protein structure prediction (Nobel Prize work)
- Veo 3 - Google’s video generation model
- Gemini - Google’s flagship AI model
- AlphaEvolve - AI system evolution