Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games | Lex Fridman Podcast #475
Demis Hassabis argues classical AI can model virtually all natural systems, challenging assumptions about quantum necessity and embodied learning.
- Hassabis conjectures any pattern in nature can be efficiently modeled by a classical learning algorithm — because natural systems have structure shaped by evolutionary or physical selection processes.
- Veo 3 models fluid dynamics, specular lighting, and material behavior by passively watching YouTube — challenging the theory that intuitive physics requires embodied, action-based learning.
- AlphaEvolve combines LLMs with evolutionary search to discover novel algorithms, potentially overcoming traditional evolutionary computing’s inability to generate genuinely new emergent properties.
- Hassabis frames P=NP as a physics question: if information is the most fundamental unit of reality (more so than energy or matter), then computability is a question about the universe’s structure.
- AI’s economic impact will be roughly 10x the industrial revolution in magnitude but 10x faster — a combined ~100x disruption — requiring new governance structures and possibly universal basic provision.
- On p(doom): refuses to give a number but calls the risk definitively non-zero and non-negligible; calls for 10x more safety research as AGI approaches.
- Classical computing, not quantum, is his bet for modeling consciousness; argues substrate difference between carbon and silicon will make AI consciousness verification fundamentally harder than behavioral tests alone.
- Post-AGI personal plans: build an open-world video game using vibe coding, and work on a personal physics theory linking information, P=NP, and the structure of reality.
Guests: Demis Hassabis, CEO of Google DeepMind, Nobel Prize winner (Chemistry 2024) · 2025-07-23 · Watch on YouTube
