Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games | Lex Fridman Podcast #475

In his second appearance on the Lex Fridman Podcast, Demis Hassabis lays out his conjecture that any pattern generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm, citing AlphaFold, AlphaGo, and the surprising physics understanding inside Google's Veo video model. He discusses his lifelong love of video games and his dream of AI-generated open-world games, the AlphaEvolve system that combines LLMs with evolutionary search, and his vision of modeling an entire virtual cell. Hassabis estimates roughly a 50% chance of AGI within five years and describes how he would test for it, including 'move 37' style lighthouse moments such as a system inventing a new scientific conjecture. He covers Google's turnaround from Gemini 1.5 to 2.5, the future of energy via fusion and solar, the disruption AI will bring to jobs, AI safety and bad actors, and consciousness and what makes humans special. The conversation closes with Lex's solo reflection on David Foster Wallace's 'This Is Water' speech.

• Hassabis explains his Nobel lecture conjecture: any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm, because natural systems have structure from evolutionary processes.
  00:02:36
• He argues Veo (V3) demonstrates genuine intuitive physics understanding of liquids, lighting, and materials learned purely from passive observation of YouTube videos, challenging the idea that embodiment is required to understand the physical world.
  00:14:32
• He envisions AI-generated open-world video games within 5-10 years, an interactive version of Veo that dynamically generates content and narrative around the player, building toward a true 'world model' needed for AGI.
  00:20:14
• Hassabis details his ~25-year dream of building a 'virtual cell' (starting with a yeast cell) to run experiments in silico, potentially 100x speeding up wet-lab work, with AlphaFold and AlphaFold 3 as building blocks.
  00:42:32
• He estimates roughly a 50% chance of AGI within the next five years (by 2030), defining it as matching the brain's general cognitive functions with consistency rather than today's 'jagged' intelligence.
  00:52:29
• He describes 'lighthouse' tests for AGI like a 'move 37' moment: back-testing whether a system given only pre-1900 knowledge could derive special and general relativity, or inventing a game as deep and elegant as Go.
  00:55:06
• Hassabis explains how Google DeepMind went from 'losing' with Gemini 1.5 to 'winning' with Gemini 2.5 by merging Google Brain and DeepMind, fostering a startup-like research culture and 'relentless progress along with relentless shipping.'
  01:18:00
• He refuses to give a 'p(doom)' number, calling it false precision, but says the risk is definitely nonzero and non-negligible, advocating proceeding with 'cautious optimism' and roughly 10x more safety research.
  01:58:08

• Hassabis says he spends his few moments of spare time working with colleagues on whether there should be a new complexity class of problems solvable by neural-network processes, viewing P vs NP as fundamentally a physics question because information is primary in the universe.
  00:06:16
• His favorite video games of all time are Civilization 1 and Civilization 2, and he deliberately avoids the most recent Civilization because it would consume too much of his time.
  00:29:01
• AlphaEvolve combines LLMs (proposing solutions) with evolutionary computing (searching novel parts of the space); AlphaGo's Monte Carlo tree search is what produced the famous never-before-seen 'move 37' strategy.
  00:32:39
• Hassabis notes Google DeepMind built the best weather prediction systems in the world (WeatherNext), outperforming traditional fluid-dynamics supercomputer methods and recently predicting cyclone and hurricane paths.
  00:49:50
• DeepMind has its own TPU line and is exploring inference-only chips, plus AI for data-center cooling, grid optimization, and plasma containment for fusion reactors (work with Commonwealth Fusion).
  01:08:04
• Hassabis cites his game design days creating games for millions of gamers as the origin of his product taste, including the early reinforcement-learning creature AI in the game Black & White.
  01:21:08
• On the talent war, he recalls that in 2010 he didn't pay himself for a couple of years because DeepMind couldn't raise money, whereas today interns are paid the equivalent of DeepMind's entire first seed round.
  01:40:24
• He predicts AI's impact will be about 10x the industrial revolution but happening 10x faster (roughly 10 years instead of 100), and calls for economists and philosophers to consider universal basic provision.
  01:45:34
• Hassabis says he and Roger Penrose 'cordially disagree' on consciousness; his bet is that the brain runs mostly classical computation, meaning consciousness phenomena should be modelable by a classical computer.
  02:06:28
• DeepMind built 'DolphinGemma,' a version of its system trained on dolphin and whale sounds, with hopes of eventually building an interpreter or translator for animal communication.
  02:09:06

Books, products and media the guest or host genuinely endorsed here — with the buy link.

Affiliate link — we may earn a commission at no extra cost to you.