Hassabis, Demis
▶ Watch on YouTubeCoverage note: full (main interview; excludes Lex's separate closing AMA segments on David Foster Wallace and his own MIT/Drexel background, which are not part of the Hassabis conversation)
Core ideas
Hassabis opens by defending a deliberately provocative conjecture from his Nobel lecture: any pattern that exists in nature — because it was shaped by some selection or "survival of the stablest" process (biological evolution, geological weathering, even planetary orbits) — should in principle be efficiently learnable by a classical neural network, because such patterns have exploitable structure rather than being random. This is why AlphaGo and AlphaFold could tame search spaces (10^170 Go positions, 10^300 possible protein structures) that are naively intractable: the model learns the underlying structure/manifold well enough to guide search into tractable, near-polynomial territory. He explicitly frames this as a physics-flavored angle on P vs. NP — since he considers information more fundamental than energy or matter, the boundary of what's efficiently learnable by classical (non-quantum) systems is, to him, a question about the structure of reality itself, and he says he's informally working on formalizing a new complexity class around it. He extends this to Veo 3's uncannily accurate rendering of fluid dynamics, materials, and lighting — traditionally considered numerically brutal (Navier-Stokes) — arguing this is evidence the model has learned real "intuitive physics" (child-like, not equation-level) purely from passive video observation, which surprised him given his prior assumption (rooted in embodied-cognition theories like "action and perception") that understanding physics required acting in the world, not just watching it.
On games, a lifelong passion predating his AI career, Hassabis describes his 1990s work (Theme Park, Black & White) on open-world simulations where players co-create the narrative, and argues today's generative video/world models (an "interactive Veo") could finally deliver true dynamic, personalized open-world storytelling rather than the illusion-of-choice branching of older games — something he'd love to build himself, possibly via vibe-coding as a side project or a literal post-AGI sabbatical (alongside working on his P-vs-NP physics theory). On AlphaEvolve, he describes it as a promising hybrid paradigm — LLMs proposing candidate solutions, evolutionary/Monte Carlo search refining and combining them (as MCTS found Move 37 in AlphaGo) — that can push past a model's known data into genuinely novel territory, though current systems still only do incremental "hill-climbing" (e.g., faster matrix multiplication) rather than inventing something structurally new like the Transformer itself; whether a future leap requires new breakthroughs or "just" more scaling is, in his view, a genuine 50/50 open question DeepMind is hedging by pursuing both simultaneously. He discusses his 25-year "virtual cell" dream (starting with a fully simulated yeast cell, building up from AlphaFold's static structures to AlphaFold 3's pairwise interaction modeling to eventual whole-pathway and whole-organism simulation) as a way to speed up biology ~100x by moving most experimentation in silico before wet-lab validation, and speculates AI could eventually help model the origin of life itself as a search process through primordial chemical soup.
On AGI, Hassabis holds roughly a 50% chance by 2030, defining it as consistent, jagged-free performance matching the brain's full generality — tested via massive batteries of cognitive tasks plus stress-testing by top domain experts (a "Terence Tao test") for a month or two to find no obvious flaws — and stresses that the real signature moment ("Move 37 for AGI") would be a system producing a genuinely novel scientific conjecture (e.g., a back-tested Einstein-level insight generated from only pre-1900 knowledge) or inventing an entirely new game as elegant as Go, since research "taste" (picking the right question, the right conjecture that meaningfully splits the hypothesis space) is what separates great from merely competent scientists and remains something today's systems clearly cannot do. He credits Google DeepMind's turnaround from lagging to leading in the LLM race to merging Google Brain and DeepMind's research cultures around Gemini, staying "still a start-up" in mentality despite scale, and treats interface design (skeuomorphic chat boxes feeling increasingly "archaic," likely replaced by multimodal, personalized, possibly AI-generated UIs) as a genuine unsolved creative problem on par with model research. On societal impact, he expects programming and other domains to see a bifurcation where practitioners who fully embrace AI tools become "10x" more productive while lower-value tasks get automated, predicts change "10x the scale of the Industrial Revolution happening 10x faster," and calls for new economic/governance structures (something like universal basic provision funded by a coming era of energy- and resource-abundance via fusion and solar) developed in parallel by economists and political scientists.
On existential risk, Hassabis refuses to give a specific "p(doom)" number, calling that false precision, but says the risk is "definitely non-zero and probably non-negligible," and describes his stance as "cautious optimism" — pursuing AI's enormous upside (curing disease, solving energy scarcity, enabling humanity's expansion into space and Carl Sagan-style "waking up the universe") while treating both misuse by bad actors and loss-of-control risk from more autonomous, agentic systems as serious, parallel problems requiring far more (he says ~10x more) safety research as AGI approaches, and possibly a CERN-like international collaborative body rather than a Manhattan-Project-style arms race. Discussing consciousness with reference to Roger Penrose (whom he respectfully disagrees with) and the late Daniel Dennett, he argues the brain is very likely doing ordinary classical computation, not exploiting quantum mechanical effects, and that a key reason humans currently assume other humans are conscious is not just similar behavior but running on the same physical substrate — meaning silicon-based AI exhibiting conscious-seeming behavior poses a genuinely new "radical empathy" problem we've never had to confront, possibly addressable in the future via brain-computer interfaces letting humans literally feel what nonbiological computation is like.
Excerpts
On why natural systems, unlike arbitrary abstract problems, should be learnable by classical neural networks:
“In nature, natural systems have structure because they were subject to evolutionary processes that shape them. And if that's true, then you can maybe learn what that structure is... it may not be possible for man-made things or abstract things like factorizing large numbers because unless there's patterns in the number space... there's no pattern to learn, there's no model to learn that will help you search. So you have to do brute force.”
On what surprised him most about Veo 3's physical realism, against his own prior theoretical expectations:
“There's a lot of theories in neuroscience, it's called action in perception, where you need to act in the world to really, truly perceive it in a deep way... But it seems like you can understand it through passive observation, which is pretty surprising to me. And again, I think hints at something underlying about the nature of reality.”
On why he refuses to state a p(doom) number, and his resulting stance:
“I don't have a P-Doom number. The reason I don't is because I think it would imply a level of precision that is not there... it's definitely non-zero and it's probably non-negligible... under those conditions of a lot of uncertainty, but huge stakes both ways... the only rational, sensible approach is to proceed with cautious optimism.”