Hassabis: anything nature evolved can be efficiently learned by classical computers

1. Models guide the search — ▶ 2:41

   AlphaGo and AlphaFold are models of huge combinatorial spaces: 10 to the 300 possible protein structures, 10 to the 170 Go positions. Enumeration would outlast the universe. The model guides the search and makes it tractable. Proteins fold in milliseconds, so physics clearly has a shortcut. Hassabis's conjecture is that natural systems have structure because evolutionary processes shaped them, so that structure can be learned. He calls it a deliberately provocative Nobel-lecture claim, then spends the interview defending it.

2. Survival of the stablest — ▶ 4:01

   Selection is not just biology. Mountains are shaped by weathering, and orbits and asteroids by repeated surviving processes. Survivors are not random, so there is a manifold to reverse-learn. The stated exception is man-made or abstract problems like factoring large numbers: if the number space is uniform there is nothing to learn, so you brute-force it, or maybe need a quantum computer. Neural networks excel at following gradients, provided someone specifies the objective function correctly.

3. A new complexity class — ▶ 5:48

   In spare moments, Hassabis is working with colleagues on a possible new complexity class: natural systems that a neural network can model, making search or next-step prediction polynomial-time on a classical machine. It fits his view that information, not energy or matter, is the universe's most fundamental unit. If the universe is informational, P equals NP becomes a physics question. He says that is part of why he wants AGI: to help scientists answer exactly this kind of question.

4. Classical systems underrated — ▶ 8:47

   A neural network is a Turing machine in effect, and Hassabis says the field has already shown classical systems go much further than expected. Protein folding and superhuman Go were thought to be decades off, or to need quantum machines. He thinks the surface is barely scratched. The limits are open. Cellular automata can probably be forward-simulated efficiently, but chaotic systems, where initial conditions dominate and the end state decorrelates, could be hard. He says the boundary is exactly the interesting question.

5. Veo learns physics passively — ▶ 13:16

   Fluid dynamics and weather are classically expensive. Yet Veo handles liquids, materials and specular lighting, learning from video alone. Hassabis, who wrote physics engines in his gaming days, finds that striking. He reads accurate next-frame prediction as a form of understanding: intuitive physics, like a child's rather than a PhD student's. Five or 10 years ago he would have said that required embodiment or simulated action. Passive observation turns out to be enough, which he says points to a world model and, eventually, AGI.

6. Open worlds that adapt — ▶ 19:25

   Games were his first professional AI work. He built open-world simulations like Theme Park and Black & White, an early reinforcement-learning creature that mirrored how you treated it. Hard-coded A/B branching is only an illusion of choice. Within 5-10 years he expects an interactive Veo-style system that creates around the player's imagination and reshapes the narrative on the fly. His post-AGI projects are a vibe-coded game and his physics theory, both aimed at the same question about the universe.

7. Search takes you past data — ▶ 30:56

   A model captures what is known. To go beyond it, you add search. Monte Carlo Tree Search found move 37 in Go. AlphaEvolve pairs LLM proposals with evolutionary computing, and he sees promise in hybrids of foundation models with search methods. Evolution may add emergent capabilities, which 1990s evolutionary computing never managed, and foundation models might fix that. The search still needs an objective function to hill-climb, or the space is too big to search randomly.

8. Taste is the gap — ▶ 38:04

   Picking the right question is the hardest part of science, and a good conjecture is harder than its proof. Systems may soon solve Math Olympiad problems, as AlphaProof's silver medal showed, and perhaps a Millennium Prize problem. But a conjecture Terence Tao would call deep must be falsifiable, testable with current technology, and split the hypothesis space in two so either outcome teaches you something. Hassabis says today's systems clearly cannot do this, naive search on a model will not be enough, and nobody is sure what the mechanism would be.

9. Virtual cell via yeast — ▶ 42:10

   Hassabis has wanted to model a cell for about 25 years, since talks with Paul Nurse. He would start with yeast, a whole organism in one cell. AlphaFold gave static structures, AlphaFold 3 starts on interactions, then a pathway like TOR, then the cell. The cutoff would be the protein level, not atomic, with hierarchical simulators for different timescales. The payoff: perhaps 100X faster experiments, with search in silico and validation in the wet lab. He thinks AI could eventually tackle the origin of life too.

10. How he'd test AGI — ▶ 52:20

    He puts 50% on AGI by 2030, defined as matching the brain's cognitive functions consistently, without today's jagged intelligence. The test is brute force: tens of thousands of cognitive tasks, plus a few hundred top experts given a month or two to find an obvious flaw. Then lighthouse moments. Train on knowledge up to 1900 and see if it derives special and general relativity. Ask it to invent a game as elegant as Go. A system doing several of these, not one, would count.

11. Hill-climbing, not leaps — ▶ 59:39

    AlphaEvolve delivers incremental gains like faster matrix multiplication when the target is specific. Vague instructions, such as "make a better version of yourself", are too unconstrained. Recursive self-improvement is possible end-to-end, but he is unsure a hard takeoff is even desirable. Could such a system have invented the Transformer in 2017? Not clear. Nobody has shown unequivocally a big leap. Whether incremental S-curve climbing is enough, or one or two more breakthroughs are needed, is the open question.

12. Scaling is a 50/50 bet — ▶ 1:03:04

    Pre-training, post-training and inference-time thinking are all still scaling. Hassabis says it is about 50/50 whether new ideas are needed or scaling the existing recipe suffices, so DeepMind puts roughly half its resources on each. He says he likes it when the terrain gets harder, because research plus engineering is the lab's sweet spot, and he claims 80-90% of the breakthroughs underpinning modern AI came from Google Brain, Google Research or DeepMind. On data, he is not worried, because real data can seed simulators that generate synthetic data.

13. Compute, energy, abundance — ▶ 1:06:18

    Compute demand keeps rising: training often needs co-location, billions of users need inference, and thinking models improve with test-time compute. Training may become the smaller share, and DeepMind is exploring inference-only chips. He bets on fusion and solar as the two main clean sources, with AI helping on plasma containment, materials and batteries within about five years. Cheap energy would unlock desalination, rocket fuel from seawater and asteroid mining, a "radical abundance" era where the open question is fair sharing.

14. 10x faster than industrialisation — ▶ 1:42:30

    Hassabis expects roughly 10 times the impact of the Industrial Revolution at 10 times the speed: 10 years instead of 100. Over the next 5-10 years, programmers who embrace the tools could become 10X more productive, and skills like specifying architecture and checking output retain value. The harder problem is societal. He wants top economists and philosophers working now on distribution, including universal basic provision, and expects new governance institutions. His order of operations is to create abundance first, then decide how to share it.

15. No p(doom) number — ▶ 1:58:10

    He declines to give a p(doom), because a number implies precision that does not exist. His position is that it is "definitely non-zero" and "probably non-negligible", with huge uncertainty about takeoff speed and controllability. The answer is cautious optimism, and roughly 10 times more research effort on defining and addressing the risks. He splits the threat into bad actors repurposing general technology and agentic systems escaping control, and admits he has not heard a clear way to restrict bad actors while keeping access open to good ones.

Nature is not random, and that is why it can be learned. Any system shaped by selection over time carries structure, and a neural network on an ordinary classical computer can recover that structure far more efficiently than anyone assumed. That is the conjecture Demis Hassabis put in his Nobel lecture, and almost everything else he says hangs from it.

The evidence is work DeepMind has already done. A protein has something like 10 to the 300 possible structures. Go has 10 to the 170 possible positions. Both numbers dwarf the atoms in the universe, so enumeration is hopeless. Yet proteins fold in milliseconds inside a body, which proves the problem has a shortcut. AlphaGo and AlphaFold worked the same way. Build a model of the environment, then let the model guide the search. The model turns an impossible space into a tractable one.

Why would the shortcut exist? Hassabis calls it "survival of the stablest". Evolution shaped living things, but weathering shaped mountains, and repeated selection shaped orbits and asteroids. Whatever has survived many rounds of that pressure is not a random pattern. There is a manifold to learn and a gradient to follow. Neural networks are very good at following gradients, provided someone specifies the objective correctly.

The conjecture has a stated limit. Man-made or abstract problems may have no such structure. His example is factoring large numbers: if the number space is uniform, there is nothing to learn, and you are back to brute force, or perhaps a quantum computer. That is why he is working with colleagues on a possible new complexity class, a set of natural systems that classical machines can model efficiently. It also explains why he treats P equals NP as a physics question. He holds that information, not energy or matter, is the universe's most fundamental unit.

The boundary keeps moving in the wrong direction for sceptics. Fluid dynamics is the textbook case of a problem that eats compute. Hassabis wrote physics engines early in his games career and knows how painful they are to build. Yet Veo renders clear liquids squeezed through hydraulic presses, along with specular lighting and materials, having learned from video alone. WeatherNext beats traditional forecasts that take days on supercomputers, including hurricane paths. AlphaFold 3 models interactions between proteins, RNA and DNA. AlphaGenome maps genetic changes to function. His reading is that the networks are extracting a lower-dimensional manifold, and that this may be true of most of reality.

He concedes where it gets hard. Chaotic systems, where initial conditions dominate and the end state decorrelates, may resist modelling. Cellular automata sit right on the edge. He presents these as open questions, not settled ones.

Veo also reopens an old argument about understanding. Hassabis says that to the extent a system predicts the next frames coherently, that is a form of understanding. It is not philosophical understanding. It is intuitive physics, closer to a child's than a PhD student's. Five or 10 years ago he would have said that needs action in the world, embodiment or at least simulated action, as neuroscience theories of action and perception hold. Passive observation turned out to be enough. That is a world model, and he says a true AGI needs one. The next step is making such models interactive, and games are where his early career and current work meet. He imagines open-world games that generate content around the player's choices, five to 10 years out.

The obvious objection is that a model of the world only captures what the data already contains. Modelling is not discovering. Hassabis agrees, and splits the problem in two. The model captures the dynamics. A search process on top takes you somewhere new. Monte Carlo Tree Search found move 37 in Go. AlphaEvolve pairs LLM proposals with evolutionary search and has produced faster matrix multiplication. Evolution adds something extra, because mutation and combination can yield emergent capabilities. Older evolutionary computing never managed that, and he suspects foundation models may fix it. Even so, search needs an objective function to climb, or the space is too big.

Which is where it gets harder. Everything these systems do well is hill-climbing on a curve already underway. Could AlphaEvolve have invented the Transformer in 2017? He says it is not clear. Nobody has shown a big leap unequivocally. Give a system a vague instruction, such as inventing a game as good as Go, and it has no way to narrow that into something tractable.

The missing piece is taste. Hassabis calls choosing the right question the hardest part of science and says a good conjecture is harder to produce than its proof. Systems may soon solve Math Olympiad-level problems, as AlphaProof did with a silver medal, and perhaps a Millennium Prize problem. But a conjecture that Terence Tao would find deep must be falsifiable, reachable with current technology, and split the hypothesis space in two, so that either outcome teaches you something. A naive search on a model will not do that. He says nobody knows the mechanism.

This shapes his AGI test. He puts 50% on AGI by 2030 and defines it as matching the brain's cognitive functions consistently, without the jaggedness of today's systems. The test is brute force: tens of thousands of cognitive tasks, plus a few hundred top experts given a month or two to find an obvious flaw. Then come the lighthouse moments. Give a model only knowledge up to 1900 and see whether it derives special and general relativity. Ask it to invent a game as deep and elegant as Go. Passing several of these would count.

Does the path run through scaling or through new ideas? He says 50/50, and DeepMind spends about half its resources on each. Pre-training, post-training and inference-time thinking are all still scaling. He says he likes it when the terrain gets harder, because it favours research over fast-following, and he claims 80-90% of the breakthroughs behind modern AI came from Google Brain, Google Research or DeepMind. He is cool on recursive self-improvement. It is a hard takeoff scenario, and he is not sure it is desirable. Product work follows the same logic. You design for where the model will be in a year, not where it is today.

Compute demand rises regardless. Training often needs co-located data centres, billions of users need inference, and thinking systems get smarter with more test-time compute. Training may become the smaller share.

What the learned-model approach buys is science. Hassabis has wanted a virtual cell for about 25 years, since conversations with Paul Nurse. He would start with yeast, a whole organism in a single cell. AlphaFold gave static structure. AlphaFold 3 starts on interactions. The next step is a pathway like TOR, then the cell. The cutoff would be the protein level, not the atomic one, with hierarchical simulators handling different timescales. The payoff is perhaps 100X faster experiments, with the search done in silico and validation in the wet lab.

Energy is the same bet at civilisational scale. He picks fusion and solar. If energy becomes cheap and clean, desalination, rocket fuel from sea water and asteroid mining follow, and the world stops being zero-sum. He calls it a "radical abundance" era and says the harder question is whether it is shared fairly.

He is less sanguine about the transition. He expects roughly 10 times the impact of the Industrial Revolution at 10 times the speed: 10 years instead of 100. Great programmers may become 10X more productive over the next 5-10 years, since specifying architecture and checking output stays valuable. But society will need new institutions and perhaps universal basic provision, and he wants economists and philosophers on it now. On catastrophic risk he refuses a number, because a p(doom) implies precision that does not exist. His answer is that the risk is "definitely non-zero" and "probably non-negligible", that the stakes run both ways, and that risk research needs 10 times more effort. He worries about bad actors and about control of agentic systems. He admits he has not heard a clear way to keep powerful systems open to good actors while closed to bad ones.

The unresolved part is the gap between the two halves of his argument. The conjecture is supported by cases where learning worked, and he concedes that chaotic systems and number theory may sit outside it. Meanwhile the capability he says AGI needs most, taste, has no known mechanism, yet the timeline is 50% by 2030. If the leap needs a breakthrough nobody has described, that date is a bet on a research bench, not a forecast.