Nature Has Structure: Demis Hassabis on Learnable Reality and AGI

As a teenager in the British games industry, Demis Hassabis wrote physics engines and graphics engines, work he later called painstakingly hard. Getting a program to simulate a clear liquid squeezed through a hydraulic press meant hand-coding surface tension, refraction, and constitutive equations. Now he runs Google DeepMind, and Veo 3, trained on YouTube video, can generate those same sequences. He told Lex Fridman he loves watching them, because only he remembers how miserable those nights used to be.

That contrast is what he smuggled into his Nobel Prize lecture. Nobel talks, he felt, have a tradition of being a little provocative, so he offered a conjecture: any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm. AlphaGo and AlphaFold are the exhibits. Go has on the order of 10 to the 170 possible positions; proteins have perhaps 10 to the 300 possible structures. Both numbers dwarf the atoms in the universe. Brute force would not finish before the heat death of everything.

What made the problems tractable was a model of the environment that guided search. Proteins fold in milliseconds in the body, which means physics already solves a problem DeepMind later solved computationally. Hassabis's explanation is that natural systems have structure because they were shaped by evolutionary processes. He sometimes calls it survival of the stablest: not only Darwinian life, but mountains weathered over millennia, asteroid shapes hammered by collisions, planetary orbits filtered by gravitational perturbations. Whatever has persisted has been acted on many times, so it is not a random pattern. There is a manifold a neural network can follow.

The converse of the conjecture is just as important. If a problem has no exploitable texture, the AlphaGo trick fails. Factorizing large numbers may be such a case: if number space is uniform, there is no model to learn, only brute force, and you might actually need a quantum computer. Most of the natural systems we care about, Hassabis argues, are not like that. Lex Fridman asked whether this deserves a new complexity class, something like learnable natural systems. Hassabis said that in his spare time, with a few colleagues, he is already poking at whether neural-network methods define a class of problems mapped onto structured physical systems.

Then he went further. He thinks information is the most fundamental unit of the universe, more fundamental than energy and matter, which can be converted into one another but can all be treated as forms of information. If the universe is an informational system, P versus NP stops being a computer-science puzzle and becomes a physics question. He ranks it among the most basic questions we have. The paradigm of model-then-search, he says, has already shown that classical Turing machines go further than many people guessed a decade or two ago: they can fold proteins and play Go at superhuman level, jobs once imagined to require quantum hardware.

He would not overclaim. Cellular automata and most emergent systems probably sit inside the modelable set; you can often just simulate them forward. Chaotic systems, where initial conditions dominate and the end state is uncorrelated, may be the hard boundary. Fluid dynamics and the Navier-Stokes equations have traditionally eaten enormous classical compute, yet Veo can already model liquids, materials, and specular lighting surprisingly well. Hassabis reads that as evidence about how the universe is structured, and as one of the reasons he wants AGI: to help scientists answer questions like P versus NP.

Fridman is less impressed by Veo's memes and ultrarealistic humans than by its physics. The cynical take on diffusion models is that they understand nothing. Hassabis's reply is operational: to the extent the system can predict the next frames coherently, that is a form of understanding. It is not a deep philosophical grasp. The model has captured enough of the dynamics to generate roughly eight seconds of consistent video that, at a glance, is hard to fault. He compares the competence to a child's intuitive physics, not a PhD student's unpacking of the equations. A child knows balls fall, water flows, and glass breaks. Veo 3 seems to live at that layer.

Five or ten years ago, Hassabis would have agreed with a dominant view in neuroscience: action in perception. To really understand the world, you had to act in it. Embodied intelligence, robotics, or at least simulated action looked like a prerequisite for common sense, the knowledge that a glass pushed off a table will shatter and spill. Veo 3 makes that look less necessary. It only watched video, and it still extracted a working model of materials and motion. If passive observation is enough, that again hints at a lower-dimensional manifold in reality itself.

The next step, especially given his games background, is to make those videos interactive: step into the generated world, walk around, push things, change it. That, he says, starts to look like a world model, a model of how the world works, its mechanics, its physics, and the things in it. And that is what a true AGI system would need. When someone on X asked "playable world models wen?" Hassabis quote-tweeted, "Now wouldn't that be something."

Games were Hassabis's first love and his first professional AI work as a teenager. Almost every game he made was open-world: a simulation, AI characters, a player who interacts, a world that adapts. On Theme Park, each player's experience was unique because the player was co-creating with the simulation. DeepMind set parameters and initial conditions; the story was written together. Black & White, which he worked on in its early stages, had an early reinforcement-learning system. You raised a mythical creature, and it treated villagers the way you treated it: mean if you were mean, protective if you were kind.

The 1990s constraint was content. Wherever the player walked, you needed something good waiting. Classical tricks such as cellular automata produced some emergence, but they were fragile and limited. Hassabis thinks the inflection is close. In five to ten years, AI systems should be able to generate around a player's imagination, rewrite the narrative on the fly, and keep the story dramatic no matter which path is taken: the ultimate choose-your-own-adventure. Fridman drew a distinction Hassabis accepted. The Stanley Parable makes the illusion of choice its subject. The Elder Scrolls game Daggerfall used random dungeon generation to feel infinite, but it was still a bag of prewritten branches. What Hassabis wants is a world where you can potentially do anything, and he thinks the only way to get there is generative systems. You cannot ship infinite AAA assets.

His favorite games of all time are Civilization I and II. He avoids the later ones because they would swallow a sabbatical. Asked whether he and Elon Musk, both serious gamers, share a deeper link, he pointed to the late-1980s home-computer era in the UK: a Spectrum, then a Commodore Amiga 500, his favorite machine, where programming games was the fun way to learn. Games fused art with the most cutting-edge programming of the decade. Graphics, physics engines, AI, hardware, even GPUs were pushed first by gaming. If he ever gets spare time, vibe coding a game is the itch; after AGI has been safely delivered, that game and a physics theory would be his two post-AGI projects, and in his mind they are related, because the game would be an open-world simulation of reality itself.

AlphaEvolve, a Google DeepMind system that still gets less attention than Hassabis thinks it deserves, uses a large language model to propose candidate solutions and evolutionary computing to hunt novel regions of the search space. He files it under hybrid systems: a foundation model plus another computational technique. Monte Carlo tree search is another such add-on; it is what found Move 37, the never-before-seen Go strategy. The model captures what is already known. Search is how you go beyond it, which is the beginning of a story about creativity, scientific discovery, and medicine.

Classical evolutionary algorithms were studied hard in the 1990s and early 2000s and hit a ceiling: they could not evolve truly new properties. What you put in, you got back as subsets and recombinations. Nature did better. Bacteria became us. Hassabis thinks bolting evolution onto foundation models might clear that old limit. He also draws a line. AlphaEvolve is excellent at incremental hill-climbing toward a crisp objective, such as faster matrix multiplication. Give it a vague high-level instruction—make a better version of yourself, invent a game as deep and beautiful as Go—and it currently does not know how to narrow the prompt. Whether such a system could have invented the Transformer architecture in 2017 is, in his view, unlikely today.

That is why the path to AGI forks. Is incremental climbing enough, or are one or two architecture-scale breakthroughs still required? He calls it fifty-fifty. DeepMind bets both ways, putting roughly half its resources on blue-sky research and half on scaling current capabilities to the limit. The piece he finds hardest to mimic is research taste: sniffing out the right direction, the right experiment, the right question. Coming up with a conjecture worthy of Terence Tao is harder than solving one. AlphaProof took a silver medal at the Math Olympiad, and systems may eventually crack a Millennium Prize problem, but inventing a question that splits the hypothesis space in two, is falsifiable, and fits today's tools is a different kind of creativity. Naive search on a model will not be enough.

Hassabis's method is to name an almost unrealistic goal, then break it into interim steps that are achievable and useful on their own. The virtual cell is the one he has been dissecting for about twenty-five years. He has talked about it since the 1990s with Paul Nurse, a mentor who founded the Crick Institute and won a Nobel Prize in 2001. They revisit the question every five years: what would you need to model a cell's full internals so that most of the search happens in silico and only validation stays in the wet lab? The dream is roughly a hundredfold speedup of experiments. He would start with yeast, a single-celled organism that is itself a complete creature and is extremely well studied, which is also what Nurse worked on.

AlphaFold solved the static three-dimensional picture of a protein. Biology's interesting events are dynamic. AlphaFold 3 is the first step toward interactions: proteins with proteins, RNA, and DNA. Next might be a whole pathway, such as the TOR pathway involved in cancer, and only later a whole cell. Two technical problems sit in the way. Events inside a cell run on wildly different clocks, from millisecond folding to hour-scale regulation, so you probably need several simulated systems interacting across timescales, or a hierarchy you can jump up and down. And every natural model needs a cutoff. For a cell, Hassabis hopes the protein level is enough, without dropping to atoms. AlphaFold then becomes the foundation, and higher-level simulations use those structures as building blocks until emergent behavior appears.

Asked whether AI could model the origin of life, he recommended Nick Lane's *The Ten Great Inventions of Evolution*, which also speaks to the Fermi great-filter question. Hassabis thinks the filter is more likely behind us than ahead: the jump from single-cell to multicell took on the order of a billion years on Earth, and bacteria were content for a very long time before capturing mitochondria. After a virtual cell, a later project might be to start from a primordial chemical soup near hydrothermal vents and ask whether something cell-like can emerge. What still bothers him is definitional. We can inspect living things down to nearly the atomic level and still lack a clean line between living and non-living. Fridman suggested there may be no line at all, only a continuum from physics through chemistry to biology. Hassabis did not argue. That mystery is one of the reasons he has spent thirty years on AI: it may be the ultimate tool for answering it. Along the way, DeepMind's WeatherNext system has already shown that even weather, a near-chaotic fluid, can be modeled by neural nets better and faster than traditional supercomputer fluid dynamics.

Hassabis's public estimate is a 50 percent chance of AGI within the next five years, by 2030. His bar is high: match the cognitive functions of a brain that already approximates a general Turing machine and that built modern civilization. Today's systems are jagged, brilliant in some places and strangely broken in others, and unstable. AGI would need consistency across the human cognitive spectrum. One blunt test is tens of thousands of cognitive tasks humans can do. Another is human red-teaming: give the system to a few hundred of the world's top experts, the Terence Taos of each field, for a month or two, and see if they find an obvious hole.

He also wants lighthouse moments, Move 37-class leaps rather than a leaderboard score. One thought experiment is a knowledge cutoff at 1900: feed the system everything written up to then and see whether it proposes special and general relativity. Another is whether it can invent a game as deep, aesthetically beautiful, and elegant as Go, and not in one domain only. Several such feats at once would be the sign. Fridman worried that experts will instantly take miracles for granted and hunt for the remaining error. Hassabis granted the point. His chess analogy is that if Garry Kasparov or Magnus Carlsen plays a brilliant move, Hassabis might never have found it, but a good explainer can show why it works. Being able to explain complex thoughts simply is itself part of intelligence. If the first reaction is that the move is a blunder, that will happen too; intuition and, in games, the win-loss result will catch up. He has been in that feedback loop since he was four.

Fridman described the same tension in generated code: the first instinct is that something subtle is wrong, then the harder question, am I the one who is wrong, is this a bug or a feature? As systems get stronger, that self-doubt gets more expensive. Hassabis expects humans to inspect the output with other AI tools, monitors distinct from the generators. On recursive self-improvement, AlphaEvolve-style loops could in principle be wired end-to-end, but he is not sure that is even desirable, because it looks like a hard takeoff. Current systems still fail at underspecified goals. They can grind for days on matrix multiplication. They cannot take "invent a game as good as Go" or "make a better version of yourself" and turn it into a tractable plan.

On scaling laws, Hassabis sees three curves running at once: pre-training, post-training, and inference-time compute. All three still have a lot of room. Whether a new breakthrough is needed remains fifty-fifty. He actually likes it when the terrain gets harder, because the work tilts from pure engineering toward research, and research plus engineering is DeepMind's sweet spot. Inventing is harder than fast-following. If an AlphaGo- or Transformer-scale leap is required, he would back DeepMind to make it; he estimates that 80 to 90 percent of the breakthroughs underwriting modern AI came from Google Brain, Google Research, and DeepMind. Data, he is not very worried about. The live question is whether you have enough real-world data to train simulators that then generate synthetic data from the right distribution. He thinks we are at that step now.

Compute will keep scaling for several reasons. Training the largest models needs co-located clusters, so even bandwidth between data centers matters. Products used by billions need a ton of inference. And thinking systems get smarter the longer you let them run at test time. Training may even become the smaller slice of the total. DeepMind is working on inference-only chips and on AI that helps with data-center cooling, grid optimization, and eventually plasma containment in fusion reactors, work already done with Commonwealth Fusion, plus reactor design. On materials, new solar compounds, room-temperature superconductors, and optimal batteries sit on his dream-breakthrough list. A solution to any one of them would be revolutionary for climate and energy, and he thinks AI systems that can materially help are perhaps five years away.

Asked to bet on humanity's main energy sources in twenty to forty years, he picks fusion and solar. Solar is the fusion reactor in the sky; the bottlenecks are batteries and transmission, perhaps eventually space-based solar and Dyson-sphere-scale collection. Fusion looks doable if the reactor design is right and plasma can be controlled fast enough. Cheap, clean energy would collapse other scarcities. Desalination already works; it is just expensive, which is why only relatively wealthy coastal countries such as Singapore and Israel use it. Split seawater into hydrogen and oxygen and you have rocket fuel. Pair that with reusable rockets and spaceflight starts to look like a bus service, with asteroid mining behind it. For the first time, Hassabis says, we would not be resource-constrained. The era would not be zero-sum: this land is mine, so it is not yours; tigers keep the forest, so the village goes hungry. Human foibles would remain. One big vector of conflict, scarcity, would be gone. The next question is fair sharing. A Type I Kardashev civilization a hundred years from now would not greatly surprise him, provided we do not destroy ourselves on the way.

A year earlier, Fridman said bluntly, Google was losing on the LLM product side with Gemini 1.5; with Gemini 2.5 it is winning. Hassabis credited people and culture first: teams led by Koray, Jeff Dean, and Oriol, plus the merger of Google Brain with the original DeepMind, gathering the best people and ideas. He called the method relentless progress paired with relentless shipping. Big companies grow layers, meetings, and managers; his counter is to keep Google DeepMind behaving like a startup, large in scale, fast in decision-making. Few places can do world-class research on Monday and plug it into products used by billions on Tuesday. The unglamorous work is constantly cutting bureaucracy so that research culture and shipping culture can both grow. Many people still meet AI first as AI mode in Search. For them, he wants the encounter to be seamless, often under the hood in Maps or Search, not a Twitter discourse.

Gemini's product and engineering teams told Fridman something Hassabis had not expected to hear: he is not only a scientist, he is a strong product person. He traces that to designing games for millions of players in the 1990s, cutting-edge technology in a consumer product. The skill is the same one science needs: the ability to imagine viscerally what using the thing feels like, and taste. You cannot design today's AI products for today's capabilities. You have to design for a year from now, which means being a deeply technical product person who can judge whether the research track will intercept the dream in six or twelve months. Today's text-box chat is likely a transitional form. Once systems are fully multimodal, these interfaces will look archaic, closer to a collaborative *Minority Report* vibe than a prompt field. Typing is low bandwidth even if you type fast. Smart glasses, earbuds, and eventually neural devices could raise input and output bandwidth by a hundred times. He also expects AI-generated interfaces, personalized to your aesthetic and the way your brain works: every knob for power users, blank space for everyone else.

A full Gemini training run plus productization takes roughly six months. During that window the team collects architectural, data, and other ideas, tests which ones will matter, packages them, and launches a giant hero training run. After pre-training comes a wide experimental post-training phase that can still yield large gains. Version numbers usually name the base model; the 2.5-style increments are often patches or post-training tricks on the same architecture. Pro, Flash, and Flash-Lite cover a Pareto frontier of performance versus cost and latency, with smaller models distilled from larger ones. Benchmarks are necessary and dangerous. The job is multi-objective: a no-regret improvement that lifts coding without hurting translation, writing, or scientific reasoning. The monitored suite keeps growing, and live product signals matter as much as leaderboards. Softer still is persona: verbose or succinct, funny or serious. Hassabis did five-factor personality research in his PhD and did not expect to need it for products. Different people want different flavors.

Asked for the probability that Google DeepMind "wins," he rejected the frame. Given how consequential the technology is, winning is the wrong way to look at it. The people at the leading edge have a responsibility to steward it safely into the world for humanity's benefit, and he hopes the international community will rally around that as AGI gets closer. He tries to stay on good terms with other lab leaders, because those channels will matter for safety. Meta's talent shopping he treats as rational for a lab that is currently behind, but he thinks people who truly believe in AGI are mostly there to stand on the frontier and influence how it lands. Money is not the deepest issue. In 2010 he went unpaid for a couple of years because DeepMind could not raise enough; intern offers now rival that entire seed round. Post-AGI, if energy is solved, what money even means will be a larger question than today's salaries.

Programming and mathematics were among the first domains to fall, which is counterintuitive if you still think of them as the hard skills. Hassabis's reason is crisp: the outputs can be verified, so you can generate synthetic training data at scale. Right and wrong have an objective test, unlike novels or strategy. Over the next five to ten years, people who fuse with these tools will become almost superhumanly productive. Great programmers may be ten times more productive than they are today, because they are best at pushing the tools to the limit, specifying architecture, and judging whether generated code is any good. Coding that gets easier also opens the door to more creatives. Value will redistribute. Frontend work is more likely to be generated outright. High-performance backends, game engines, and system architecture, where the job is making tradeoffs under constraints, will concentrate human value.

The shock, he did not soften. The internet, mobile, and the industrial revolution all remade jobs and grew occupations no one had imagined. This time the adaptation window is crushed. He puts the effect at about ten times the industrial revolution's impact and ten times the speed: a hundred years of change compressed into ten, a hundredfold combined hit. He wants top economists and philosophers working on it now, including mechanisms such as universal basic provision, sharing a floor of the productivity gains as services, while people who want more still cultivate rare skills and make themselves unique.

Fridman pushed into political philosophy: government is itself a technology, and if it cannot iterate, rapid change will tear it, while politicians are skilled at converting division into votes. Hassabis agreed that new governance structures and institutions will be needed. He insisted on sequence. First raise the total of resources, moving from zero-sum to positive-sum. Then argue about distribution. Without the first, the second has nothing to allocate.

Benjamin Labatut's novel *The MANIAC* shadows Hassabis and centers on John von Neumann, a haunting study of genius, madness, and the double edge of discovery. Von Neumann helped build the atomic bomb, pioneered the architecture of modern computers, and foresaw learning machines. Hassabis notes a detail he loves: von Neumann spoke of systems that are grown, as opposed to programmed. He would not be shocked by 2025; Hassabis thinks much of this was already visible to him in the 1950s, and he would have delighted in AlphaGo, being a games and game-theory mind. The difference from nuclear technology is breadth. AI is a general-purpose tool that might cure disease and end scarcity, with risks attached. Hassabis guesses von Neumann saw both faces.

On the book's claim that the mad dreams of reason are not enough to guide humanity, Hassabis agrees. Call it a spiritual dimension or a humanist one; it need not be religious. The spark that makes us human, perhaps bound up with consciousness, has to sit at the center of the endeavor. He sides with Richard Feynman: science and art are companions. Understanding why a flower's colors evolved makes it more beautiful, not less. Leonardo da Vinci, in Hassabis's reading, did not split science, art, and religion. One of his favorite philosophers is Spinoza, who treated understanding the universe and our place in it as a way of understanding religion. Technologists, he warns, often know only the technology. That is one reason he supports AI summits and government involvement. The chatbot era has a democratic upside: ordinary people can touch frontier AI and are forced into the debate.

Would states run a Manhattan Project and treat the technology as a weapon? He hopes not. That would be dangerous and a misuse. He would rather see something more like CERN: research-focused, the world's best minds finishing the last steps carefully before deployment. Current geopolitics make cooperation hard, but things change. At minimum, researchers should stay in contact. Science, Fridman added, has always been a connector; multicultural, multidisciplinary teams working on shared questions are the hopeful pattern, along with two-way education and immigration so that people actually mix. Hassabis has no p(doom) number, because a number would imply precision that is not there. What he can say is that the probability is definitely non-zero and probably non-negligible, which is already sobering. Under huge uncertainty and huge stakes both ways, the only sensible path is cautious optimism: take the benefits, and use the scientific method to define and address the risks more precisely. He thinks that effort needs to be at least ten times larger as we near AGI.

The risks sit on different clocks and are both important. One is ordinary bad actors, individuals or rogue states, repurposing a general-purpose technology. He is a committed open-science advocate; AlphaFold and the other science projects were released for the research community. How you keep powerful systems from bad actors while letting good actors build on them is a problem he has not heard a clear solution to. The other clock is autonomous systems approaching AGI: keeping guardrails on, keeping them under human control. Fridman worries more about people, including how to reduce the number of bad actors, not only how to keep tools from them. Hassabis noted that AI itself might provide early warning on bio or nuclear threats, if the warning system is reliable, an interlocking problem. Some basic standards, at least between China and the United States, will likely be required.

Fridman brought the conversation back to Lee Sedol's Move 78 against AlphaGo, often called the last flash of purely human genius, a move that briefly broke the machine's evaluation. Hassabis remembers it as mutual inspiration. The DeepMind team was moved by Sedol's brilliance and nobility; Sedol, perhaps stirred by AlphaGo's play, conjured the move. The AlphaGo documentary captures it well. He expects versions of that exchange to continue. For the foreseeable future, humans will still supply the ingenuity of asking the right question, then use the tools to crack it.

What makes humans special, if anything, is a question Hassabis has been chasing since childhood. He did a neuroscience PhD on imagination and memory, focused on the hippocampus, because philosophy and even experiments on real brains were never going to be enough. The best test is to build an intelligent artifact and compare it with the human mind. He suspects there is a difference, including possibly between carbon and silicon as they process information. One definition of consciousness he likes is that it is how information feels when we process it, not a very scientific explanation, but an intuitively interesting one. He and Roger Penrose have talked often and cordially disagree. Penrose has worked with neuroscientists looking for quantum-mechanical behavior in the brain and, to Hassabis's knowledge, has not found convincing evidence. Hassabis's bet is that the brain is mostly classical computation, which would mean the relevant phenomena can in principle be modeled or mimicked on a classical computer. Qualia, the feeling of what it is like, might still be bound to the substrate.

A neural interface, perhaps needed just to keep up with AI, might let us feel what it is like to compute on silicon and settle the question from the inside. He once debated the late Daniel Dennett about why we attribute consciousness to one another: similar behavior, and the same substrate. With silicon the second clause fails. An AI may look conscious and even claim to be; we would not know how it felt, and it would not know how we felt, until perhaps a superintelligence builds a bridge. Fridman called that a test of radical empathy humanity has never faced. Hassabis noted we already stretch toward dolphins, orcas, elephants, dogs, and monkeys, animals that show aspects of consciousness even if they are not high-IQ. DeepMind's DolphinGemma, trained on dolphin and whale sounds, might one day become a translator.

What gives him hope is two human traits. First, almost limitless ingenuity: there is nothing better than watching someone at the top of sport, science, or art enter flow. Brains are general systems with a very high ceiling. Second, extreme adaptability. Minds evolved to hunt buffalo on the tundra now fly, record podcasts, and inhabit virtual simulations, which is already absurd. Society has adopted chatbots faster than he expected, with a shrug of, I talk to a chatbot, that is fine. Fridman joked that this podcast will be replaced by AI and that he is waiting. Hassabis told him not to the level Lex can do it. That compliment, Fridman laughed, is what humans do for each other. Underneath the joke is Hassabis's actual method: take an unrealistic goal, cut it into deliverable steps, and ship. The virtual cell has been in pieces for twenty-five years. AGI is a sequence of six-month hero runs. Energy is materials, batteries, transmission, and plasma control. He is not so much predicting the future as putting it on an engineering schedule.