Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games

This is Demis Hassabis's second appearance on the Lex Fridman Podcast. Fridman introduces him as the leader of Google DeepMind and now a Nobel Prize winner. The conversation opens with a conjecture Hassabis put forward in his Nobel Prize lecture about what classical learning algorithms can capture about the natural world.

“any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm.”

Fridman asks which domains the claim covers, listing biology, chemistry, physics, maybe cosmology and neuroscience. Hassabis first explains why he framed it the way he did. He says he sees it as a tradition of Nobel lectures to be "a little bit provocative", and he wanted to follow that tradition. So he is presenting the statement as a deliberately bold conjecture, not an established result.

His evidence comes from DeepMind's own "Alpha X" projects, chiefly AlphaGo and AlphaFold. He describes both as building models of combinatorial, high-dimensional spaces where brute force fails. If you enumerated every possible Go move or every possible protein shape, "there wouldn’t be enough time in the time of the universe." In both cases the team built a model of the environment, and that model guided the search "in a smart way", which made the problem tractable.

Protein folding is his central example of a natural system. Proteins fold in milliseconds in the body, so physics somehow solves a problem that DeepMind has now also solved computationally. His explanation is that natural systems have structure because evolutionary processes shaped them. He states the step to learnability as a condition: "if that’s true, then you can maybe learn what that structure is."

Fridman offers a blunt paraphrase: anything that can be evolved can be efficiently modeled. Hassabis agrees that there is some truth to it and widens the idea beyond biology. He calls it "survival of the stablest." Mountains are shaped by weathering over thousands of years. On a cosmological scale, the orbits of planets and the shapes of asteroids are also the products of processes that have acted on them many, many times.

From this he infers that such systems should contain a pattern that can be "reverse" learned, which he describes as a kind of manifold. That manifold would help a search reach the right solution or shape, and would allow efficient prediction because the pattern is not random. Later he extends the selection idea even to which chemical elements are more stable. Everything around us, he says, is subject to "some kind of selection process pressure."

Hassabis also sets out a limit. The argument may not hold for man-made or abstract problems, and his example is factorizing large numbers. If the number space is uniform and has no patterns, there is no model to learn that could guide the search, so brute force is required. For that kind of problem, he suggests, "you maybe need a quantum computer." He adds a hedge: there might be patterns in the number space after all. He holds that most natural things of interest are not like this, because their structure "evolved for a reason and survived over time." Even so, his conclusion is cautious: such structure is "potentially learnable by a neural network."

Fridman restates the idea as nature running a search process that happens to produce systems that can be efficiently modeled. Hassabis agrees, and adds that such systems can be efficiently "rediscovered or recovered because nature’s not random."

Fridman, pointing to Hassabis's interest in theoretical computer science, half-jokingly proposes a new entry for the "complexity zoo": a class of learnable natural systems, which he labels "LNS." These would be natural systems that classical machines can model efficiently.

Hassabis takes the proposal seriously. He says he has long been fascinated by the P equals NP question and by what classical systems can model, which he equates with non-quantum systems and, in effect, Turing machines. He says this is "exactly what I’m working on actually in my few moments of spare time with a few colleagues": whether there should be a new class of problems solvable by this kind of neural network process, mapped onto natural systems that exist in physics and have structure. He presents it as ongoing exploratory work, and describes it as "a very interesting new way of thinking about it" without claiming any result.

He links this to his broader view of physics. He holds that information is the most fundamental unit of the universe, more fundamental than energy and matter. He thinks all three can be converted into each other, and he sees the universe as "a kind of informational system." This is stated as personal belief, not demonstrated physics.

Fridman draws out a consequence: if the universe is informational, then P versus NP becomes a physics question. Hassabis agrees, calling it "one of the most fundamental questions" under that view of physics. He expects that the answer will be "very enlightening."

Fridman acknowledges that parts of the discussion may sound "kind of crazy right now." He compares it to Christian Anfinsen's Nobel Prize speech, which he says made a controversial claim that sounded crazy, before Hassabis went on to win a Nobel Prize with John Jumper for solving the problem. He then asks more precisely whether spending polynomial-time or constant-time compute up front to build a huge model could make some extremely hard problems solvable in a formal, theoretical computer science sense.

Hassabis answers that "a huge class of problems" could be framed the way AlphaGo and AlphaFold were. You model the dynamics, properties and environment of the system, and that model makes the search for a solution, or the prediction of the next step, efficient: "Basically polynomial times." That keeps the problem tractable on a classical system, which a neural network running on ordinary computers is. He calls the question of "how far can that paradigm go?" one of the most interesting there is.

His main evidence is how expectations have shifted historically. He says DeepMind and the wider AI community have shown that classical systems can go much further than previously thought. They can model protein structures and play Go better than world champions. He recalls that 10 or 20 years ago many people might have thought those achievements were decades away, or that protein folding might need quantum systems. His conclusion is that we "haven’t really even sort of scratched the surface yet" of what classical systems can do.

He presents AGI, built on neural networks running on classical computers, as "the ultimate expression" of this paradigm. On this view, the limits of what such a system can do bear directly on the P equals NP question.

Fridman asks what might fall outside the conjecture, and suggests emergent phenomena such as cellular automata, where very simple rules produce complexity. Hassabis places those systems "right on the boundary." He thinks most emergent systems, cellular automata included, are probably modelable classically, because a forward simulation would "probably be efficient enough."

He names chaotic systems as the harder case. In those systems initial conditions matter a great deal and lead to "some uncorrelated end state," and they "could be difficult to model." He openly calls these "kind of the open questions," which is the clearest admission in this segment of where his conjecture could fail.

Fridman adds his own examples of surprising progress. He says it is surprising that AlphaFold 3 makes any headway on molecular interactions, and that AlphaGenome can map genetic code to function. In both cases a huge number of combinatorial options turns out to have "the kernel that is efficiently model-able." Hassabis explains this in terms of structure: there is some landscape, such as an energy landscape, with "some gradient you can follow." Neural networks are very good at following gradients. If such a gradient exists and the objective function is specified correctly, he says, the system does not need to handle all of the apparent complexity. He suggests we may have "naively" thought about these problems for decades by imagining full enumeration, which makes them look totally intractable.

Fluid dynamics is the sharpest test case. Hassabis notes that the Navier-Stokes equations have traditionally been seen as very difficult, intractable problems for classical systems. They require enormous amounts of compute and underlie weather prediction. Yet Veo, DeepMind's video generation model, can model liquids "surprisingly well," along with materials and specular lighting. He points to user-made videos of clear liquids being squeezed out of hydraulic presses.

He draws on his own background. He wrote physics and graphics engines early in his gaming career and knows how "painstakingly hard" it is to program those effects by hand. Yet these models seem to be reverse engineering the behaviour "from just watching YouTube videos." His explanation is phrased as inference, not established fact: presumably the model is extracting underlying structure about how materials behave. He also suggests Veo 3's rendering of physics and lighting may be telling us "something quite fundamental about how the universe is structured," and he labels that view as his own opinion.

So perhaps there is some kind of lower dimensional manifold that can be learned if we actually fully understood what’s going on under the hood. That’s maybe true of most of reality.

Fridman's framing captures the caution that runs through the exchange. Humans struggle to make clean predictions about highly nonlinear dynamical systems, yet "we might be very surprised" by what classical learning systems can do, even with fluids. Hassabis ties this back to his larger aim: part of why he wants to build AGI is to help scientists answer questions like P equals NP.

Limitations and open questions raised in this section: the conjecture is deliberately provocative and still unproven; problems without structure, such as factorizing large numbers if the number space is uniform, may still need brute force or quantum computers; chaotic systems that are sensitive to initial conditions could be hard to model; the work on a new complexity class is a spare-time exploration with colleagues and has no stated results; and the claims that Veo learns underlying physical structure, and that information is more fundamental than matter and energy, are presented as presumption and opinion.

Demis Hassabis frames his conjecture with the scale of the search spaces involved. He cites roughly 10 to the 300 possible protein structures and 10 to the 170 possible Go positions, both far more than the number of atoms in the universe. That raises the question of how anyone could find the right solution or predict the next step. His answer is that it turns out to be possible, and that nature already does it. If we understood how physics achieves this and could model that process, he argues, it should be possible on classical systems. That, he says, is basically what the conjecture is about.

Proteins do fold.

Lex Fridman takes the argument to highly nonlinear dynamical systems, especially anything involving fluid. He mentions a recent conversation with Terence Tao, who works on systems with singularities that break the mathematics. Fridman notes that humans find it hard to make clean predictions about such systems. He suggests we might still be very surprised by what classical learning systems can do, even for fluid.

Hassabis agrees. He notes that fluid dynamics and the Navier-Stokes equations are traditionally seen as very difficult, intractable problems for classical systems. They take enormous amounts of compute, and weather prediction systems depend on them. His counterexample is Veo, Google's video generation model, which he says models liquids "surprisingly well," along with materials and specular lighting. He singles out user-generated videos of clear liquids being squeezed through hydraulic presses. He also draws on his own background: he wrote physics and graphics engines early in his gaming career, and knows how painstakingly hard such programs are to build. Yet these systems are, in his words, reverse engineering this behavior "from just watching YouTube videos."

He offers his interpretation cautiously. Presumably, he says, the model is extracting some underlying structure in how these materials behave. He adds that this may be true of most of reality.

So perhaps there is some kind of lower dimensional manifold that can be learned if we actually fully understood what’s going on under the hood.

Fridman says the aspect of Veo 3 that holds his attention is its physics. He lists features others tend to highlight: the comedic content, the ultra-realistic depiction of humans, and native audio. Still, he calls the physics the most compelling part. He grants that it is "not perfect, but it’s pretty damn good," and asks what the model must understand about the world to do this. He cites the cynical view of diffusion models, that they cannot understand anything, and rejects it. He does not think you can generate that kind of video without understanding. This, he says, brings our own philosophical notion of understanding to the surface.

Hassabis gives a qualified answer. Predicting the next frames in a coherent way is, he says, a form of understanding, but he explicitly limits the claim. He says these systems have modeled enough of the dynamics to generate about eight seconds of consistent video. In that footage, problems are hard to spot "by eye, at least at a glance."

not in the anthropomorphic version of, it’s not some kind of deep philosophical understanding of what’s going on, I don’t think these systems have that

He points to the pace of progress. He asks listeners to imagine the results two or three more years out, given how far the technology has come from early versions one or two years ago. Like Fridman, he says the stand-up comedy videos capture a lot of human dynamics and body language well. What impresses him most, though, is the physics: lighting, materials, and liquids. He takes this as evidence of at least some notion of intuitive physics, a sense of how things are supposed to work.

maybe the way that a human child would understand physics, right, as opposed to a PhD student really being able to unpack all the equations.

Fridman calls intuitive physics the base layer, the thing people sometimes call common sense. He says it surprised many people, himself included: he did not think this level of realism was possible without understanding. He contrasts it with the view that a system can only understand the physical world through embodiment, as a robot interacting with that world. He says Veo 3 feels like it directly challenges that view.

Hassabis says passive observation has changed his view of how intuitive physics can be learned. Five or ten years ago, he thought embodied intelligence, robotics or at least simulated action might be needed to acquire it. His example: knowing that a glass pushed off a table may shatter and spill its liquid. He describes neuroscience theories, including "action in perception," which hold that you need to act in the world to truly perceive it in a deep way. He also mentions theories that this would require embodied intelligence, robotics, or at least simulated action.

But it seems like you can understand it through passive observation, which is pretty surprising to me.

He frames the significance as his own opinion: it hints at something underlying the nature of reality, beyond the videos themselves. He describes possible next stages as hypotheticals. Videos might become interactive, so a person could step into them and move around, which he says would be really mind-blowing given his games background. He says this begins to approach what he would call a world model: a model of how the world works, including its mechanics, its physics, and the things in it. He calls that what a true AGI system would need.

Fridman turns to video games, prompted by an exchange on X. A user named Jimmy Apples tweeted: "Let me play a video game of my Veo 3 videos already. Google cooked so good. Playable world models wen?" Hassabis responded: "Now, wouldn’t that be something?" Fridman asks how hard it is to build game worlds with AI, and what games might look like five to ten years out.

Hassabis calls games his first love. He says doing AI for games was his first professional work, in his teenage years, and included the first major AI systems he built. He says he wants to return to it one day and thinks he will. He also wonders what he would have made in the nineties with today's AI systems, and believes the results could have been "absolutely mind-blowing games."

He says all the games he has made are open world games: simulations with AI characters that adapt to how the player plays. He gives Theme Park as an example. In it, each player's experience was unique because the player co-created the game with the simulation, while the designers set the parameters and initial conditions. He is frank about the limits of that approach. Open world games are very hard to program, because content must be compelling in whichever direction the player goes. Classical techniques such as cellular automata did produce emergent behavior, but he says they were always "a little bit fragile, a little bit limited."