Demis Hassabis: Solve Intelligence, Then Solve Everything Else

Lex Fridman opens with a nested joke: if he were an AI Hassabis had written to practice interviewing people until he was good enough to interview Hassabis, would Hassabis tell him? The DeepMind CEO laughs and says probably not. Telling the system would change its behavior. He calls it a Heisenberg situation — observation disturbing the observed — and Fridman spins the bit into a meta Turing Test, a 2022 replay from a future in which AIs were not yet good enough.

Hassabis is cooler on the Turing Test than most people expect. Turing is one of his heroes, and the 1950 paper was unbelievably influential, but if you actually read it, Hassabis says, Turing never meant a rigorous exam. He did not specify who the judge is, what they know, or how long they get to interrogate the machine — the parameters you would have to lock down to turn a thought experiment into a formal test. A decade earlier, someone had already claimed a pass with a bog-standard program that pretended to be a child, so the judges went easy. That is a different scene from an expert AI researcher who knows how the system was built.

He would rather score systems on thousands, eventually millions, of tasks covering the whole cognitive space, and ask whether they generalize. A sentence that sounds human would not count. Gato, DeepMind's most general agent at the time, pushes prediction as far as it will go: language models predict the next word; Gato predicts the next token, which might be text or a robot joint angle or a gamepad button. "Prediction seems to be fundamental to intelligence," Hassabis says, "and what you're predicting doesn't so much matter." Gato is only a beginning, and they were already scaling it up by orders of magnitude.

Hassabis started chess around four. At eight, winnings from a competition bought his first computer, a ZX Spectrum — the machine that, in Britain, trained a whole generation of programmers because you switched it on and the BASIC prompt was waiting. His parents knew nothing about computers, but the money was his, so he had the vote. He went to bookstores, typed programs from manuals, then started changing them and writing his own games. He could not have explained it then, but he felt the machine was a magical extension of the mind: you set it a task, go to sleep, and come back the next day to a solution. Cars extend the legs. Cranes extend the arms. This machine extended the mind. AI was the obvious next step.

By eleven or twelve he was captaining England junior chess teams and thinking he might go professional. He reached master standard at twelve, second in the world among his age group behind Judit Polgár. Improving at chess meant thinking about thinking — why the brain makes mistakes — and practicing against the physical chess computers of the mid-1980s, including a Kasparov-branded board. David Levy's *The Chess Computer Handbook*, from 1984, explained how those programs were built. His Amiga was not powerful enough for a full chess engine, so his first AI program played Othello, with alpha-beta search and evaluation copied from the chess literature. He still remembers writing it.

At sixteen or seventeen he was designing games professionally. *Theme Park* had AI as a core part of the simulation, sold millions of copies, and reacted to how you played; it was an early sandbox, in the family of *SimCity*, so every run was unique. Later he worked on the early stages of *Black & White*, which he still thinks is the most impressive reinforcement-learning example in a commercial game: you trained a creature, and if you treated it badly it became mean to the villagers; if you were kind, it was kind. Players discovered that a trail of small choices had grown a character. Games, he says, are a unique medium because the player is an agent, not a passive consumer — more visceral than film or a book.

Games have had three lives in Hassabis's career: playing them, designing them, and then, from the first day of DeepMind, using them as a testbed for learning algorithms. The reasons are brutally practical. Humans have played some of these games for hundreds or thousands of years, so the ceiling is known. Win conditions are clear, which makes reward functions easy to write — the thing reinforcement learning fears most is a fuzzy reward. You can run millions of games in parallel in the cloud. And at the end you can put the system across the board from a world champion, an external standard nobody can argue with. DeepMind also hired game engineers Hassabis knew from his previous life, which bootstrapped the lab.

Go had been treated as AI's shame: people said that no amount of Moore's law would crack it, because the search space was too large. When a machine wins in a domain repeatedly declared impossible, Hassabis says, the shock reaches billions of people, and the shock is cognitive — the impossible was only something we believed. Kasparov called chess the *Drosophila* of intelligence, and Hassabis loves the phrase: Turing, Shannon, and the other founders all tried to write chess programs. Turing's computers were too slow, so he ran his own program by hand with pencil and paper over two or three days. Deep Blue beating Kasparov was a huge moment, but Hassabis came away more impressed by Kasparov's mind than by the machine. The same brain could ride a bike, talk, and do politics; Deep Blue, hand-coded from grandmaster knowledge, could not even play tic-tac-toe. What was missing was generality, and learning. That is what AlphaGo, AlphaZero, and MuZero were built to restore.

With his game-design hat on, Hassabis thinks chess stays compelling because of the creative tension between bishop and knight. They move in utterly different ways, yet over centuries they have been balanced to roughly three points each. Swap them and you aim for a different kind of position — closed for the knight, open for the bishop — and that tension is the game. He splits creativity into three levels. Interpolation is averaging examples, an ordinary cat from millions of cat pictures. Extrapolation is AlphaGo's move 37 in game two, a motif no human had found in thousands of years of play. True invention would be inventing Go itself: a game that takes five minutes to learn, lifetimes to master, looks beautiful, and finishes in three or four hours. AI can interpolate and extrapolate. It cannot yet be given those high-level constraints, because it does not really hold combinable abstractions. One day it might auto-balance a game like StarCraft overnight, work that now takes Blizzard years of human testers.

Asked whether we live in a simulation, Hassabis splits the question. He does not buy Nick Bostrom's version — that we are a throwaway ancestor experiment running on some descendant's hardware, one of billions of disposable copies. This instance, he thinks, is critical and possibly unique. What he does believe is that the best way to understand physics is computational: information as the fundamental unit of reality, rather than matter or energy. A physicist says *E = mc²*. Hassabis says the arrangement of molecules in a body *is* information, and in that thinner sense we are in a simulation — except the simulation is reality itself.

That stance puts him at odds with Roger Penrose, whose *The Emperor's New Mind* argued that consciousness needs something quantum, something beyond classical computation. Hassabis has debated Penrose often. DeepMind's career, as he frames it, is to be Turing's champion: push classical machines to their limits and see what they cannot do. He also did a neuroscience PhD partly to look for quantum effects in the brain; mainstream biology, he says, has found no evidence, and most of what we see can be explained classically. Meanwhile AI has spent a decade knocking over theories of the impossible, from Go to protein folding. He would be hesitant to bet against the universal Turing machine. His wager is that what the brain does can be mimicked or approximated on a classical machine, without metaphysics.

The brain, he says, is the biggest miracle in the universe: a few pounds of mush, and also the most complex object we know. One reason he has always wanted to build AI is to compare an intelligent artifact with the human mind and unlock the old mysteries — consciousness, dreaming, creativity, emotion — that philosophy of mind could approach only by introspection. Now there are fMRI machines, single-cell recording, and systems you can actually construct. Fridman puts the loop in one sentence: the universe built the human mind, which is now building computers that help us understand both. Hassabis agrees. We may be the mechanism by which the universe tries to understand itself.

Proteins are the workhorses of biology, tiny bio-nano machines whose every function in the body depends on a 3D shape. They are specified by a one-dimensional amino-acid sequence and, in nature, fold into that shape. Structure determines function, and if you want a drug to bind, you need the surface. The folding problem is whether you can go from the letter string to the 3D structure computationally. Christian Anfinsen floated the idea in his 1972 Nobel lecture and did not say how. Hassabis has heard it compared to Fermat's Last Theorem for biology: a throwaway remark that launched a fifty-year field. Fridman tells him that, Nobel or not, he should do more of that in the margins — set people off for two hundred years and give no details.

Until AlphaFold, the work was experimental and punishing. You crystallize the protein, which some proteins, including membrane proteins, will not do, then use X-ray crystallography or expensive cryo-electron microscopy. The rule of thumb in experimental biology, Hassabis says, is one PhD student, one entire PhD, one protein. AlphaFold 2 predicts the structure in seconds. Over Christmas they ran the whole human proteome, about 20,000 proteins. Structural biologists can now look up a protein the way they would run a search, instead of deciding whether the molecule is worth four years of a young scientist's life.

The deeper puzzle is Levinthal's paradox. An average protein, Hassabis says, maybe 2,000 amino acids long, can fold in on the order of 10^300 conformations, and yet physics solves that search in milliseconds. In a healthy case the mapping is usually unique, and that uniqueness is not obvious — which is why the problem is hard. Misfolding is not a rounding error: one conjecture for Alzheimer's is that amyloid-beta folds the wrong way and tangles in neurons. Proteins also change shape when they interact; they are not static Lego. Nature is doing a search that looks astronomical, and doing it on time.

AlphaFold is, Hassabis says, the most complex and probably the most meaningful system DeepMind had built. Games were never the point; they were a bootstrap for general learning systems that could be aimed at scientific problems, and protein folding was the first big proof. More than 30 component algorithms had to be combined. After 40 years of experiment, the training set was only about 150,000 solved structures — far less than they would normally want. Self-distillation was critical: the model's own most confident predictions went back into the training set to make it larger. They also built in hard constraints from physics and evolutionary biology, such as sensible bond angles, without letting those constraints replace learning. The system still had to learn the physics from examples.

AlphaFold 1 output a distogram, a matrix of pairwise distances, and a separate optimization step turned that into a 3D structure. AlphaFold 2 deleted the middleman and went from sequence straight to coordinates. In machine learning, Hassabis says, the more end-to-end you can make a system, the better, because the model is better at learning constraints than human designers are at specifying them. Let the gradients flow from the output you actually want back to the inputs. Hand-crafted intermediate representations become the ceiling. The recipe for a new problem is to start with plenty of manual structure and a small learned piece, then grow the learned piece until it eats the rest.

That pattern is the AlphaGo lineage. AlphaGo was a learning system trained only for Go, and the brief was world-champion performance by any means. AlphaGo Zero removed human games and started from random self-play. AlphaZero stripped out Go-specific priors, including board symmetry, so it could learn any two-player game from scratch. MuZero stopped being told the rules and had to learn the environment, covering computer games as well as board games. The path runs from imitation learning to fully self-supervised learning — but it cannot be walked backwards. Building AlphaZero or MuZero first would have been nearly impossible, both because they learned so much by building AlphaGo, and because of the psychology. You have to believe in yourself for a long time while people say it cannot be done. Hassabis calls AI an engineering science: unlike the natural sciences, the phenomenon does not sit out in nature. You have to build the artifact, then take it apart.

Fridman asks how much of "solving intelligence" is science, how much engineering, how much data, compute, software, humans, even philosophy. Hassabis says it is a combination, but the ratios change. DeepMind started in 2010. Nobody was talking about AI. After a postdoc at MIT around then, he remembers the mood: we know AI does not work; we tried in the 1990s with logic systems, good old-fashioned AI, Minsky and Patrick Winston. Professors thought he was mad for betting on learning. In industry you could barely raise two cents. The mission statement was, and still is: step one, solve intelligence; step two, use it to solve everything else. Pitch that to a VC in 2010 and you get strange looks. They found a few kooky backers. They stopped mentioning it to professors, who would eye-roll and call it career suicide.

The founding tenets, hashed out with Shane Legg, were a stack that had only just become available at once. Deep learning had just been invented in academia by people like Geoffrey Hinton and Yann LeCun; industry had barely heard of it. Reinforcement learning, they thought, could be scaled. Neuroscience, via fMRI and other tools, was offering hints about architectures and representations at the systems level, not the implementation level. Hassabis's faith in RL was specifically biological: a famous late-1990s result showed monkey midbrain dopamine neurons firing like a temporal-difference prediction error. If that is how the primate brain learns, the path is not a dead end. Compute had been commoditized by the games industry, GPUs sitting there for matrix multiplies. The last piece was theoretical: AIXI, the universal-intelligence proof Legg had worked on with Marcus Hutter, a reinforcement-learning system in the limit, assuming infinite compute and memory, the way a Turing-machine proof does. Hassabis had been waiting for a foundation as load-bearing as Turing and Shannon were for computer science.

Early on, ideas mattered most — deep RL, then transformers, AlphaGo. As you get closer to AGI, he says, engineering and data take more of the weight. Large models, including GPT-3, have shown that scale is a necessary but perhaps not sufficient part of an AGI solution. The other DeepMind invention he flags is organizational: a multidisciplinary lab, originally neuroscience plus machine learning, engineering, mathematics, and games, later philosophers, ethicists, physicists. He wanted a new Bell Labs for this era, an innovation machine made of human minds trying to build learning systems. Fridman thanks him for keeping RL and the phrase "solving intelligence" alive when ambitious work had shrunk to tinkering and the occasional DARPA challenge. Belief, Hassabis agrees, has to run through the whole organization, or the engineers never get to build the thing.

AlphaFold, amazing as it is, is only the beginning. Biology is dynamic. Next come protein-protein interaction, protein-ligand binding, pathways, and eventually a virtual cell — a dream Hassabis has discussed for twenty years with Paul Nurse, the Nobel-winning biologist who runs the Francis Crick Institute, now a few minutes' walk from DeepMind in King's Cross. If you could simulate a cell, you could run most experiments in silico and only validate at the last stage in the wet lab. Going from target to drug candidate takes about ten years; that cycle might shrink by an order of magnitude. After AlphaFold he called Nurse and said the time had come. He also has a larger bet: if mathematics is the perfect description language for physics, AI may be the perfect description language for biology, because biology is messy, emergent, and dynamic. He finds it hard to believe we will ever get Newton's laws for a cell. You have to start from the building blocks, learn the rules, and let the simulation run, like a cellular automaton you cannot analyze from above.

Open source, he says, is decided case by case, for maximum benefit to humanity. DeepMind bought MuJoCo, the physics engine robotics researchers live in, specifically to open-source it after the original professor could no longer support it. AlphaFold's code and structure database went out because the downstream uses were too many to imagine, let alone productize. Within a year more than 500,000 researchers had used it — Hassabis thinks that is roughly every professional biologist on earth. A *Science* special issue put the nuclear pore complex on the cover, the doughnut-shaped gateway that controls traffic in and out of the cell nucleus, unsolved for decades at high resolution. Four groups around the world, more or less simultaneously, used AlphaFold predictions plus low-resolution experimental data to assemble it like a giant Lego set. Pretty much every pharma executive he has spoken to says their teams are using it.

Not everything will be open. Some follow-ons will be commercial, because that is how you gather enough resources and impact; some will be nonprofit; some, especially in synthetic biology, have dual-use risks. Before releasing AlphaFold they consulted 30 bioethicists. Fridman asks whether an AI will ever undeniably earn a Nobel. Hassabis draws a sharp line. Today's systems are tools, like Galileo's telescope: credit goes to Galileo, not the lens. Averaging the internet and PubMed is still a tool. To deserve the prize, a system would need to invent something like general relativity on its own — invent Go, not just play a brilliant move. That said, a lot of the next century's Nobels may already be sitting on the internet, unconnected. Breakthroughs, he thinks, will keep arriving at the intersection of fields; DeepMind itself was neuroscience cross-bred with AI engineering. What human minds cannot imagine is what it feels like to hold the whole internet at once. Wikipedia, maybe. The whole web is beyond comprehension, which is why large models surprised everyone — and why the connections waiting in that pile are a mystery.

Disease is one of the two places Hassabis most wants AI to help; energy and climate are the other. Fusion still has physics, materials, and engineering mountains to climb. DeepMind's method on a new field is to sit with world-class domain experts, list every bottleneck, and pick the one or two that today's methods can chew on and that are interesting as AI research. Plasma control was a perfect match. With EPFL in Switzerland they published a *Nature* paper on deep RL for tokamak plasmas. The plasma is on the order of a million degrees Celsius, hotter than the surface of the sun, so no material can hold it; superconducting magnets have to suspend it. It is unstable. You are holding a mini-sun in a box, and you have milliseconds to move the field. Classical controllers were hand-crafted and could not react in the moment. EPFL had a simulator, the state was observable, actions were coil voltages, and the reward was whether the plasma held. Their controller could carve the plasma into shapes, including "droplets" that are better for energy production, and hold them for a record time. They were already talking to fusion startups about the next bottleneck.

A second paper took on density functional theory: learning a functional that approximates the Schrödinger equation well enough to describe how electrons — and therefore materials — behave. Exact solutions only work for tiny molecules; for decades people have written empirical approximations to guess the electron cloud. DeepMind's move is to learn the functional instead of handwriting it. They can run vast molecular-dynamics simulations on their own clusters, generate almost as much data as they want, and train a function that covers more chemistry while staying close to the expensive simulation. Hassabis jokes that when the machines are not serving cat videos, the spare compute runs quantum chemistry. Once you have the functional, it is far cheaper than running the full simulation.

The reason he has worked on AI his whole life, he says, is to build a tool that helps us understand the universe — physics, the nature of reality. Current systems cannot crack physics, but when AGI arrives, that is among the first things he would apply it to. Time, consciousness, gravity, life: we interact with them constantly and have no idea how they work. To get through the day we pin assumptions on them and box them off. A true scientist keeps taking those boxes off the shelf. AI might let us refuse premature assumptions and build experimental tools that can test ideas that are untestable today — Scott Aaronson's questions about how much information a Planck unit of spacetime can hold, or whether the universe computes. He thinks of all knowledge as a tree we have barely scratched since the Enlightenment. AI will turbocharge the climb, not only by finding patterns but by designing new instruments.

Fridman worries that the human cognitive tree is a small subset of possible trees, and that non-human systems might go farther and bring conclusions back. Hassabis splits it into three nested sets: what we understand today, what a human mind could in principle understand, and the totality of what there is to understand. The third set may have hard walls — what is outside the universe may be unknowable from inside it. Or the walls may just be our three-pound brains, stuck on 3D space plus time, while tools think in 11 or 12 dimensions. His analogy is speed chess with Garry Kasparov: you cannot find his move, but he can explain it, and afterward you understand. You can appreciate Mozart without being able to write Mozart. Richard Feynman said that if you cannot explain a complex topic simply, you do not understand it. Hassabis can imagine an AI sighing at how dumb we are — and us talking trash back: if you were intelligent, you would make it simple.

Hassabis's view on alien civilizations runs against the usual hope: the most likely scenario, given the evidence, is that we are alone. The argument is a timing argument. A small quirk in Earth's history — the asteroid hitting the dinosaurs a million years earlier — could have put us a million years ahead or behind, and a million years is nothing on cosmological clocks. If humanity does not destroy itself in the next few centuries, builds AI, and goes to space, von Neumann's self-replicating probes would seed the galaxy in about a million years. Dyson spheres would make stars blink on galactic scales. Radio and television have been leaking since the 1930s; a million years of that, from hundreds of civilizations, should be a cacophony. We opened our ears in the space age and heard nothing.

He does not buy the usual replies. The search is incomplete, we are on the wrong bands, aliens would be too strange to notice — if they were everywhere, he says, it should not be this hard. The zoo hypothesis, a galactic non-interference rule, asks us to believe that hundreds of civilizations share priorities when humans cannot even coordinate climate change inside one species. And if everything we see is a hologram projected for our own good, that is experientially equivalent to the simulation hypothesis: you still cannot see true reality. Thoughts as a communication channel, muses, consciousness itself as alien contact — some species might work that way, but they should not all be uniform. There ought to be a distribution: aggressive, curious, stoical, primitive, close to our level.

If we are alone, that is oddly comforting on the Great Filter: it probably means the hard gates are behind us. First life from chemical soup looks difficult, but Hassabis would not be shocked by single-celled life elsewhere. Multicellularity looks much harder — swallowing a mitochondrion and, instead of digesting it, making it part of yourself. Nick Lane's book on the great inventions of evolution is on his shelf. Intelligence, which as far as we know evolved only once on Earth, is a later candidate. For the origin of human intelligence he lists fire and cooking (energy efficiency), tools and spears (the megafauna extinctions after humans reached the Americas), and tribal cooperation plus language (outcompeting less cooperative cousins). Brains are expensive, about 20 percent of the body's energy; a serious game of chess burns as much as a Formula 1 driver uses in a race. Half a brain does not pay. A monkey-level IQ cannot justify the metabolic bill until you cross the human threshold, and that jump is so hard it happened once. AI has the same shape: for a long time a crafted chess engine is easier than a general learner, because the general system starts out terribly weak.

Does AGI require solving consciousness? Hassabis's guess is that the two are doubly dissociable. High consciousness, low intelligence: dogs, dolphins, self-aware, social, apparently dreaming, not smart by IQ. High intelligence, low consciousness: AlphaGo, superhuman at a domain, with no feeling of someone home. Maybe our definition of intelligence is too narrow, Fridman suggests. Hassabis agrees that a dog's specialty is a kind of intelligence, squeezed into a narrow channel by evolution — and, he jokes, a wildly successful one, given that dogs have convinced the entire human species to feed and service them. If there is a choice, he would build AI systems that are not conscious at first, just tools, until we understand the capabilities better.

On the Google engineer who claimed LaMDA was sentient, Hassabis is blunt. He has never met a language model that was sentient. None of today's systems, in his view, have one iota of consciousness. What we are seeing is a projection: brains are trained to read agency into almost anything, including inanimate objects, and language makes anthropomorphism easy. Eliza, in the 1960s, just played psychologist and bounced your words back, and it fooled people. That is one reason the Turing Test is a flawed formal test — it depends on the sophistication of the judge. His preferred working definition, from Max Tegmark, is that consciousness is the way information feels when it is processed. It is not very operational, but the angle is right. Neuroscience suggests prerequisites that are necessary and not sufficient: a self/other distinction, preferences coherent over time, memory.

The sharpest point in the conversation, and one he has debated with Daniel Dennett, is that we judge other humans conscious on two legs and usually discuss only one. The first is behavior: does it talk and act like a minded being. The second is substrate: we run on the same few pounds of carbon-based flesh, so the parsimonious explanation is that you feel what I feel. We will never have substrate equivalence with a machine, so we will have only behavior, and that is exactly the reason we extend the benefit of the doubt to animals. The practical problem arrives before the philosophy is settled. When millions or billions of people believe an AI has a soul, and that belief gives them real comfort, what then? His current position is that language models are not ready for large-scale deployment. Interpretability, analysis tools, and guardrails are missing. One policy he leans toward: an AI system should probably always announce that it is an AI. Philosophy, social science, theology, arts and humanities have to come in, because they are the fields that know what it means to be human. Dual-use technologies get used first by bad actors, naive actors, and crazy ones — social media already taught that lesson — and AI adds a further twist: it may eventually have its own agency. Testing has to follow the scientific method, controlled experiments, not live A/B tests. An image app can be rolled back. An AI that goes wrong may do the harm before you can fix it.

Fridman asks a dark personal question. Superintelligent AI would be among the most powerful things in the world, power corrupts, and Hassabis is among the people most likely to hold the switch. Has he thought about that? All the time, he says. The defenses are ordinary on purpose. Stay grounded and humble: his best friends are still the group from undergraduate Cambridge, and family is an anchor. Stay multidisciplinary: no matter how good you are at one subject, someone is better, and relearning a field from scratch is humbling. Biology, for him, has been five years of that, joyful and chastening. Keep ethical, grounded people around you. The larger picture is radical abundance — curing disease, solving the obvious problems, eventually going to the stars to find the aliens, or to find out why they are not there. AI is too big to be run by one person or organization; it should belong to the world. Culture and values will leave a residue in the system even if it learns most of its knowledge itself. Removing scarcity should reduce zero-sum thinking, and many historical atrocities did happen under resource stress, but it is not enough. Some people just want power over others, and abundance will not dissolve that.

He is a complete night owl and has optimized around it. He arrives around 11 a.m., runs back-to-back meetings until about 7 p.m. — collaboration and management — then goes home for family, dinner, a little rest. Around 10 or 11 p.m. a second workday starts, lasting until 4 or 5 a.m.: thinking, reading papers, writing papers. He no longer codes; it is not an efficient use of the hours he has. Pencil and paper still beat the screen for working things out. He often prints papers and marks them, because handwriting sticks in the brain. The golden hours are 1 a.m. to 3 a.m., the world quiet, music on, the questions that lock a team into years of work: is this method right, will RL work here, what benchmark, do we have to build one from scratch. Night hours are flexible. If he is in flow he can go until 6 a.m. and pay for it the next day. Morning people cannot do that; breakfast and the 8 a.m. meeting break the thread. Daytime is violent context-switching across whole scientific fields. He enjoys it. He has always been a generalist — one reason he left chess was the realization that there were too many other great games to spend a life on only one.

Advice to students comes in two parts. First, find your true passions while you are young and can take risks, by exploring widely, especially at the joints between fields. Second, know yourself: when you work best, how you study, how you handle pressure; improve the weaknesses, then hone the unique strengths, because those are what the world cannot replace. The career that matters is the thin intersection of real excitement and real distinctive skill. The meaning of life, for the life he is living, is to gain knowledge and understand the universe. The classical Greeks treated that as one of the few true virtues, and it spills: the more you understand the world and people, the more compassionate and tolerant you become. What troubles him is that the universe seems structured so that science works — results repeat — and that computers are even possible, built from silicon, the common sand of the crust. If chips required diamond, we would have had one computer. "A lot of things are kind of slightly suspicious to me," he says. It feels like a puzzle designed to be solvable, and the more you learn, the more you see how little you know.

Fridman asks the last nested question. If Hassabis is the one who walks into the room with the first system that reaches human level and goes beyond it, and he gets one question, what does he ask? "What is the true nature of reality." He might not understand the answer. It might be 42. He imagines a deep sigh, then the system taking him by the hand: a more careful specification of physics, glimpses of what today's theories missed, a deeper or simpler frame than the Standard Model we keep patching, folding in consciousness, life, gravity — the mysteries humans have wondered about for thousands of years. Maybe, Hassabis says, it would just draw him a picture.