Published Jul 6, 20262:34:56 video46 min readAdded Jul 11, 2026Open on YouTube →
At a glance
This is the Lex Fridman Podcast episode 475, the second appearance of Demis Hassabis, the Nobel Prize winning leader of Google DeepMind. For roughly two and a half hours Fridman and Hassabis push on one idea from every side: that the universe is built out of information, that most of nature is therefore learnable by ordinary computers, and that this is why systems like AlphaFold and Veo 3 work at all. From there they walk the whole map, the P versus NP question as a physics question, video games as co created worlds, AlphaEvolve and evolutionary search, the dream of a simulated cell, the origin of life, a fifty percent bet on AGI by 2030, scaling laws, compute, fusion and solar, the war for talent, the future of programming, John von Neumann, p(doom), and consciousness.
Hassabis argues for a specific worldview. Nature has structure because it was shaped by selection over billions of years, so its patterns are not random, so a neural network can find the low dimensional manifold that generated them. That is why protein folding turned out to be tractable on classical machines, and it is the seed of a possible new complexity class he is quietly working on in his spare time. Fridman keeps pulling the conversation toward the deepest mysteries, life, time, consciousness, and Hassabis keeps insisting that building artificial intelligence is the sharpest tool humanity has ever had for answering them. The final half hour is Fridman alone, on David Foster Wallace, attention, and a personal note about his years at MIT and Drexel. This page rebuilds the entire conversation in order, keeps the numbers and named systems, and turns the core ideas into diagrams you can read.
The deep explanation
The Nobel conjecture: any pattern nature can make, a classical machine can learn
Fridman opens on a line from Hassabis's Nobel lecture, a deliberately provocative conjecture: any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm. Hassabis says being a little provocative is a tradition of Nobel lectures and he wanted to honor it. Step back and look at the Alpha X projects, AlphaGo and AlphaFold, and what they really do is build models of combinatorially enormous spaces. If you tried to brute force the best move in Go or the exact shape of a protein by enumeration, there would not be enough time in the lifetime of the universe. Instead both systems learned a model of the environment that guides the search and makes it tractable.
Why should that even be possible? Proteins fold in milliseconds inside our bodies, so physics already solves the problem that AlphaFold now solves computationally. Hassabis's answer is that natural systems have structure because they were subject to evolutionary processes that shaped them, and if that is true the structure can be learned. He calls the general version survival of the stablest. There is evolution for living things, but there is also geological time shaping mountains through weathering over thousands of years, and cosmological time shaping the orbits of planets and the shapes of asteroids. Everything around us, even the more stable chemical elements, has survived some selection pressure, so it carries a non random pattern that a network can reverse learn as a manifold that guides the search to the right answer. The exception is man made or abstract things such as factoring large numbers. If the number space has no pattern and is uniform, there is nothing to learn and you are back to brute force, which is where a quantum computer might be needed. But most of nature is not like that.
Figure 1. Hassabis's core intuition. The space of possible protein shapes is larger than the number of atoms in the universe, yet a real protein slides down an energy funnel to one native structure in milliseconds. Because selection carved that funnel, its shape is a low dimensional manifold a classical neural network can learn and descend, which is why AlphaFold works and why he thinks most of nature is efficiently modelable.
Information as the fundamental unit, and P versus NP as a physics question
This leads Hassabis to a proposed new complexity class, a set of learnable natural systems, alongside the classic entries in the complexity zoo. It is exactly what he says he works on in his few spare moments with a few colleagues: whether there is a class of problems solvable by this kind of network mapped onto natural systems that have physical structure. It fits how he thinks about physics in general. Information is primary, more fundamental than energy and matter, all convertible into each other, with the universe best understood as an informational system. If physics is informational, then P versus NP becomes a physics question, and its answer would be enlightening.
Fridman notes the pattern: Christian Anfinsen said something that sounded crazy in his own Nobel speech, and decades later Hassabis and John Jumper got a Nobel for solving the folding problem it pointed at. Hassabis thinks a huge class of problems can be couched the AlphaGo and AlphaFold way, where you model the dynamics of the system so that the search for a solution or the next step becomes efficient, effectively polynomial time and tractable on a classical Turing machine, which is what a neural network runs on. The AI community has already shown classical systems can go much further than people thought a decade or two ago, when many assumed protein folding might need quantum machines. AGI built as a neural network on a neural network on a classical computer would be the ultimate expression of that, and the bounds of what such a system can do speak directly to P versus NP.
They probe the edges. Cellular automata and most emergent systems are probably modelable by a forward simulation, sitting right on the boundary; genuinely chaotic systems, where initial conditions dominate and the end state is uncorrelated, are the hard case. Even highly nonlinear dynamical systems and fluids, the Navier Stokes regime traditionally thought intractable and compute hungry, may surprise us. Fridman brings up his conversation with Terence Tao about singularities that break the mathematics, and Hassabis points at Veo, which models liquids, materials, and specular lighting surprisingly well. He used to write physics and graphics engines in his gaming days and knows how painstaking that was, yet these models reverse engineer it from watching YouTube videos, presumably by extracting a lower dimensional manifold of how materials behave. Maybe, he says, that is true of most of reality.
Veo 3 and what it means to understand reality
The physics is what fascinates Hassabis about Veo 3, more than the comedy, the native audio, or the uncanny realism of the humans. It is not perfect but it is pretty good, and the real scientific question is what the model has to understand to do it. A cynic says a diffusion model understands nothing, but you cannot generate coherent video like that without some form of it, which drags our own philosophical notion of understanding to the surface. Hassabis draws the line carefully: to the extent it can predict the next frames coherently, that is a form of understanding, not the deep anthropomorphic kind, but enough modeled dynamics to produce eight seconds of consistent video that is hard to fault at a glance. Wind that forward two or three years. He calls it intuitive physics, the way a human child grasps that a pushed glass shatters and spills, rather than a PhD unpacking the equations. It is the base layer people call common sense.
That result challenges a belief Hassabis himself held. Theories of action and perception in neuroscience say you need to act in the world, through embodiment or robotics or at least simulated action, to truly perceive it. Five or ten years ago he would have agreed you need that to learn intuitive physics. Yet Veo learns it through passive observation, which he finds surprising and, again, a hint about the underlying nature of reality. The next stages make those videos interactive, so you can step in and move around, which given his games background would be mind blowing, and takes you toward what he calls a world model, a model of the mechanics of the world and the things in it, which is exactly what a true AGI needs.
Video games: open worlds, co creation, and the dream unbuilt
Fridman brings up a post on X from Jimmy Apples asking for playable world models built from Veo 3 videos, which Hassabis quote posted with "Now, wouldn't that be something?" Games were his first love and his first professional AI work as a teenager. He dreams about what he could have built in the 1990s with today's systems. His favorite genre is the open world game, a simulation plus AI characters plus a player whose actions the simulation adapts to, so every player co creates a unique experience. He worked on Theme Park at Bullfrog, where each player's game was their own. The problem is that open worlds are hard to author because you must have compelling content in whatever direction the player goes, and the classical cellular automata tricks that gave emergent behavior were always fragile and limited. Now, within five to ten years, AI may be able to generate a dramatic narrative on the fly around whatever the player chooses, the ultimate choose your own adventure, an interactive Veo wound forward.
Fridman sharpens the point. Real open world design is not the illusion of choice that The Stanley Parable satirizes, nor the two hardcoded doors of most games, but a world where opening any door genuinely defines what you see. He loved the random dungeon generation of The Elder Scrolls II: Daggerfall, where you might be the only person to ever see a given room. Hassabis wants better than random and better than a hardcoded A or B branch, which requires generative systems producing content on the fly, since you cannot pre build infinite assets and AAA games are already expensive. He points to Black & White, which had an early reinforcement learning creature you nurtured, and which mirrored your behavior toward the villagers, mean if you were mean, protective if you were good. His whole career is a follow on from those hardcoded beginnings to fully general learning systems chasing the same goal.
He and Elon Musk are both gamers itching to build a game. The sad irony is that success in serious science may leave him only building the tools others use to make the game he dreamed of. Two ways out: vibe coding good enough to build one in his spare time, which he is itching to try, or a post AGI sabbatical after the technology is safely stewarded into the world, alongside his physics theory. Asked to choose between solving P versus NP and making a game, he says in his world they are the same question, because the ultimate game would be an open world simulation as realistic as possible, which is just another way of asking what the universe is. Games get looked down on, but as AI absorbs the boring work we call jobs, games may be where humans find meaning and richer, more diverse ways of living. His favorite game of all time is Civilization, specifically Civ 1 and Civ 2, and he has avoided the latest one because it would consume him. He learned to program on a ZX Spectrum and then his beloved Commodore Amiga 500, in a UK era when the forefront of computing, AI, graphics, physics engines, and even GPUs, was being pushed forward by games fused with art.
AlphaEvolve: language models proposing, evolution searching
Hassabis calls AlphaEvolve one of the most underappreciated recent results. It is a Google DeepMind system that evolves algorithms: the evolutionary process does the search while a large language model proposes where to look. He frames it as a template, foundation models combined with other computational techniques, evolutionary computing here, but also Monte Carlo tree search or other reasoning algorithms layered on top. The pattern is always a model of the system's dynamics plus a search on top to reach a novel region. AlphaGo used Monte Carlo tree search, and that is how it found move 37, a never before seen Go strategy, going beyond everything already known. Bolt a fairly simple search onto a good model, add an objective function to hill climb, and you can reach genuinely new discoveries, which is the point for science and medicine.
What excites him is that evolution is good not just at natural selection but at combining components and building increasingly complex hierarchical systems through mutation and recombination, in the space of programs, which generalizes almost everything. The historical limit is that traditional evolutionary methods, well studied in the 1990s and 2000s, could never evolve genuinely new emergent properties, only reshuffle the ones you put in. Natural evolution clearly overcame that, going from bacteria to us, so combining it with foundation models might finally unlock new emergent capabilities. Fridman keeps a Tiktaalik skull on his desk, one of the early creatures that crawled from water onto land, as a daily reminder of what a relatively simple algorithm produced over four billion years running on the physics substrate of the universe.
Figure 2. The hybrid pattern Hassabis keeps returning to. A foundation model proposes, a search process (evolutionary here, Monte Carlo tree search in AlphaGo) explores, and an objective function selects. The loop hill climbs a known landscape but, run long enough, can jump to a genuinely novel region, the kind of leap that produced move 37.
Research taste, conjectures, and splitting the hypothesis space
Fridman raises the AI 2027 essay by Daniel Kokotajlo, Scott Alexander and others, which sketches steps toward superintelligence through a superhuman coder and a superhuman AI researcher, and its notion of research taste. Can an AI system have taste? Hassabis thinks this is the hardest thing to mimic, the judgment that separates great scientists from merely good ones, all of whom are technically competent. Taste is sniffing out the right direction, the right experiment, the right question, and picking the right question and hypothesis is the hardest part of science, which today's systems cannot do. His refrain: it is harder to come up with a really good conjecture than to solve it. Systems may soon solve hard conjectures, and AlphaProof already reached silver medal level on International Mathematical Olympiad problems, perhaps one day a millennium prize problem, but coming up with a conjecture that Terence Tao would call a deep question about the nature of mathematics is a far harder creativity, the leap of imagination Einstein made toward special and then general relativity.
A good conjecture, they agree, hits a sweet spot: interesting, falsifiable, amenable to proof with available technology, and splitting the hypothesis space roughly in two so that either outcome teaches you something. That is why Hassabis says there is no such thing as failure in real blue sky research, as long as your experiments and hypotheses meaningfully split the hypothesis space. A well designed experiment that does not work tells you as much as one that does, and the whole enterprise is a search that uses that information to decide where to go next, essentially a binary search through ideas.
The virtual cell: simulating a biological organism
Hassabis's method is to take a grand dream and break it into achievable, individually useful steps. The virtual cell is a dream he has held for about twenty five years, discussed since the 1990s with his biology mentor Paul Nurse, who founded the Francis Crick Institute and won a Nobel Prize in 2001. The goal is to model enough of a cell's internals to run experiments in silico, potentially a hundredfold speedup, with the wet lab reserved for validation. He would start with a yeast cell, which Nurse studied, because it is a full organism that is also a single cell, the simplest case, and very well understood.
AlphaFold gives the static picture, the 3D structure of a protein. The interesting biology is in the dynamics and interactions, which is what AlphaFold 3 begins to model, first pairwise, proteins with proteins, RNA and DNA. The next step would be a whole pathway, perhaps the TOR pathway implicated in cancer, and then eventually a whole cell. A cell operates at many time scales, so you would need several interacting simulations or a hierarchical system that can jump up and down the temporal stages. A key modeling decision is the granularity cutoff, and for a cell Hassabis hopes the protein level is enough, without going down to the atomic or quantum level, with AlphaFold as the basis and higher level simulations built on top to give emergent behavior. AlphaGenome, which predicts how small genetic changes link to function, is another piece creeping toward that complexity.
1990s Hassabis builds AI for games at Bullfrog, including Theme Park, then the learning creature in Black & White, his first major AI systems.
2010DeepMind is founded in London to solve intelligence and use it to solve everything else.
2014Google acquires DeepMind.
2016AlphaGo beats Lee Sedol 4 to 1. Move 37 in game 2 is a new strategy no human had played; Lee Sedol answers with move 78 in game 4.
2017AlphaZero masters Go, chess and shogi from self play. The Transformer is published, work Hassabis credits to Noam Shazeer and colleagues.
2020AlphaFold 2 effectively solves single protein structure prediction at CASP14.
2021 The AlphaFold Protein Structure Database is opened to the world.
2024AlphaFold 3 predicts interactions across proteins, RNA, DNA and ligands. AlphaProof hits silver medal level at the math olympiad. Hassabis and John Jumper win the Nobel Prize in Chemistry.
2025AlphaGenome, AlphaEvolve, Veo 3, WeatherNext, Gemini 2.5 and DolphinGemma arrive in a single relentless year.
Figure 3. The Alpha X lineage the conversation keeps circling, from Hassabis's game AI in the 1990s to the 2024 Nobel and the flood of 2025 systems. Each one is the same recipe applied to a new natural system.
The origin of life and the continuum from physics to biology
Could AI help simulate the origin of life, the birth of a living organism from non living chemistry? Hassabis loves this area and recommends Nick Lane and his book Life Ascending: The Ten Great Inventions of Evolution, which also bears on where the great filter might lie. He suspects the filters are mostly behind us, given how unlikely any life at all seems, and how the jump from single cell to multicellular life took roughly a billion years, with bacteria content for a very long time before one captured mitochondria through endosymbiosis. He sees no reason AI could not help by searching a combinatorial space: give it the primordial chemical soup near hydrothermal vents and some initial conditions, and see if something cell like emerges. Fridman wishes for a move 37 for the origin of life, and they converge on the idea that there may be no real line between non living and living, only a continuum connecting physics, chemistry and biology, the same process running from the Big Bang to today.
This is Hassabis's whole reason for working on AI: the ultimate tool for answering questions we still cannot answer, like a rigorous definition of life, the nature of time, consciousness, gravity, and quantum weirdness. He says these mysteries have been screaming in his face since childhood, and getting louder. Humans have always coped with deep mysteries, the sun, the rain, by staying busy, arguably distracted. Weather is the classic hard, near chaotic system, and Google DeepMind has built the best weather prediction systems in the world with WeatherNext, better than traditional fluid dynamics run on supercomputers that take days, including recent cyclone track prediction that matters enormously and must be fast as well as accurate. Fridman connects this to the storm chasers he met in Texas, extremely tech savvy people who ride the edge between living organisms in the eye of the storm and the cutting edge of forecasting, and who might one day use a DeepMind model. Hassabis would love to join a chase.
The path to AGI: a 2030 bet and the tests that would prove it
Hassabis puts a fifty percent chance on AGI within five years, so by around 2030. Much depends on the definition, and his has always been a high bar: matching the cognitive functions of the human brain, which is close to a general Turing machine and built modern civilization. To call something a true AGI, it would have to have all those capabilities consistently, not be a jagged intelligence that is brilliant at some tasks and badly flawed at others, which is exactly what today's systems are. It would also need the invention and creativity capabilities that are currently missing. One test is brute force: tens of thousands of cognitive tasks humans can do. Another is to hand 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 can find an obvious flaw. If they cannot, you can be fairly confident it is general.
Fridman pushes back that humans quickly take new capabilities for granted and zero in on the one flaw, and that even brilliant humans have limits across domains. Hassabis agrees there are also lighthouse moments he would watch for, move 37 style. One would be inventing a genuinely new conjecture or hypothesis about physics the way Einstein did, and you could even back test it by giving the system everything known up to a 1900 knowledge cutoff and seeing whether it derives special and general relativity. Another would be inventing a game as deep, elegant and beautiful as Go, not just a new move within Go. A system that could do several of these, plus pass the consistency checks, would be his sign of AGI. If it produced a real new discovery, he imagines being in the room, two or three months before announcing, trying not to tweet, validating it with world experts while the system explains its own workings. He is not worried it would be totally mysterious to the best scientists, giving the chess analogy that Garry Kasparov or Magnus Carlsen can explain a brilliant move afterward even if you could not have found it, and that explaining simply is itself part of intelligence.
On recursive self improvement, AlphaEvolve shows the shape of it, but Hassabis is cautious and not even sure end to end self improvement is desirable, since that is a hard takeoff scenario. Current systems keep a human in the loop, and today's models cannot design their own architecture from a vague, high level instruction. "Invent a game as good as Go" or "make a better version of yourself" is too underspecified, whereas honed tasks like faster matrix multiplication they handle well with incremental improvement. Fridman offers a softer path, a long sequence of incremental gains that slows as it goes, giving a gradual set of S curves rather than an explosion. The open question is whether these systems could ever produce a leap like the Transformer architecture in 2017, and Hassabis is honest that no one has shown systems that make those big leaps yet, only ones that hill climb the S curve they are already on. That leap would be the move 37 of AGI.
Scaling laws, compute, and whether progress hits a wall
Hassabis sees plenty of room left, and importantly across all three scalings happening at once, pretraining, post training, and inference time. It comes down to how innovative you can be, and he prides Google DeepMind on the broadest and deepest research bench, naming Noam Shazeer of Transformer fame and David Silver who led AlphaGo. If a new breakthrough on the order of AlphaGo or Transformers is required, he would back his team to produce it, and he actually prefers it when the terrain gets harder because that shifts the work from engineering toward research, their sweet spot. He puts it at roughly fifty fifty whether new ideas are needed or whether scaling the current approach is enough, and pushes both hard, with about half of resources on blue sky ideas. He notes that over the last ten to fifteen years something like eighty to ninety percent of the breakthroughs underpinning modern AI came from Google Brain, Google Research and DeepMind. On data he is not worried about running out of high quality human data, partly because simulations can generate synthetic data from the right distribution once you have enough real world data to build the generators, which is where he thinks we are now.
Compute is crucial and increasingly geopolitical, entangled with supply chains and energy. Training needs co located compute, and even bandwidth between data centers is a constraint. But now that AI systems are in products used by billions, inference compute is enormous, and the new thinking systems get smarter the more inference time you give them at test time, so training may become the smaller part of the total. The Veo 3 success spawned a meme about sweating servers, and DeepMind made a joke video of servers frying eggs. On hardware they push their TPU line, including inference only chips, and they build AI to cut energy use: data center cooling efficiency, grid optimization, plasma containment for fusion reactors with Commonwealth Fusion Systems, reactor design, and materials design for new solar materials, room temperature superconductors and optimal batteries, any one of which would be revolutionary. He thinks AI could materially help with these within five years.
The future of energy and radical abundance
Asked to bet on the main energy source in twenty or thirty years, Hassabis picks fusion and solar. Solar is the fusion reactor in the sky, and its real problems are batteries and transmission, plus more efficient materials, perhaps eventually space based solar or Dyson sphere ideas. Fusion is doable with the right reactor design and fast enough plasma control. Both get solved, giving clean, renewable, almost free energy. On a hundred year scale he would not be surprised to reach a type one Kardashev civilization. Cheap clean energy cascades: water access is solved by desalination, currently too expensive for all but wealthy coastal states like Singapore and Israel; you get effectively unlimited rocket fuel by splitting seawater into hydrogen and oxygen, which combined with self landing rockets turns space into a bus service; asteroid mining becomes real; and you approach Carl Sagan's vision of waking up the universe by bringing consciousness to it. A tourist flying past Earth would see traffic in space like London traffic, floating solar, a more technological surface, and the surplus energy used to preserve nature. For the first time humanity would not be resource constrained, breaking the zero sum trap into a radical abundance era, after which the big question becomes distributing that abundance fairly.
Human nature: games, conflict, and the meaning of mastery
Fridman notes a stubborn fact about human nature, the Borat style neighbor who starts trouble, that we do start conflicts. Games throughout history, he is learning, served to channel that away from hot war, and with civilization ending weapons now available it is far better to fight over a chessboard or a football pitch. Hassabis loves football as a visceral, tribal outlet that meets a human need to belong in a healthy, constructive way. He argues games like chess are great for children because they are microcosm simulations of the world, simplified versions of real situations, poker, Go, chess, diplomacy, where you can practice decision making. You only get a dozen or so genuinely big life decisions, so a safe, repeatable environment to improve your decision process is valuable. Just as important is practicing losing and winning. Fridman brings up Brazilian Jiu Jitsu, where you get your ass kicked safely and are reminded that losing is a fundamental part of life, and you can still be friends afterward. Hassabis says martial arts and chess, done in a healthy way, are about self improvement and self knowledge rather than the other person, not getting carried away by victory and staying humble through losses, always knowing there is more to learn. The deepest source of meaning is mastery, the number going up on the skill tree, the color belts, the quiet joy of doing something you could not do before, and games and sports are beautiful precisely because they measure that progress in creatures who are themselves hill climbing systems.
Google and the race to AGI: from losing to winning, and the art of the interface
A year ago, with Gemini 1.5, Google was widely seen as losing on the LLM product side; now, with Gemini 2.5, it is winning, and Hassabis took the helm. He credits an absolutely world class team, naming Jeff Dean and Oriol Vinyals among the Gemini leadership, plus great compute, and above all a research culture built by merging Google Brain with the old DeepMind and pulling together the best people and ideas. His mantra is relentless progress plus relentless shipping of that progress. On the bureaucracy of a giant company, he still runs Google DeepMind like a startup, large but decisive, chasing the best of both worlds, world class research plugged into product surfaces that reach billions the next day, while continually cutting away bureaucracy so the research and shipping cultures can flourish responsibly.
He tells a humbling story about scale. Irving Finkel, the British Museum cuneiform expert, has never heard of ChatGPT or Gemini, and his first encounter with AI is AI Mode in Google Search. A large part of the world has not met this technology yet, even as X feeds and pockets of the valley think about nothing else, which makes that first interaction a real responsibility, whether in rural India or anywhere, often as invisible improvements under the hood of Maps or Search. The Gemini team praised Hassabis as a great product person, which he traces to designing games for millions in the 1990s, the same fusion of cutting edge technology and product, and the same reliance on taste. He sees no boundary between art and science or product and research, only a continuum, and only works on products with cutting edge technology under the hood.
His product lessons: continually simplify and get out of the way of the model, because the model train is coming down the track and improving unbelievably fast, 2.5 versus 1.5 being a gigantic jump. You have to design not for what the technology can do today but for what it will do in a year, which makes you a deeply technical product person judging whether the research track will intercept your dream in six months or a year. He doubts the text box chat will survive contact with truly multimodal systems, expecting something more like Minority Report where you vibe collaboratively, and predicts we will look back on today's interfaces as archaic within a couple of years. Typing is low bandwidth, so he expects smart glasses, audio earbuds, and eventually neural devices to raise input and output bandwidth perhaps a hundredfold. Interface design is an underappreciated art form that unlocks the intelligence underneath, and he invokes Steve Jobs and the pursuit of simplicity, beauty and elegance, an interface as beautiful as Go, and foresees AI generated interfaces personalized to your aesthetic and the task, minimal for some, every parameter exposed for keyboard driven power users like Fridman.
On version numbers and the road to Gemini 3, he explains the cadence. Roughly every six months they collect the best research ideas on architecture and data, package and test them, then start a giant hero training run, followed by pretraining and an extensive post training and experimentation phase where a lot of the gains come. Version numbers refer to the base pretrained model, while interim 2.5 releases are usually patches or post training on the same architecture, and the sizes, Pro, Flash and Flash Lite, are often distilled from the biggest, Flash from Pro. They think in terms of a Pareto frontier of performance against cost and latency, aiming to have a model on every point of that frontier so any user or developer finds one that fits. Cool side quests like Veo spin off along the way, but you cannot take too many or you drown in versions, so they keep converging upstream, folding ideas from product surfaces and post training back into the core model training for the next run, which makes the main Gemini track more general and, eventually, AGI. One hero run at a time.
Figure 4. How Hassabis describes the Gemini family. Performance rises with cost and latency, and the sizes sit along a frontier so that whatever trade off a user or developer wants, one model satisfies it. Flash and Flash Lite are distilled from the biggest Pro model rather than trained separately.
Benchmarks, he says, are necessary but must not be overfit. He cites LMArena as the way that organically became a main test for chatbots, alongside academic benchmarks for math, coding, language and science, and their own internal suite. It is a multi objective optimization, and they aim for no regret improvements, gains in coding that do not degrade translation or other capabilities, which is the hard part, since more of one kind of data can quietly hurt another. Ultimately they trust direct usage signals from real users, because usefulness is hard to reduce to a number, a vibe across many people, and it would be terrifying to ship a smarter model that just feels off. Even more esoteric is persona and style, whether the system is verbose, succinct or humorous, which drifts into the personality research Hassabis did in his PhD, the five factor model, and raises genuinely new product questions about what an assistant should be like, which people can partly steer through prompt engineering. Asked the probability of Google DeepMind winning, he rejects the frame, calling winning the wrong way to look at something this consequential; the job is to steward the technology safely into the world for the benefit of humanity, and he hopes the international community rallies around that as AGI nears. He stays deliberately collaborative and on good terms with the other lab leaders, keeping communication channels open for cooperation on safety, and wishes more labs used AI for science the way DeepMind does, pointing to AlphaFold and to Isomorphic Labs, spun out of it to accelerate drug discovery.
Competition, the war for talent, and the future of programming
On the talent war and Meta paying huge salaries, Hassabis is measured. He thinks the real believers in the AGI mission mostly want to be on the research frontier so they can influence how it goes, and Meta is not currently at that frontier, so its aggressive hiring is rational for a company that is behind. There are more important things than money, though people must be paid market rates, which keep rising. He has known for thirty plus years that AGI is probably the most important technology ever, and it is striking that leaders are only now realizing it. He remembers not paying himself for a couple of years when DeepMind started in 2010 and could not raise money, whereas interns today are paid what his entire first seed round was; the world has inverted. And he keeps returning to the bigger question of what money will even mean post AGI, when energy is solved and the economy has to be rethought.
On jobs, programming is the sharp case, because AI is already strong at it and improving. Counterintuitively, coding and math turned out easier for AI because you can generate synthetic data and verify its correctness. For the next five to ten years, he thinks people who embrace these tools and become at one with them, in creative or technical fields, become superhumanly productive, and the great programmers become even better, perhaps ten times what they are today, by exploiting the tools fully. Cheaper coding opens programming to far more creatives, while the top programmers keep a large advantage in specifying architecture, framing questions, guiding coding assistants and checking their output. Value shifts, front end web design may be more automatable than game engine design or high performance back end work, which moves where humans are needed most and is scary to adjust to. He compares the disruption to the internet, mobile and the Industrial Revolution, with new jobs we cannot yet imagine, but warns this time it may be ten times the impact of the Industrial Revolution and ten times faster, ten years rather than a hundred, a combined hundredfold that society will find hard to absorb. He urges economists and philosophers to think now about ideas like universal basic provision, sharing the productivity gains as services with a basic floor, while rare skills still command more. Fridman broadens it to government as a technology that must adapt to represent people's pain without fueling division, and Hassabis agrees new governance structures and institutions will be needed, but insists the first job is to create abundance, because you cannot distribute what does not yet exist.
John von Neumann and the mad dreams of reason
Hassabis recommended Fridman the book The MANIAC by Benjamín Labatut, which even contains a strange biographical thread about Hassabis himself, blurring fiction and reality, but whose central figure is John von Neumann. Von Neumann contributed to quantum mechanics, worked on the Manhattan Project, and pioneered the modern computer and AI through the von Neumann architecture that all our machines still use. He watched physics become the atomic bomb and foresaw a similar arc for computing. Hassabis would love to meet him, praising his time at Princeton's Institute for Advanced Study, his polymath range and his foresight, and guesses von Neumann would have loved AlphaGo, having done game theory, and would barely be surprised, having foreseen learning machines that are grown rather than programmed back in the 1950s. He suspects there is lost wisdom from that era about dialogue and picking up the phone to talk to those the politicians call enemies. AI, he stresses, is a multi use technology aimed at curing disease and solving energy and scarcity, but with real risks, and von Neumann probably foresaw both, telling his wife computers would be even more impactful than the bomb.
A takeaway from the book, they agree, is that reason alone, the mad dreams of reason, is not enough to guide humanity as it builds powerful technology; there is something else, a spiritual or humanist dimension that pulls at the human spirit in a way cold reason does not. Hassabis embraces that, not necessarily as religion but as a sense of soul, of the spark that makes us human, perhaps consciousness itself. He casts technology as the enabler, the tools that let us flourish and understand, and sides with Richard Feynman that science and art are companions, that knowing why a flower's colors evolved makes it more beautiful, not less. He points to the Renaissance and Leonardo da Vinci, who saw no boundary between science, art and the sacred, and names Spinoza as a favorite philosopher for uniting the drive to understand the universe with our place in it. He worries many researchers in the field are too narrow and only understand the technology, which is why he welcomes society debating AI, the AI summits, governments engaging, and the chatbot era letting ordinary people feel cutting edge AI for themselves. Asked whether there will be a Manhattan Project style escalation, states weaponizing the technology, he hopes not, preferring something more like CERN, a research focused effort where the best minds complete the final steps responsibly before deployment, and he argues for education and immigration flowing both ways between the West and China so human ties make it harder for warmongers to divide us, with science as the connector.
p(doom): cautious optimism under deep uncertainty
Hassabis refuses to give a p(doom) number, calling the precision it implies a little ridiculous and questioning how people even derive theirs. What he will say is that the risk is definitely nonzero and probably not negligible, which is sobering, and that everything is hugely uncertain: what these systems will do, how fast they take off, how controllable they are. Some things may prove far easier than feared, others far harder. Given huge stakes both ways, curing disease, solving energy and scarcity, reaching the stars on one side, and catastrophe on the other, the only rational stance is cautious optimism. He would actually be more worried for humanity if AI were not coming, because climate, disease, aging and resources look hard to solve without it. The right response is the scientific method, more research to define and address the risks precisely, and he thinks that effort needs to be perhaps ten times larger than it is now as we approach the AGI line. He splits the danger into two: bad actors repurposing a general purpose technology, which sits in tension with his instinct for open science and open source, since it is genuinely hard to give good actors access while restricting individuals or rogue states; and, on a different time scale, increasingly agentic and autonomous systems whose guardrails must hold. AI itself might help, giving early warning on biological or nuclear misuse, as long as the AI you rely on is itself reliable, an interlocking problem that may need at least basic standards agreed between the United States and China.
What makes humans special, consciousness, and the friendly disagreement with Penrose
Returning to The MANIAC, they discuss Lee Sedol's move 78, the hand of God move in game 4, perhaps the last time a human's pure genius beat AlphaGo, a moment where human and machine seemed to inspire each other, captured in the documentary. What makes humans special as AI improves? It feels, perhaps with bias, that we are deeply special, and maybe it is not our intelligence but something outside the mad dreams of reason. Hassabis did a neuroscience PhD on imagination and memory in the hippocampus, and always believed that building an intelligent artifact and comparing it to the human mind is the best way to uncover what, if anything, is special about us. He suspects something is, and expects this journey to help define it, including any difference between carbon and silicon substrates processing information. One definition he likes: consciousness is the way information feels when we process it, not a rigorous scientific explanation but an intuitive one.
Fridman invokes Feynman, "what I cannot create I do not understand," and turns to Roger Penrose. Is consciousness a computation, and if it is information processing, is it modelable by a classical computer or is it quantum mechanical? Hassabis calls Penrose one of the greatest thinkers of the modern era and says they cordially disagree. Penrose collaborates with neuroscientists to find quantum mechanical behavior in the brain and, to Hassabis's knowledge, they have not found anything convincing, so his bet is that the brain is mostly classical computing, which would mean all its phenomena are modelable or mimicable by a classical computer. There may still be a final mystery, the felt quality of consciousness, the qualia, that is unique to the substrate, and we might approach understanding it through neural interfaces that let us feel what it is like to compute on silicon. He recounts a debate with the late Daniel Dennett about why we assume each other are conscious: partly shared behavior, but also, often overlooked, that we run on the same substrate, which makes it parsimonious to assume you feel what I feel. With a silicon AI we lose that second reason; even if it behaves conscious and claims to be, we would not know how it felt, and it could not know how we feel, at least until superintelligence perhaps bridges the gap.
Question
Hassabis
Penrose
Is consciousness computation?
Yes, information processing
Not ordinary computation
Substrate of the mind
Mostly classical computing in the brain
Quantum effects in neurons
Modelable by a classical computer?
Yes, phenomena are mimicable
No, needs new physics
Evidence so far
No convincing quantum mechanism found in the brain
Searching with neuroscientists for one
The remaining mystery
Qualia may be unique to the substrate; feel it via neural interfaces
Consciousness is fundamental and non computable
Figure 5. The cordial disagreement at the heart of the final act. Hassabis bets the brain is a classical machine, so its behavior, and perhaps eventually its feelings, are modelable; Penrose holds that consciousness needs physics beyond computation. Both agree we do not yet have the answer.
This demands what Fridman calls radical empathy, the ability to empathize across a different substrate, something humanity has never had to confront. Brain computer interfaces might let us truly feel what it is like for information to be computed on something other than carbon. Hassabis notes we already extend some empathy to higher animals, killer whales, dolphins, dogs, monkeys and elephants, which have aspects of consciousness even without high IQ, and mentions DolphinGemma, a version of their system trained on dolphin and whale sounds that might one day become a translator. What gives him hope is the almost limitless ingenuity of the best human minds, and human adaptability, the fact that hunter gatherer brains built for the tundra now fly on planes, record podcasts and inhabit virtual simulations, and have already absorbed talking to chatbots as normal. Fridman jokes that this very podcast may be replaced by AI and that he is waiting for it, and closes grateful for the human capacities for curiosity, adaptability, compassion and love, calling Hassabis one of the truly special humans in the world.
Coda: Fridman on David Foster Wallace, MIT, and Drexel
The last half hour is Fridman alone. Prompted by the twentieth anniversary of the This Is Water commencement speech David Foster Wallace gave at Kenyon College, he calls it one of the greatest and most unique such speeches ever, alongside the one by Steve Jobs, and Wallace one of his favorite writers, a man whose tragic honesty read like notes from the front lines of a battle with his own mind. He retells the parable of the two young fish who do not know what water is, and Wallace's point that the most obvious, important realities are the hardest to see and talk about, a banal platitude that can carry life or death importance in the trenches of adult existence. His takeaways: question everything, especially your basic assumptions, a deeply personal project; the central spiritual battles are fought in mundane daily moments, not on a mountaintop; and we too easily surrender our attention to distractions, the insatiable black holes of attention, when the call is to find meaning in the mundane and to be, in Wallace's phrase, unborable. He connects it to Feynman's flower, that a scientist sees more beauty in a flower than an artist alone, not less, and to Wallace's Alaskan bar story about the atheist and the religious man, whose lesson is that everything is a matter of perspective and that wisdom needs the humility to keep expanding it.
Fridman then speaks candidly about being attacked and lied about online, which breaks his heart but which he accepts as the cost of the path he has chosen. He states the facts about a recurring attack: he proudly earned his bachelor's, master's and doctorate at Drexel University, and he has been a paid research scientist at MIT for over a decade, from 2015 to today, in LIDS, the Laboratory for Information and Decision Systems, listed in the directory, with peer reviewed papers on his Google Scholar profile. He taught many lectures for fun, not as research, worked intensively his first four years, took leave in 2019 for AI and robotics projects and the podcast, and has not actively published since 2020 because, like the podcast, real research is a full time effort, which eats at him because building systems that people use is a genuine source of happiness he hopes to return to. At Drexel he took hard math and theoretical computer science courses, programmed robots, optimization, computer vision, wireless protocols and simulations in C and C plus plus, fell in love with programming, Emacs and the Kinesis keyboard, played guitar, wrote crappy poetry, and trained Judo and Jiu Jitsu that humbled him daily through his twenties. He closes by hoping people caught up in online mobs do not lose themselves, insisting there is more good than bad in people, admitting he is flawed, human, and that we are all in this beautiful mess together.
Key takeaways
Hassabis's Nobel conjecture is that any pattern nature can generate or find can be efficiently modeled by a classical learning algorithm, because selection over billions of years leaves non random structure a neural network can learn as a low dimensional manifold. He calls it survival of the stablest.
He treats information as more fundamental than energy or matter, which reframes P versus NP as a physics question, and he is quietly working on a possible new complexity class of learnable natural systems.
Veo 3's grasp of intuitive physics from passive observation challenges the belief that understanding the physical world requires embodiment, and points toward interactive world models needed for true AGI.
The Alpha X recipe is one idea reused: model a system's dynamics, then add a search on top (Monte Carlo tree search in AlphaGo, evolution in AlphaEvolve) to reach novel regions like move 37.
The hardest thing to replicate is research taste, picking the right question and conjecture, which he says is harder than solving it; today's systems can hill climb an S curve but have not shown the leaps like the Transformer.
His AGI estimate is roughly fifty percent by 2030, defined at a high bar of consistent, general, inventive cognition, tested by tens of thousands of tasks plus expert probing and move 37 style lighthouse moments.
He puts scaling at fifty fifty, betting DeepMind's deep research bench to supply any new breakthroughs, and expects inference compute to eventually dwarf training.
Fusion and solar plus cheap energy could break the zero sum resource trap into radical abundance, after which fair distribution becomes the central question.
He gives no p(doom) number but calls the risk nonzero and not negligible, arguing for cautious optimism, ten times more safety research, and CERN like cooperation over a Manhattan Project style arms race.
He bets consciousness is classical computing and therefore modelable, cordially disagreeing with Penrose, while granting that qualia may be unique to the substrate.
Chapters
0:00 Introduction
0:29 Sponsors, comments, and reflections
8:40 Learnable patterns in nature
12:22 Computation and P versus NP
21:00 Veo 3 and understanding reality
25:24 Video games
37:26 AlphaEvolve
43:27 AI research
47:51 Simulating a biological organism
52:34 Origin of life
58:49 Path to AGI
1:09:35 Scaling laws
1:12:51 Compute
1:15:38 Future of energy
1:19:34 Human nature
1:24:28 Google and the race to AGI
1:42:27 Competition and AI talent
1:49:01 Future of programming
1:55:27 John von Neumann
2:04:41 p(doom)
2:09:24 Humanity
2:12:30 Consciousness and quantum computation
2:18:40 David Foster Wallace
2:25:54 Education and research
Notable quotes
"Any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm." Hassabis, quoting his own Nobel lecture (8:40)
"I sometimes call it survival of the stablest." Hassabis, on why nature is learnable (9:20)
"I think information is primary. Information is the most sort of fundamental unit of the universe, more fundamental than energy and matter." Hassabis (12:40)
"To the extent that it can predict the next frames in a coherent way, some of that is a form of understanding." Hassabis, on Veo 3 (21:20)
"It's harder to come up with a conjecture, a really good conjecture, than it is to solve it." Hassabis (44:30)
"There's no such thing as failure really, as long as you're picking experiments and hypotheses that meaningfully split the hypothesis space." Hassabis (46:30)
"My estimate is sort of fifty percent chance in the next five years, so by 2030." Hassabis, on AGI (58:49)
"I think it's going to be ten times at least of the industrial revolution, and ten times faster as well." Hassabis, on the coming disruption (1:53:20)
"One hero run at a time." Hassabis, on the road to AGI (1:41:00)
"Winning is the wrong way to look at it, given how important and consequential what it is we're building." Hassabis (1:40:10)
"It's definitely non-zero, and it's probably non-negligible. The only rational, sensible approach is to proceed with cautious optimism." Hassabis, on p(doom) (2:05:40)
"One of the best definitions I like of consciousness is it's the way information feels when we process it." Hassabis (2:11:40)
"The most obvious, important realities are often the ones that are hardest to see and talk about." Fridman, quoting David Foster Wallace (2:20:10)
The following is a conversation with
Demis Hassabis, his second time on the
podcast.
He is the leader of Google DeepMind and
is now a Nobel Prize winner.
Demis is one of the most brilliant and
fascinating minds in the world today
working on understanding and building
intelligence and exploring the big
mysteries of our universe.
This was truly an honor and a pleasure
for me.
And now a quick few second mention of a
sponsor. Check them out in the
description or at
lexfridman.com/sponsors.
It's the best way to support this
podcast. We got Hampton for connecting
with founders and CEOs, Finn for AI
customer service, Shopify for building
e-commerce businesses, Element for daily
electrolytes, and AG1 for your health.
Choose wisely, my friends. And now on to
the full ad reads. I do try to make them
interesting, but if you must skip,
friends, please still check out our
sponsors. I enjoy their stuff, maybe you
will, too. And also, to get in touch
with me for whatever reason, go to
lexfridman.com/contact.
All right, let's go.
This episode is brought to you by
Hampton, a private community for
high-growth founders and CEOs. That's
the interesting thing about starting a
company and running a company,
especially one that's growing really
quickly, has to hire a lot, has to scale
a lot. It's perhaps a little bit
counterintuitive, but for the founder,
it can be deeply lonely. I suppose
that's one of the reasons they recommend
to have a co-founder, but even outside
of that, there's just a deep loneliness
with putting it all on the line, risking
everything, knowing that the chances of
success are low.
But if you do succeed, the gains are
huge, and you have your heart in it, you
have your dreams in it, you believe in
it, but also there's a constant roller
coaster of fear and doubt and hope and
moments of triumph, and moments of
failure. All those go back and forth and
just this is a constant psychological
turmoil. Anyway,
through all that, it's just nice to
connect with other people that are going
through the same thing, and that's what
Hampton is about. It does a thing where
every month eight founders face-to-face
have real conversations about their
daily struggles. Groups are forming in a
bunch of places, New York City, Austin,
San Francisco, LA, Miami, Denver, and so
on.
If you are a founder who's tired of
carrying it all alone, visit
joinhampton.com/lex
to see if it's a fit for you. That's
joinhampton.com/lex.
This episode is also brought to you by
Finn. It's an AI agent for customer
service. So, they are focused,
laser-focused on the customer service
application, and they want to do that
better than
anybody else in the world.
In fact, if you measure by the metric of
the number of resolutions, so when you
have the agent resolve a customer
service issue, that's resolution,
they have a 59%
average resolution rate,
which makes it the highest-performing
customer service agent on the market.
It's uh trusted by over 5,000 customer
service leaders, and even uh top AI
companies including Anthropic. The way
they design the system is uh it can
continuously improve from the
interactions, so you can uh continuously
analyze, train, test, and deploy.
Also, probably important to say, they uh
give you a 90-day money-back guarantee.
Go to fin.ai/lex to learn more about
transforming your customer service and
scaling your support team. That's
fin.ai/lex.
Shopify, a platform designed for anyone
to sell anywhere with a great-looking
online store. Even I figured out how to
create an online store at
lexfridman.com/store.
I put up a few shirts. Haven't done
anything with it since because I'm not a
serious person. There's a lot of serious
people that build real businesses on top
of Shopify. It's a platform that
connects you with millions of people
that want to buy stuff and gives you all
the tools you need and all the
integrations that you need to do just
that at scale. As we talked about with
DHH about the incredible beauty and
power of Ruby on Rails that Shopify is
powered by. I have not yet build a
serious sort of medium-scale project on
Rails. I need to. It's just I need to
actually find things that I need to do
web dev type of stuff with to inspire
myself to build something useful. I
don't want to build some weird variant
of a to-do list, especially now with the
help of LLMs, you can generate so much
of the code. So, I need to figure out
how to learn new framework and new
programming languages when LLMs can
generate so much of it. And I don't want
to do it exclusively by vibe coding
because I feel like that's not a way to
learn fully a thing, but vibe coding
does remove some of the friction of
learning. So, balancing that out is a
tricky thing to do. Anyway, that's about
the programming language and the
framework that powers Shopify. But
Shopify itself
connects buyers and sellers in an
incredible scale that's awe-inspiring.
Sign up for a $1 per month trial period
at shopify.com/lex.
That's all lowercase. Go to
shopify.com/lex to take your business to
the next level today.
Element, my daily zero sugar and
delicious electrolyte mix. I've been
traveling recently and I have a lot of
Element packets with me and I bring that
and I bring bands,
uh you call them. I don't know what
they're called. They're like rubber
bands for like um basic shoulder
exercises. So, if I have to do a lot of
either heavy lifting or uh heavy jujitsu
training, I like to warm up the
shoulders really well cuz probably cuz I
have issues with shoulders from many
years of playing tennis and many years
of doing stupidly
bench press. Anyway, I think Element is
a critical component of my workout
routine. Hydrate before, rehydrate
after, fully embrace the deliciousness
of watermelon salt flavor.
The flavor of champions, the one I
recommend. It's been quite a while since
I tried the others. They're all good,
but for me, I'm a man of focus and
dedication.
>> [gasps]
>> And I'm dedicated to watermelon salt. I
think they have actually, uh I saw a
lemonade flavor. I think a lot of people
love lemonade. So, maybe that's your
thing. For me, I'm sticking to
watermelon salt. Get a free 8-count
sample pack with any purchase. Try it to
drinkelement.com/lex.
This episode was also brought to you by
AG1, an all-in-one daily drink to
support better health and peak
performance. I travel with it. It makes
me feel like I take a little piece of
home with me. I drink it at least once a
day, very often twice a day. And they
keep innovating, they keep improving it.
They recently introduced uh AG1 next
gen, improving every aspect, more
vitamins and minerals, and upgraded
probiotics. It's funny how a morning
routine can be the source of peace and
happiness.
Cuz I uh I find that if I check my phone
at all in the first couple hours of the
day,
I get this weird anxiety
that ultimately morphs into
unhappiness. And if I don't, I'm much
more likely to sort of maintain that
deep focus. And a part of that early in
the morning is some coffee or
caffeinated drink, and then a few hours
on is AG1. And it's just
many hours of deep focus in between. It
makes me feel happy, makes me feel at
one with the universe, and it helps me
get done.
Anyway, they'll give you a 1-month
supply of fish oil when you sign up at
drinkAG1.com/lex.
This is a Lex Fridman podcast. To
support it, please check out our
sponsors in the description or at
lexfridman.com/sponsors,
and consider subscribing, commenting,
and sharing the podcast with folks who
might find it interesting.
I promise to work extremely hard to
always bring you nuanced and long-form
conversations with a wide variety of
interesting people from all walks of
life.
And now, dear friends, here's Demis
Hassabis.
>> [music]
>> In your Nobel Prize lecture,
you proposed what I think is a super
interesting conjecture that, quote, any
pattern that can be generated or found
in nature can be efficiently discovered
and modeled by a classical learning
algorithm. What kind of patterns or
systems might it be included in that?
Biology, chemistry, physics,
maybe cosmology,
>> Yeah.
>> neuroscience, what what are we talking
about?
>> Sure. Well, look, I felt that it's sort
of a tradition, I think, of Nobel Prize
lectures that you're supposed to be a
little bit provocative, and I wanted to
follow that tradition. What I was
talking about there is if you take a
step back and you look at all the work
that we've done, especially with the
Alpha X projects, so I'm thinking Alpha
Go, of course, AlphaFold. What they
really are is we're building models of
very combinatorially high-dimensional
spaces that, you know, if you try to
brute-force a solution, find the best
move in Go, or find the the exact shape
of a protein, and if you enumerated all
the possibilities, you there wouldn't be
enough time in the in the you know the
time of the universe. So, you have to do
something much smarter. And what we did
in both cases was build models of those
environments.
Um and that guided the search in a in a
smart way. And that makes it tractable.
So, if you think about protein folding,
which is obviously a natural system, you
know, why should that be possible? How
does physics do that? You know, proteins
fold in milliseconds in our bodies. So,
somehow physics solved this problem that
we've now also solved computationally.
And I think the reason that's possible
is that in nature, natural systems have
structure because they were subject to
evolutionary processes that that shaped
them. And if that's true, then you can
maybe learn
uh uh what that structure is.
>> So, this perspective, I think, is a
really interesting one.
You've hinted at at it, which is almost
like uh crudely stated, anything that
can be evolved
can be efficiently modeled. Think
there's some truth to that?
>> Yeah, I sometimes call it survival of
the stabilist or something like that
because yeah it you know it's it's of
course there's evolution for life, uh
living things, but there's also, you
know, if you think about geological
time, so the shape of mountains, that's
been shaped by weathering processes but
over thousands of years. But then you
can even take it cosmological, the
orbits of planets, the um shapes of
asteroids. These have all been survived
kind of processes that have acted on
them many many times. So, if that's
true, then there should be some sort of
pattern um that you can kind of reverse
learn and uh a kind of manifold really
that helps you
uh search to the right solution, to the
right shape, um and actually allow you
to predict things about it uh in an
efficient way. Because it's not a random
pattern. Right? So, um it may not be
possible for for man-made things or
abstract things like factorizing large
numbers because unless there's patterns
in the number space, which there might
be, but if there's not and it's uniform,
then there's no pattern to learn.
There's no model to learn that will help
you search. So, you have to do brute
force. So, in that case, you you know,
you maybe need a quantum computer,
something like this. But, in most things
in nature that we're interested in are
not like that. They have structure
that evolved for a reason and survived
over time. And if that's true, I think
that's potentially learnable by a neural
network.
>> It's like nature's doing a search
process.
And it's so fascinating that it's in
that search process it's creating
systems that could be efficiently
modeled.
>> That's right. Yeah.
>> So interesting.
>> So, they can be efficiently rediscovered
or recovered
because nature's not random, right?
These everything that we see around us,
including like the elements that are
more stable, all of those things,
they're subject to
some kind of selection process pressure.
>> Do you think, because you're also a fan
of theoretical computer science and
complexity, do you think we can come up
with a kind of complexity class, like a
complexity zoo type of class, where
maybe it's the set of learnable systems,
the set of learnable natural systems,
LNS?
>> This is a Demis Hassabis new class of
systems that could be actually learnable
by classical systems in this kind of
way, natural systems that can be
modeled efficiently.
>> Yeah. I mean, I've always been
fascinated by the P equals NP question
and what is modelable by classical
systems, I non-quantum systems, you
know, Turing machines in effect. And
that's exactly what I'm working on
actually in kind of my few moments of
spare time with a few colleagues about
is should there be, you know, maybe a
new class of problem that is solvable by
this type of neural network process and
kind of mapped onto these natural
systems. So, you know, the things that
exist in physics and have structure. So,
I think that could be a very interesting
new way of thinking about it. And it
sort of fits with the way I think about
physics in general, which is that, you
know, I think information is primary.
Information is the most sort of
fundamental unit of the universe, more
fundamental than energy and matter. I
think they can all be converted into
each other, but I think of the universe
as a kind of informational system.
>> So, when you think of the the universe
as an informational system, then the P
equals NP question is a
is a physics question.
>> That's right.
>> It is a question that can help us
actually
solve the entirety of this whole thing
going on.
>> Yeah, I think it's one of the most uh
fundamental questions, actually, if you
think of physics as informational.
Uh and and the answer to that, I think,
is going to be, you know, very
enlightening.
>> More specific to the P and NP question,
this again, some of the stuff we're
saying is kind of crazy right now,
just like the Christian Anfinsen Nobel
Prize speech controversial thing that he
said sounded crazy, and then you went
and got a Nobel Prize for this with John
Jumper, solved the problem. So, let me
let me just stick to the P equals NP. Do
you think there's something
in this thing we're talking about that
could be shown if you get can do
something like uh polynomial time or
constant time compute ahead of time, and
construct this gigantic model,
then you can solve some of these
extremely difficult problems in a
theoretical computer science kind of
way.
>> Yeah, I think that there are I actually
a huge class of problems that could be
couched in this way, the way we did
AlphaGo and the way we did AlphaFold,
where, you know, you you model what the
dynamics of the system is, the the the
the properties of that system, the
environment that you're trying to
understand, and then that makes the
search for the solution or the
prediction of the next step efficient,
basically polynomial time, so tractable
by a uh classical system, uh which a
neural network is. It runs on normal
computers, right? Classical computers,
uh Turing machines, in effect. And um I
think it's one of the most interesting
questions there is is how far can that
paradigm go? You know, I think we've
proven
and the AI community in general that
classical systems, Turing machines, can
go a lot further than we previously
thought. You know, they can do things
like model the structures of proteins
and play Go to better than world
champion level. And
you know, a lot of people would have
thought maybe 10, 20 years ago that was
decades away or maybe you would need
some sort of quantum machines to to
quantum systems to be able to do things
like protein folding. And so, I think we
haven't really
even sort of scratched the surface yet
of what classical systems, so-called,
could do. And of course, AGI being built
on a on a neural network system on top
of a neural network system on top of a
classical computer would be the ultimate
expression of that. And I think the
limit that you know, the the what what
the bounds of that kind of system, what
it can do, it's a very interesting
question and and and directly speaks to
the P equals NP question.
>> What what do you think, again,
hypothetical might be outside of this?
Maybe
emergent phenomena?
Like if you look at cellular automata,
some of the you have the extremely
simple systems and then some complexity
emerges.
>> Yes.
>> Maybe that would be outside or even
would you guess even that might be
amenable to efficient modeling by a
classical machine?
>> Yeah, I think those systems would be
right on the boundary, right? So,
I think most emergent systems, cellular
automata, things like that, could be
modelable by a classical system. You
just sort of do a forward simulation of
it and it probably be efficient enough.
Of course, there's the question of
things like chaotic systems where the
initial conditions really matter and
then you get to some, you know,
uncorrelated end state. Now, those could
be difficult to model. So, I think these
are kind of the open questions, but I
think when you step back and look at
what we've done with the systems and the
and the problems that we've solved, and
then you look at things like VO3 on like
video generation, sort of rendering,
physics and lighting and things like
that, you know, really core fundamental
things in physics.
Um it's pretty interesting. I think it's
telling us something quite fundamental
about how the universe is structured, in
my opinion. Um so, you know, in in a
way, that's what I want to build AGI for
is to help uh us uh as scientists answer
these questions
uh like P equals NP.
>> Yeah, [snorts] I think it will maybe
continuously surprised about what is
modelable by classical computers. I
mean, AlphaFold 3 on the interaction
side is surprising that you can
make any kind of progress on that
direction. Alpha genome is surprising
that you can map the genetic code to the
function.
Kind of playing with the emergent kind
of phenomena, you think there's so many
combinatorial options that and then here
you go. You can find the kernel that is
efficiently modelable.
>> Yes, because there there's some
structure, there's some landscape, you
know, in the energy landscape or
whatever it is that you can follow, some
gradient you can follow. And of course,
what neural networks are very good at is
following gradients. And so, if there's
one to follow an object and you can
specify the objective function
correctly, you know, you don't have to
deal with all that complexity, which I
think is how we maybe have naively
thought about it for decades, those
problems. If you just enumerate all the
possibilities, it looks totally
intractable. And there's many, many
problems like that. And then you think,
well, it's like 10 to the 300 pop
possible protein structures, uh it's 10
to the 170
possible go positions. All of these are
way more than atoms in the universe. So,
how could one possibly find the the
right solution or predict the next step?
And and it but it turns out that it is
possible, and of course, reality and
nature does do it. Right? Proteins do
fold. So, that that gives you confidence
that there must be if we understood how
physics was doing that uh in a sense,
uh then and we could mimic that process,
i.e. model that process,
uh it should be possible on our
classical systems is is is basically
what the conjecture's about.
>> And of course there's non-linear
dynamical systems, highly non-linear
dynamical systems, everything involving
fluid.
>> Yes, right.
>> I had a recent conversation with Terence
Tao who mathematically
uh it contends with a very difficult
aspect of systems that have some
singularities in them
that break the mathematics.
And it's just hard for us humans to make
any kind of clean predictions about
highly non-linear dynamical systems. But
again, to your point, we may be very
surprised what classical learning
systems might be able to do about even
>> Yes, exactly. I mean, fluid dynamics,
Navier-Stokes equations, these are
traditionally thought of as very very
difficult, intractable kind of problems
to do on classical systems. They take
enormous amounts of compute, you know,
weather prediction systems, you know,
these kind of things all involve fluid
dynamics calculations. And
um but again, if you look at something
like VEO, our video generation model, it
can model liquids quite well,
surprisingly well, and materials,
specular lighting. I love the ones
where, you know, there's there's people
who generate videos where there's like
clear liquids going through hydraulic
presses and then it's being squeezed
out. I I used to write uh physics
engines and graphics engines in in my
early days in gaming and I know it's
just so painstakingly hard to build
programs that can do that. And yet
somehow these systems are, you know,
reverse engineering from just watching
YouTube videos. So, presumably what's
happening is it's extracting some
underlying structure around how these
materials behave. 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, you know, maybe true
of most of reality.
>> Yeah, I've been continuously precisely
by this aspect of VEO 3. I think a lot
of people highlight different aspects
including the comedic and the meaning
and all that kind of stuff.
>> And then the ultra-realistic ability to
capture humans in a really nice way
that's compelling and get feels close to
reality and then combine that with
native audio. All of those are marvelous
things about VO3, but the exactly the
thing you're mentioning, which is the
physics.
>> It's not perfect, but it's pretty damn
good. And then the really interesting
scientific question is what is it
understanding about our world in order
to be able to do that? Because if the
cynical I take with the diffusion models
there's no way it understands anything.
But it seems I mean I don't think you
can generate that kind of video without
understanding and then our own
philosophical notion of what it means to
understand then is like brought to the
surface. Like do to what degree do you
think VO3 understands our world?
>> I think to the extent that it can
predict the next frames, you know, in a
coherent way. That's some of that is a
form, you know, of understanding, right?
Not in the anthropomorphic version of
you know, it's not the some kind of deep
philosophical understanding of what's
going on. I don't think these systems
have that. But they they certainly have
modeled enough of the dynamics, you
know, put it that way, that they can
pretty accurately generate whatever it
is, 8 seconds of consistent video that
by eye, at least you know, at a glance
is quite hard to distinguish what the
issues are. And imagine that in two or
three more years time. That's the thing
I'm thinking about and how incredible
that would they would look
given where we've come from, you know,
the early versions of that
one or two years ago. And so
the rate of progress is incredible and I
think
I'm like you, it's like a lot of people
love all of the the the the stand-up
comedians and the the the actually
captures a lot of human dynamics very
well and and body language, but actually
the thing I'm most impressed with and
fascinated by is the physics behavior,
the lighting and materials and liquids
and it's pretty amazing that it can do
that. And I think that shows it that it
has some notion of at least intuitive
physics, right?
How things are supposed to work
intuitively, maybe the way that a human
child would understand physics, right?
As opposed to a you know, a PhD student
really being able to unpack all the
equations. It's more of an intuitive
physics understanding.
>> Well, that intuitive
physics understanding, that's the base
layer. That's the thing people sometimes
call like a common sense. Like it it
really understands something. I think
that really surprised a lot of people.
It blows my mind that
I just didn't think it would be possible
to generate that level of realism
without understanding.
>> Mhm.
>> You there's this notion that you can
only understand the physical world by
having an embodied AI system, a robot
that interacts with that world. That's
the only way to construct an
understanding of that world.
>> But VO3 is directly challenging that, it
feels like.
>> And it's very interesting, you know,
even though if we if you would ask me 5
10 years ago, I would have said even
though I was a master in all of this, I
would have said, well, yeah, you
probably need to understand intuitive
physics, you know, like if I push this
off the table, this glass, it will maybe
shatter, you know, and the and the
liquid will spill out, right? So, we
know all of these things.
But I thought that, you know, there's a
lot of theories in neuroscience, it's
called action and perception where, you
know, you you need to act in the world
to really truly perceive it in a deep
way. And there was a lot of theories
about you need embodied intelligence or
robotics or something or maybe at least
simulated action so that you would
understand things like intuitive
physics. But it seems like
you can understand it through passive
observation, which is pretty surprising
to me. And and again, I think hints at
something underlying about the nature of
reality in in in my opinion, beyond
just the you know, the cool videos that
it generates.
And and of course, there's next stages
is maybe even making those videos
interactive. So, uh, one can actually
step into them and move around them. Um,
which would be really mind-blowing,
especially given my games background.
So, you can imagine. Uh, and then and
then I think you know, you're we're
starting to get towards what I would
call a world model, a model of how the
world works, the mechanics of the world,
the physics of the world, and the things
in that world. And of course that's what
you would need for a true AGI system.
>> I have to talk to you about video games.
So, you you were being a bit trolly.
>> [laughter]
>> I I think you're you're having more and
more fun on Twitter on X, which is great
to see. So, guy named Jimmy Apples
tweeted, "Let me play a video game of my
VO3 videos already. Uh, Google cooked so
good playable world models when?"
spelled w e n {question mark}. Uh, and
then you quote tweeted that with, "Now,
wouldn't that be something?"
So, how how hard is it to build game
worlds with AI? Maybe can you look out
into the future
uh, of video games, hm, 5 10 years out?
What do you think that looks like?
>> Well, games were my first love really,
and doing AI for games was the first
thing I did professionally in my teenage
years, and and was the first major AI
systems that I built. And, uh, I always
want to I have I want to scratch that
itch one day and come back to that. So,
you know, and I will do, I think, and
um, I think I sort of dream about, you
know, what would I've done back in the
'90s if I'd had access to the kind of AI
systems we have today, and I think you
could build absolutely mind-blowing
games. Um, and I think the next stage is
I always used to love making all the
games I've made are open world games.
So, they're games where there's a
simulation, and then there's AI
characters, and then the player, uh,
interacts with that simulation, and the
simulation adapts to the way the player
plays. And I always thought they were
the coolest games because, uh, so games
like Theme Park that I worked on where
everybody's game experience would be
unique to them. Right? Because you're
kind of co-creating the game. Right? We
set up the parameters, we set up initial
conditions, and then you as the player
immersed in it, and then you are
co-creating it with the with the
simulation. But of course, it's very
hard to program open world games. You
know, you've got to be able to create
content whichever direction the player
goes in, and you want it to be
compelling no matter what the player
chooses.
Um and so, it was always quite difficult
to build things like cellular automata
actually type of those kind of classical
systems which created some emergent
behavior. But they're always a little
bit fragile, a little bit limited. Now
we're maybe on the cusp in the next few
years, 5-10 years, of having AI systems
that can truly create around your
imagination,
can now sort of dynamically change the
story, and story tell the narrative
around and make it dramatic no matter
what you end up choosing. So, it's like
the ultimate choose your own adventure
sort of game. And
you know, I think maybe we're within
reach if you think of a kind of
interactive version of VO,
and then wind that forward 5 to 10
years, and you know, imagine how good
it's going to be.
>> Yeah, so you said a lot of super
interesting stuff there. So, one, the
open world
built into that is a deep
personalization, the way you've
described it. So, it's not just that
it's open world, but you can open any
door and there'll be something there.
It's that the choice of which door you
open
in an unconstrained way defines the
world you see. So, some games tried to
do that, they give you choice,
>> but it's really just an illusion of
choice because
you're only
like like Stanley Parable is game I
played. It's it's it's really there's a
couple of doors and it really just takes
you down a narrative. Stanley Parable is
a great video game, I recommend people
play, that kind of in a meta way
mocks
the illusion of choice and there's
philosophical notions of free will and
so on. But uh I do like one of my
favorite games of Elder Scrolls is
Daggerfall, I believe,
that they really played with a like
random generation of the dungeons.
>> Of if you can step in and they give you
this feeling of an open world. And
there, you mentioned interactivity,
you don't need to interact That That's a
first step cuz you don't need to
interact that much. You just When you
open the door, whatever you see is
randomly generated for you.
>> And that's already an incredible
experience cuz you might be the only
person to ever see that.
>> Yeah. Exactly. And And so, but what
you'd like is a little bit better than
just sort of a random generation, right?
So, you'd like uh and and also better
than a simple AB hard-coded choice,
right? That's not really uh open world,
right? Like as you say, it's just giving
you the illusion of choice. What you
want to be able to do is is potentially
anything in that game environment. Um
and I think the only way you can do that
is to have uh generated systems, systems
that uh will generate that on the fly.
Of course, you can't create infinite
amounts of game assets, right? It's
expensive enough already how AAA games
are made today. And that was obvious to
to us back in the '90s when I was
working on all these games. I think
maybe Black & White uh was the game that
I worked on early stages of that that
had the still probably the best AI
learning AI in it. It was an early
reinforcement learning system that you,
you know, you were you were looking
after this mythical creature and growing
it and nurturing it. And depending how
you treated it, it would treat the
villagers in that world in the same way.
So, if you were mean to it, it would be
mean. If you were good, it would be
protective. And so, it was really a
reflection of the way you played it. So,
actually all of the uh I've been working
on sort of simulations and AI
uh through the medium of games at the
beginning of my career. And And really
the whole of what I do today is still a
follow-on from uh those early more
hard-coded ways of doing the AI to now,
you know, fully general learning systems
that that are trying to achieve the same
thing.
>> Yeah, it's been interesting, hilarious,
and fun to watch you and Elon obviously
itching to create games cuz you're both
gamers.
And one of the sad aspects of your
incredible success in so many domains of
science, like serious adult stuff.
>> That you might not have time to really
create a game. You might end up creating
the tooling that others would create the
game and you have to watch
>> others create the thing you've always
dreamed of. Do you think it's possible
you can somehow in your extremely busy
schedule actually find time to create
something like black and white, some
some
an actual video game where like you
could
make the childhood dream
>> come become
>> Well, you know, there's two things ways
to think about that is maybe that with
vibe coding as it gets better, there's a
possibility that I could, you know, one
could do that actually in the in your
spare time. So, I'm quite excited about
that as a as that would be my project if
if I got the time to do some vibe
coding.
Um I'm actually itching to do that. And
then the other thing is, you know, maybe
it's a sabbatical after AGI has been
safely stewarded into the world and
delivered into the world. You know, that
and then working on my physics theory as
we talked about at the beginning, those
would be the two my my two post-AGI
projects, let's call it that way.
>> I would I would love to see which
post-AGI
would you choose, solving
the problem that some of the smartest
people in human history contended with,
so P equals NP,
>> or creating a cool video.
>> Yeah. Well, they they might but in my
world they'd be related because it would
be an open world simulated game
as realistic as possible. So, you know,
what what is what is the universe?
That's that's that's speaking to the
same question, right? NP equals NP, I
think all these things are related, at
least in my mind.
>> I mean, in a really serious way, so of
video games sometimes are looked down
upon. It's just this fun side activity.
But especially as AI does more and more
of the difficult boring
tasks, something we in in modern world
called work.
You know, video games is the thing in
which we may find meaning, in which we
may find like what to do with our time.
You could create incredibly rich
meaningful experiences. Like that's what
human life is. And then in video games
you can create
more sophisticated
more diverse
ways of living.
>> Yeah. I think so. I mean, those of us
who love games and I still do is is is
um
you know, it's almost can let your
imagination run wild, right? Like I I
used to love games um and working on
games so much because it's the fusion,
especially in the '90s and two early
2000s, the sort of golden era, maybe the
'80s of of of game of the games
industry. And it was all being
discovered, new genres were being
discovered. We weren't just making
games, we felt we were we were creating
a new entertainment medium that never
existed before. Like especially with
these open world games and simulation
games where you were co-create, you as
the player were co-creating the story.
There's no other media uh entertainment
media where you do that, where you as
the audience actually co-create the the
story. And of course now with
multiplayer games as well it can be a
very social activity and can explore all
kinds of interesting worlds in that. But
on the other hand, you know, it's very
important to
um also enjoy and experience uh the
physical world. But the question is
then, you know, I think we're going to
have to color confront the question
again of what is the fundamental nature
of reality. Uh what is the going to be
the difference between these
increasingly realistic simulations and
uh multiplayer ones and emergent um and
what we do in the real world.
>> Yeah, there's clearly a huge amount of
value to experiencing the real world
nature. There's also a huge amount of
value in experiencing other humans
directly in person, the way we're
sitting here today.
But, we need to really scientifically,
rigorously answer the question why.
>> And which aspect of that can be mapped
into the virtual world?
>> Exactly.
>> And it's not it's not enough to say,
"Yeah, you should go touch grass and
hang out in nature." It's like, "Why
exactly is that valuable?"
>> Yes. And I guess that's maybe the thing
that's been uh haunting me, obsessing me
from the beginning of my career. You can
If you think about all the different
things I've done, that's they're all
related in that way. This simulation,
nature of reality, and what is the
bounds of, you know, what can be
>> Sorry for the ridiculous question, but
so far, what is the greatest video game
of all time? What's up there? What what
makes
>> favorite one of all time is
Civilization. I I have to say. That that
was the the the Civilization 1 and
Civilization 2, my favorite games of all
time.
Um
>> I can only assume you've avoided the
most recent one because
it would probably you would That would
be your sabbatical. That would you would
disappear.
>> Yes, [laughter] exactly. They take a lot
of time, these Civilization games. So,
uh I've got to be careful with them.
>> Fun question. You and Elon seem to be
somehow solid gamers.
Uh is there a connection between being
great at gaming and and uh being great
leaders of AI companies?
>> I don't know. I It's an interesting one.
I mean, uh we both love games and uh
it's interesting he wrote games as well
to start off with. It's probably
especially in the era I grew up in where
home computers were just became a thing,
you know, in the late '80s and '90s,
especially in the UK. I had a Spectrum
and then an a Commodore Amiga 500, which
was my my favorite computer ever. And
that's where I learned all my
programming. And of course, it's a very
fun thing uh to program is to program
games. So, I think it's a great way to
learn programming. Probably still is.
And um
and then of course, I immediately took
it in directions of AI and simulations,
which so I mean was able to express my
interest in in games and my sort of
wider scientific interests all together.
And then the final thing I think that's
great about games is it fuses
um artistic design, you know, art with
the the the most cutting-edge
programming. Um, so again in the '90s,
all of the most interesting uh technical
advances were happening in gaming,
whether that was AI, graphics, physics
engines, uh hardware, even GPUs of
course were designed for gaming
originally. Um, so everything that was
pushing computing forward in the in the
'90s was due to gaming. So
interestingly, that was where the
forefront of research was going on, and
it was this incredible fusion with with
art. Um,
you know, graphics, but also music and
just the whole new media of
storytelling. And I love that. For me,
it's this sort of multi-disciplinary
kind of effort is again something I've
enjoyed my whole my whole life.
>> I have to ask you. I almost forgot about
one of the many and I would say one of
the most incredible things recently uh
that somehow didn't yet get enough
attention is AlphaGo Evolve. We talked
about evolution a little bit, but it's
the Google DeepMind system that evolves
algorithms.
>> Are these kinds of evolution-like
techniques promising as a component of
future super intelligence systems? So if
people don't know, it's kind of um I
don't know if it's fair to say it's
LLM-guided
evolution search.
>> So the evolutionary algorithms are doing
the search and LLMs are telling you
where.
>> Yes, exactly. So LLMs are kind of
proposing some possible solutions and
then you do you use evolutionary
computing on top to to to find some
novel part of the of the search space.
So actually I think it's an example of
very promising directions where you
combine LLMs or foundation models with
other computational techniques.
Evolutionary methods is one, but you
could also imagine Monte Carlo tree
search. Basically, many types of search
algorithms or reasoning algorithms sort
of on top of or using the foundation
models as a basis. So, I actually think
there's quite a lot of interesting
things to be discovered probably with
these sort of hybrid systems, let's call
them.
>> But, not to romanticize evolution, I'm
only human, but you you think there's
some value in whatever that mechanism
is? Cuz we already talked about natural
systems. Do you think where there's a
lot of low-hanging fruit of us
understanding being being able to model
being able to simulate evolution and
then using that whatever
we understand about that nature-inspired
mechanism to to then do search better
and better and better.
>> Yes. So, if you think about again
breaking down the sort of systems we've
built to their really fundamental core,
you've got like the model of the of the
underlying dynamics of the system.
And then if you want to discover
something new, something novel that
hasn't been seen before,
then you need some kind of search
process on top to take you to a novel
region of the of the of the search
space. And you can do that in a number
of ways. Evolutionary computing is one.
With AlphaGo, we just used Monte Carlo
tree search, right? And that's what
found move 37, the new kind of never
seen before strategy in Go. And so,
that's how you can go beyond potentially
what is already known. So, the model can
model everything that you currently know
about, right? All the data that you
currently have, but then how do you go
beyond that? So, that starts to speak
about the ideas of creativity. How can
these systems create something new, find
discover something new? Obviously, this
is super relevant for scientific
discovery or pushing science and
medicine forward, which we want to do
with these systems. And you can actually
bolt on some uh, fairly simple search
systems on top of these models and get
you into a new region of space. Of
course, you also have to, um, make sure
that uh, you're not searching that space
totally randomly, or it would be too
big. So, you have to have some objective
function that you're trying to optimize
and hill climb towards and that guides
that search.
>> But, there's some mechanism of evolution
that are interesting. Maybe in the space
of programs, but then the space of
programs is an extremely important space
cuz you can probably generalize the
the everything, you know, that. But, you
know, for example, mutation.
This is not just
Monte Carlo tree search where it's like
a search.
You could every once in a while
>> Combine things, yeah.
>> combine things out there like sub like a
components of a thing.
>> So, then, you know, what evolution is
really good at is not just the natural
selection.
It's combining things and building
increasingly complex hierarchical
systems. So, that component is super
interesting, especially like with Alpha
Evolve in the space of
>> Yeah, exactly. So, there's a you can get
a bit of an extra property out of
evolutionary systems, which is some new
emergent capability may come about.
Right? Of course, like it happened with
life. Interestingly, with naive, uh,
sort of traditional evolution competing
methods without LLMs and the modern AI,
the problem with them they there was a
they were very well studied in the '90s
and and and and early 2000s and some
promising results, but the problem was
they could never work out how to evolve
new properties. New emergent properties.
You always had a sort of subset of the
properties that you put into the system,
but maybe if we combine them with these
foundation models, perhaps we can
overcome that limitation. Obviously, uh,
natural evolution clearly did cuz it it
did evolve new capabilities. Right? So,
bacteria to where we are now. So,
clearly that it must be possible with
evolutionary systems to generate, uh,
new patterns,
going back to the first thing we talked
about and uh new capabilities and
emergent properties. And maybe we're on
the cusp of discovering how to do that.
>> Yeah, listen, uh AlphaFold is one of the
coolest things I've ever seen. I've I've
On my desk at home, you know, most of my
time is spent behind a computer just
programming.
And next to the the three screens is a a
skull of a Tiktaalik, which is one of
the early organisms that crawled out of
the water onto land. And I just kind of
watch
that little guy.
>> It's like you whatever the competition
mechanism of evolution is
is quite incredible. Yeah, it's truly
truly incredible. Now, whether that's
exactly the thing we need to do to do
our search, but never never dismiss the
power of nature with what it did here.
>> Yeah. And it's amazing. Um with with a
relatively simple algorithm, right,
effectively, and it can generate all of
this immense complexity. It emerges.
Obviously running over, you know, 4
billion years of time, but but it's it's
it's you know, you can think about that
as again a a process process that ran
over the physics substrate of the
universe for a long amount of
computational time. But then it
generated all this incredible rich
diversity.
>> So, so many questions I want to ask. So,
one, you do have a dream.
One of the natural systems you want to
try to model is is a cell.
That's a beautiful dream. Uh I could ask
about that. I also just for that purpose
on the AI scientist front just broadly.
So, there's a essay from Daniel
Kokotajlo, Scott Alexander, and others
that outline steps along the way to get
to ASI. And has a lot of interesting
ideas in it. One of which is
including a superhuman coder and a
superhuman AI researcher.
And in that there's a term of research
taste That's really interesting. So, in
everything you've seen, do you think
it's possible for AI systems to have
research taste?
To help you in the way that AI
co-scientist does to help steer
human
um human [clears throat] brilliant
scientists and then potentially by
itself to figure out what are the
directions
where you want to generate truly novel
ideas? Cuz that seems to be like a
really important of component how to do
great science.
>> Yeah, I think that's going to be one of
the hardest things to to uh mimic or
model is is this this idea of taste or
or judgment. I think that's what
separates the you know, the the great
scientists from the good scientists.
Like all all professional scientists are
good technically, right? Otherwise, they
wouldn't have been made it uh that far
in in academia and things like that. But
then, do you have the taste to sort of
sniff out what the right direction is,
what the right experiment is, what the
right question is? So, the is the is the
picking the right question is is the
hardest part of science. Um and and
making the right hypothesis. And um
that's what you know, today's systems
definitely they can't do. So, you know,
I often say it's harder to come up with
a conjecture, a really good conjecture,
than it is to solve it. So, we may have
systems soon that can solve pretty hard
conjectures.
Um you know, I I am in math olympiad
problems. Well, we we you know,
AlphaProof last year, our system got you
know, silver medal in that. Really hard
problems. Maybe eventually we'll be able
to solve a millennium prize kind of
problem. But could a system have come up
with a conjecture worthy of study that
someone like Terence Tao would have
gone, you know what? That's a really
deep question about the nature of maths
or the nature of numbers or the nature
of physics. And that is far harder type
of creativity. And we don't really know
today's systems clearly can't do that.
And we're not quite sure what that
mechanism would be. This kind of leap of
imagination. Like like Einstein had when
he came up with, you know, special
relativity and then general relativity
with the knowledge he had at the time.
>> As for as for conjecture,
the
you want to come up with a thing that's
interesting is amenable to proof.
>> So like it's easy to come up with a
thing that's extremely difficult. Yeah.
It's easy to come up with a thing that's
extremely easy, but that at that very
edge
>> That sweet spot, right? Of of basically
advancing the science and splitting the
hypothesis space into two ideally,
right? Where the if it's true or not
true, you you've learned something
really useful. And
um and and that's hard. And and and and
making something that's also
uh you know, falsifiable and within sort
of the technologies that you have you
currently have available. So it's a very
creative process actually, highly
creative process that um I I think just
a kind of naive search on top of a model
won't be enough for that.
>> Okay, the idea of splitting the
hypothesis space into super interesting.
So
uh I've heard you say that there's
basically no failure in or failure is
extremely valuable if it's done if you
construct the questions right, if you
construct the experiments right, if you
design them right, that failure or
success are both useful.
So perhaps because it splits the
hypothesis space into two, it's like a
binary search.
>> Yes. That's right. So when you do like,
you know, real blue sky research,
there's no such thing as failure really
as long as you're picking experiments
and hypotheses that that that that
meaningfully split the hypothesis space.
So, you know, and you learn something
you can learn something kind of equally
valuable from an experiment that doesn't
work. That should tell you if you've
designed the experiment well and your
hypotheses are are interesting, it
should tell you a lot about where to go
next. And um and then it's you're
effectively doing a search process
um and using that information in in you
know, very helpful ways.
>> So
to go to your dream
>> [snorts]
>> of uh modeling a cell, uh what are the
big challenges that lay ahead for us to
make that happen? We should maybe
highlight that AlphaFold I mean, there's
just so many leaps.
>> So, AlphaFold solved, if it's fair to
say, protein folding, and there's so
many incredible things we could talk
about there, including the open
sourcing. Uh
the everything you've released.
AlphaFold 3 is doing protein RNA DNA
interactions,
>> which is super complicated and and
fascinating that's amenable to modeling.
AlphaGenome uh predicts uh how small
genetic changes, like if we think about
single mutations, how they link to
actual
uh function. So, um
those are it seems like it's creeping
along to a sophisticated to much more
complicated uh things like a cell. But,
a cell has a lot of really complicated
components.
>> Yeah. So, what I've tried to do
throughout my career is I have these
really grand dreams, and then I try to,
as you've noticed, and then I try to
break but I try to break them down any
you know, it's easy to have a kind of a
crazy ambitious dream, but the the trick
is how do you break it down into
manageable, achievable interim steps
that are meaningful and useful in their
own right. And so, virtual cell, which
is what I call the project of modeling a
cell, I've had this idea, you know, of
wanting to do that for maybe more like
25 years. And I used to talk with Paul
Nurse, who is a bit of a mentor of mine
in biology. He runs the you know,
founded the Crick Institute and and won
the Nobel Prize in in 2001. Uh it it is
is we've been talking about it since,
you know, before the year you know, in
the '90s. And um and I come used to come
back to every 5 years. It's like, what
would you need to model of the full
internals of a cell so that you could do
experiments on the virtual cell, and
what those experiment, you know, in
silico, and those predictions would be
useful for you to save you a lot of time
in the wet lab, right? That would be the
dream. Maybe you could 100x speed up
experiments by doing most of it in
silico. The search in silico and then
you do the validation step in the wet
lab. That would be that's the that's the
dream. And so, but maybe now finally so
I was trying to build these components
AlphaFold being one that that would
allow you eventually to model the full
interaction a full simulation of a cell.
And I'd probably start with the yeast
cell and partly that's what Paul Nurse
studied because the yeast cell is like a
full organism that's a single cell,
right? So, it's the kind of simplest
single cell organism. And so, it's not
just a cell it's a full organism. And um
and yeast is very well understood. And
so, that would be a good candidate for
a kind of full simulated model. Now,
AlphaFold is the is the solution to the
kind of static picture of what is a what
does a protein look 3D structure of
protein look like? A static picture of
it, but we know that biology all the
interesting things happen with the
dynamics, the interactions. And that's
what AlphaFold 3 is is the first step
towards is modeling those interactions.
So, first of all pairwise, you know,
proteins with proteins, proteins with
RNA and DNA. But then the next step
after that would be modeling maybe a
whole pathway, maybe like the TOR
pathway that's involved in cancer or
something like this. And then eventually
you might be able to model, you know, a
whole cell.
>> Also, there's another complexity here
that stuff in a cell happens at
different time scales.
Is that tricky? It's like the you know,
protein
folding is you know, super fast.
>> Um I don't know all the biological
mechanisms.
>> But some of them take a long time. And
so, is that that's a level So, the
levels of interaction has a different
temporal scale that you have to be able
to model.
>> So, that would be hard. So, you'd
probably need several simulated systems
that can interact at these different
temporal dynamics. Or at least maybe
it's like a hierarchical system. So,
you can jump up and down the the
different temporal stages.
>> So,
can you avoid I mean one of the
challenges here is
not avoid simulating. For example,
the quantum mechanical aspects of any of
this, right? You want to not over model.
You could skip ahead to just model the
really high-level things that get you a
really good estimate of what's going to
happen.
>> you you go to make a decision when
you're modeling any natural system, what
is the cutoff level of the granularity
that you're going to model it to that
and then captures the dynamics that
you're interested in. So, probably for a
cell, I I would hope that would be the
protein level. Uh and that one wouldn't
have to go down to the atomic level.
Um so, you know, of course, that's where
AlphaFold stuck kicks in. So, that would
be kind of the basis and then you'd
build these um higher-level simulations
that um take those as building blocks
and then you get the emergent behavior.
>> Apologies for the pothead questions
ahead of time, but will
do you think uh we'll be able to
simulate a model the origin of life?
So, being able to simulate the first
from from non-living organisms, the the
birth of a living organism.
>> I think that's a one of the of course,
one of the deepest and most fascinating
questions. Um I love that area of
biology, you know, uh there's people
like there's a great book by Nick Lane,
one of the top top experts in this area
called the the 10 great inventions of of
of evolution. I think it's fantastic and
it also speaks to what the great filters
might be, but you know, prior or are
they ahead of us? I think I think
they're most likely in the past if you
read that book of how unlikely to go,
you know, have any life at all and then
single cell to multi-cell seems an
unbelievably big jump that took like a
billion years, I think, on Earth to do,
right? So, shows you how hard it was.
>> Bacteria were super happy for a very
long time.
>> time before they captured mitochondria
somehow, right? I don't see why not why
AI couldn't help with that. Some kind of
simulation again, it's again, it's a bit
of a search process through a
combinatorial space. Here's like all the
you know, the chemical soup that that
you start with, the primordial soup that
you know, maybe was on Earth near these
hot vents. Here's some initial
conditions. Can you generate something
that looks like a cell? So perhaps that
would be a next stage after the virtual
cell project is well, how how could you
something like that emerge from the
chemical soup?
>> Well, I would love it if there was a
move 37 for the origin of life.
>> I think that's one of the sort of great
mysteries. I think ultimately what we
will figure out is there a continuum.
There's no such thing as a line between
non-living and living. But if we can
make that rigorous.
>> That that the very thing from the big
bang to today has been the same process.
If we can break down that wall that
we've constructed in our minds of the
actual origin
of from non-living to living and it's
not a line that it's a continuum that
connects physics and chemistry and
biology.
>> As there's no line.
>> I mean, this is my whole reason why I
worked on AI and H I'm my whole life
because I think it can be the ultimate
tool to help us answer these kind of
questions. And I don't really understand
why um you know, the average person
doesn't think like worry about this
stuff more. Like how [laughter] how how
can we not have a good definition of
life and not and not living and
non-living and
>> the nature of time and let alone
consciousness and gravity and all these
things. It's it's just and quantum
mechanics weirdness. It's just to me
it's I've always had this this sort of
screaming at me in my face. And that's
it's getting louder. You know, it's like
how what is going on here? You know, in
in and I mean that in the deepest sense.
Like in the you know, the nature of
reality which has to be the ultimate
question.
>> That would answer all of these things.
It's sort of crazy if you think about
it. We can stare each other and all
these living things all the time. We can
inspect it in microscopes and take it
apart almost down to the atomic level
and yet we still can't answer that
clearly in a a way that question of how
do you define living?
It's kind of amazing.
>> Yeah, living you can kind of talk your
way out of thinking about but like
consciousness like we have this very
obviously subjective conscious
experience like we're at the center of
our own world and it it feels like
something and then how how how are you
not screaming
>> Yeah. [laughter]
>> at the mystery but all whatever. I mean
but really humans have been contending
with the mystery of the world around
them uh for a long long There's a lot of
mysteries
like what's up with the sun and and the
rain
>> like what's that about and then like
last year we had a lot of rain and this
year we don't have rain like what did we
do wrong? Humans have been asking that
question for a long time.
>> Exactly. So we're quite I guess we've
we've developed a lot of mechanisms to
cope with this these deep mysteries that
we can't fully we can see but we can't
fully understand and we have to have to
just get on with daily life and and and
we get we keep ourselves busy right? In
a way do we keep ourselves distracted?
>> I mean weather is one of the most
important questions of human history. We
still it's that's the go-to small talk
direction of the weather.
>> especially in England.
>> And then it's which is you know,
famously is an extremely difficult
system to model and
even that system
the Google DeepMind has made progress
>> Yes, we've we have we've created the the
best weather prediction systems in the
world and they're better than
traditional fluid dynamics sort of
systems that usually calculated on
massive supercomputers takes days to
calculate it we've managed to model a
lot of the weather dynamics with neural
network systems with our weather next
system and again it's interesting that
those kinds of dynamics can be modeled
even though they're very complicated
almost bordering on chaotic systems in
some cases a lot of the interesting
aspects of that
can be modeled by these neural network
systems including very recently we had
you know, cyclone prediction of where,
you know, parts of hurricanes might go.
Of course, super useful, super important
for the world. And and and it's super
important to do that very timely and
very quickly and as well as accurately.
And uh I think it's very promising
direction again of, you know, simulating
and uh so they can run forward
predictions and simulations of very
complicated real-world systems.
>> As you mentioned that uh I've got a
chance in Texas to meet a community of
folks called the storm chasers.
>> And what's really incredible about them,
I need to talk to them more, is they're
extremely tech-savvy. Because what they
have to do is they have to use models to
predict where the storm is.
>> So they're it's this
>> it's this this beautiful mix of like
crazy enough to like go into the eye of
the storm and like
in order to protect your life and
predict where the extreme events are
going to be, they have to have
increasingly sophisticated models of of
weather.
>> Yeah, it's it's a a beautiful balance of
like
being in it as living organisms and the
the cutting edge of science. So, they
actually might be using uh DeepMind
system. So, that's
>> Yeah, they I would hopefully they are
and I'd I'd love to join them on one of
those chases. [laughter] They look
amazing, right? To actually experience
it one time.
>> Exactly. And then also to experience the
correct prediction where something will
come and how it's going to evolve. It's
incredible.
>> You've estimated that we'll have AGI by
2030.
so, there's interesting questions around
that. How will we actually know that we
got there?
Uh and uh what may be the move, quote,
move 37 of AGI.
>> My estimate is sort of 50% chance by in
the next 5 years. So, you know, by 2030,
let's say. And uh so I think there's a
good chance that that could happen. Part
of it is what what is your definition of
AGI? Of course, people are arguing about
that now and and uh mine's quite a high
bar and always has been of like can we
match the cognitive functions that the
brain has, Right, so we know our brains
are pretty much general Turing machines,
approximate, and of course we created
incredible modern civilization with our
minds. So, that should also speak to how
general the brain is. And for us to know
we have a true AGI, we would have to
like make sure that it has all those
capabilities. It isn't kind of a jagged
intelligence where some things it's
really good at, like today's systems,
but other things it's really
flawed at. And and that's what we
currently have with today's systems.
They're not consistent. So, you'd want
that consistency of intelligence across
the board. And then we have some
missing, I think, capabilities. Like
sort of the true invention capabilities
and creativity that we were talking
about earlier. So, you'd want to see
those. How you test that? I think you
just test it. One way to do it would be
a kind of brute force test of tens of
thousands of cognitive tasks that,
you know, we know that humans can do.
And maybe also make the system available
to a few hundred of the world's top
experts, the Terrence Taos of each each
subject area, and see if they can find,
you know, give them give them a month or
two and see if they can find an obvious
flaw in the system. And if they can't,
then I think you're you're pretty, you
know, pretty you can be pretty confident
you we have a a fully general system.
>> Maybe to push back a little bit. It
seems like humans are really incredible
as the the intelligence improves across
all domains to take it for granted.
Uh like you mentioned Terrence Tao,
uh these brilliant experts, they might
quickly, in a span of weeks, take for
granted all the incredible things it can
do and then focus in well, "Haha, right
there."
You know, I I consider myself a
first of all human.
>> Uh
>> I identify as human. Um
is it
I you know, some people listen to me
talk and they're like, "That guy's not
good at talking. The stuttering, the you
know
>> So like even humans have obvious across
domains limits. Even just outside of
mathematics and physics and so on.
It
I I I wonder if it will take something
like a move 37. So on the positive side
versus like a barrage of 10,000
cognitive tasks where it would be one or
two where it's like
>> Holy This is personal.
>> are exactly. So I think there's the sort
of blanket testing to just make sure
you've got the consistency. But I think
there are the sort of lighthouse moments
like the move 37 that I would be looking
for. So one would be inventing a new
conjecture or a new hypothesis about
physics like Einstein did. So maybe you
could even run the back test of that
very rigorously. Like have a cutoff of
knowledge cutoff of 1900 and then give
the system everything that was you know
that was written up to 1900 and then and
then see if it could come up with
special relativity and general
relativity, right? Like Einstein did.
That that would be an interesting test.
Another one would be can it invent a
game like Go? Not just come up with move
37 a new strategy, but can it invent a
game that's as deep as aesthetically
beautiful as elegant as Go? And those
are the sorts of things I would be
looking out for.
And probably a system being able to do
several of those things, right? As for
it to be very general.
Not just one domain. And so I think that
would be the signs at least that I would
be looking for that we've got a system
that's AGI level. And then maybe to fill
that out you would also check their
consistency. You know, make sure there's
no holes in that system either.
>> Yeah, something like a new conjecture or
scientific discovery. That would be a
cool feeling.
That would be amazing. So it's not not
just helping us do that, but actually
coming up with something brand new.
>> And you would be in the room for that.
And so it would be like
probably two or three months before
announcing it.
>> And you would just be sitting there
trying not to tweet.
>> Something like that. Exactly. It's like
what is this amazing new, you know,
physics idea and then we would probably
check it with world experts in that
domain, right? And validate it and kind
of go through its workings and it I
guess it would be explaining its
workings, too.
Um yeah, be an amazing moment.
>> Do you worry that we as humans, even
expert humans like you, might miss it,
might miss it?
>> it may be pretty complicated. So, it
could be the analogy I give there is I
don't think it will be um
totally mysterious to the to the best
human scientists, but it may be a bit
like, for example, in chess, if I was to
talk to Garry Kasparov or Magnus Carlsen
and play a game with them and they make
a brilliant move, I might not be able to
come up with that move, but they could
explain why afterwards that move made
sense. And we would be able to
understand it to some degree, not to the
level they do, but in you know, if they
were good at explaining, which is
actually part of intelligence, too, is
being able to explain in a simple way
that what you're thinking about,
I think that that would be very possible
for the best human scientists.
>> But I wonder, maybe you can you can
educate me on the side of Go, I wonder
if there's moves from Magnus or Garry
where they at first will dismiss it as a
bad move.
>> Yeah. Sure. It could be, but then
afterwards they will figure out with
their intuition that that this why this
works and then and then and then
empirically, the nice thing about games
is one of the great things about games
is you can it's it's a sort of
scientific test. Does it do win the game
or not win? And then um that tells you
okay, that move in the end was good.
That strategy was good. And then you can
go back and analyze that and and and and
explain even to yourself a little bit
more why, explore around it. And that's
how chess analysis and things like that
work. So, perhaps that's why my brain
works like that cuz I I've been doing
that since I was four and you're train
you know, train it's sort of hardcore
training in that way.
>> But even even now like when I generate
code
there is this kind of nuanced
fascinating contention that's happening
where I might at first identify as a set
of generated code as incorrect in in
some interesting nuanced ways. But then
I'm always have to ask the question
is there a deeper insight here that
that I'm the one who's incorrect.
And that's going to as the systems get
more and more intelligent, you're going
to have to contend with that. It's like
what what what do you mean? Is this a
bug or a feature what you just came up
with?
>> and they're going to be pretty
complicated to do, but of course it will
be you can imagine also AI systems that
are producing that code or whatever that
is and then human programmers looking at
but also not unaided with the help of AI
tools as well. So it's going to be kind
of an interesting you know maybe
different AI tools to the ones that the
more
kind of monitoring tools than the ones
that generated it.
>> So if we look at a AGI system
sorry to bring it back up but Alpha
evolve
super cool. So Alpha evolve enables on
the programming side something like
recursive
self-improvement
potentially. Like what if we can imagine
what that AGI system maybe not the first
version but
a few versions beyond that. What does
that actually look like? Do you think it
will be simple? You think it will be
something like a self-improving program
and a simple one?
>> I mean potentially that's possible I
would say. I'm not sure it's even
desirable because that's a kind of like
hard takeoff scenario. But but you
these current systems like Alpha evolve
they have you know human in the loop
deciding on various things. They're
separate hybrid systems that interact.
One could imagine eventually doing that
end-to-end. I don't see why that
wouldn't be possible. But right now
you know I think the systems are not
good enough to do that in terms of
coming up with the architecture of the
code. Um and again, it's a little bit
reconnected to this idea of coming up
with a new conjectural hypothesis. How
like they they're good if you give them
very specific instructions about what
you're trying to do.
Um but if you give them a very vague
high-level instruction, that wouldn't
work currently. Like uh and I think
that's related to this idea of like
invent a game as good as Go, right?
Imagine that was the prompt. That's
that's pretty underspecified. And so,
the current systems wouldn't know I
think what to do with that, how to
narrow that down to something tractable.
And I think they're similar like look,
just make a better version of yourself.
That's too that's too unconstrained. But
we've done it in you know in and as you
know with AlphaFold like things like
faster matrix multiplication. So, when
you when you hone it down to very
specific thing you want, um it's very
good at incrementally improving that.
But at the moment, these are more like
incremental improvements, sort of small
iterations. Whereas if you know if you
wanted a big leap in uh understanding,
you'd need a you need a much larger uh
advance.
>> Yeah, but it could also be sort of to
push back against hard takeoff scenario.
It could be just a sequence of
incremental improvements like matrix
multiplication. Like it has to sit there
for days thinking how to incrementally
improve a thing and that it does so
recursively. As you do more and more
improvement, it'll slow down. So,
there'll be like a like the path AGI
won't be like a
it'll be a gradual improvement over
>> if it was just incremental improvements,
that's how it would look. So, the
question is could it come up with a new
leap like the Transformers architecture?
Like could it have done that back in
2017 when you know we did it and Brain
did it? And it's it's not clear that
that these systems something like
AlphaFold wouldn't be able to do make
such a big leap. So, for sure these
systems are good. We have systems I
think that can do incremental hill
climbing and that's a kind of bigger
question about is that all that's needed
from here or do we actually need one or
two more
uh uh big breakthroughs?
>> And can the same kind of systems provide
the breakthroughs also? So, maybe get a
bunch of S curves. Like incremental
improvement, but also every once in a
while leaps.
>> Yeah, I don't think anyone has systems
that can have shown unequivocally those
big leaps. That that that right? We have
a lot of systems that do the hill
climbing of the S curve that you're
currently on.
>> Yeah. And that would be the move 37 is a
>> Yeah, I think it would be a leap
uh something like that.
>> Uh do you think the scaling laws are
holding strong on the pre-training,
post-training, test and compute? Uh do
you
uh on the flip side of that anticipate
AI progress hitting a wall?
>> We certainly feel there's a lot more
room just in the scaling. So,
um actually all steps pre-training,
post-training, and
inference time. So, uh there's sort of
three scalings that are happening
concurrently.
Um and we again there it's about how
innovative you can be. And we you know,
we pride ourselves on having the
broadest and um deepest research bench.
Uh we have amazing you know, incredible
uh researchers and uh people like Noam
Shazeer who you know, came up with
Transformers and and Dave Silver you
know, who led the AlphaGo project and so
on. And um it's it's it's that research
base means that if some new new
breakthrough is required like an AlphaGo
or Transformers, uh I would back us to
be the place that does that. So, I
actually quite like it when the terrain
gets harder, right? Because then it
veers more from just engineering to to
true research and you know, research or
research plus engineering and that's our
sweet spot. And I I think that's harder
it's harder to invent things than to
than to um you know, fast follow. And um
so, you know, we don't know. I would say
it's a it's kind of 50/50 whether new
things are needed or whether the scaling
the existing stuff is going to be
enough. And so, in true kind of
empirical fashion, we're pushing both of
those as hard as possible. The new blue
sky ideas, and you know, maybe about
half our resources are on that, and then
and then uh scaling to the max the the
current the current capabilities. And um
we're still seeing some, you know,
fantastic progress on uh each different
version of Gemini.
>> That's interesting the way you put it in
terms of the deep bench
that if uh progress towards AGI
is more than just scaling compute. So,
the engineering side of the problem and
is more on the scientific side where
there's breakthroughs needed. Then you
feel confident DeepMind as well, but
Google DeepMind as well positioned to
kick ass in that domain.
>> Well, I mean, if you look at the history
of the last decade or 15 years, um it's
been I mean, you know, maybe I don't
know, 80-90% of the breakthroughs that
more that underpins modern AI field was
from, you know, originally Google Brain,
Google Research, and DeepMind. So, yeah,
I would back that to continue,
hopefully. [laughter]
>> Uh so, on the data side, are you
concerned about running out of
high-quality data, especially
high-quality human data?
>> I'm not very worried about that, partly
because I think there's enough data uh
and it's been proven to get the systems
to be pretty good. And this goes back to
simulations again. If you Do you have
enough data to make simulations or so
that you can create more synthetic data
that are from the right distribution.
Obviously, that's the key. So, you need
enough real-world data in order to be
able to uh create those kinds of
generator data generators. And um I
think that we're at that step at the
moment.
>> Yeah, you've done a lot of incredible
stuff in the side of science and
>> Doing a lot with not so much data.
>> I mean, it's still a lot of data, but I
guess enough
to get that going.
>> Exactly. So, exactly.
>> Uh how crucial is the scaling of compute
to building AGI? This is a question
that's an engineering question. It's a
almost a geopolitical question
because it also integrated into that
is supply chains and energy, a thing
that you care a lot about, which is
potentially fusion. So, innovating on
the side of energy also. Do you think
we're going to keep scaling compute?
>> I think so for several reasons. I think
compute There's There's the amount of
compute you have for training. How often
it needs to be co-located. So, actually
even like, you know, bandwidth
constraints between data centers can
affect that. So, it's it's it's There's
additional constraints even there. And
that that's important for training
obviously the largest models you can.
But there's also because now AI systems
are in products and being used by
billions of people around the world, you
need a ton of influence compute now. Um
and then on top of that, there's the
thinking systems, the new paradigm of
the last year that where they get
smarter the longer amount of inference
time you give them at test time. So, all
of those things need a lot of compute.
And I don't really see that slowing
down.
Um and as AI systems become better,
they'll become more useful and there'll
be more demand for them. So, both from
the training side The training side
actually is is only just one part of
that. It may even become the smaller
part of of what's needed
in the overall compute that that's
required.
>> Yeah, that's one sort of almost meme-y
kind of thing, which is like the success
in the incredible aspects of VL3.
There's
uh people kind of make fun of like the
more successful it becomes, the you
know, the servers are sweating.
>> Yes. [laughter] Can't have inference.
Yeah, yeah, exactly. We did a little
video of of us of the servers frying
eggs and things. And
um that's right. And And And we're going
to have to figure out how to do that. Um
there's a lot of interesting hardware
innovations that we do. As you know, we
have our TPU line. And we're looking at
like inference-only things,
inference-only chips, and how we can
make those more efficient. We're also
very interested in building AI systems,
and we have done that help with energy
usage. So, help um data center energy
like for the cooling systems be
efficient, um grid optimization,
um and then eventually things like
helping with uh plasma containment
fusion reactors. We've done lots of work
on that with Commonwealth Fusion, and
also uh one could imagine reactor
design, um and then material design I
think is one of the most exciting.
New types of solar material, solar panel
material, super room temperature
superconductors has always been on my
list of dream breakthroughs, and um
optimal batteries. And I think a
solution to any, you know, one of those
things would be absolutely revolutionary
for, you know, climate and energy usage.
And we're probably close, you know, and
again in the next 5 years to having AI
systems that can materially help with
those problems.
>> So, if you were to bet, sorry for the
ridiculous question, but what what is
the main source of energy
uh
in like 20, 30, years? Do you think it's
going to be nuclear fusion?
>> I think fusion and solar are the two
that I I would bet on. Um solar, I mean,
you know, it's the fusion reactor in the
sky, of course. And I think really the
the problem there is is is batteries and
transmission. So, you know, as well as
more efficient more more efficient solar
material, perhaps eventually, you know,
in space, you know, these kind of Dyson
sphere type ideas. And fusion I think is
definitely doable, seems, uh if we have
the right design of reactor and we can
control the plasma and uh fast enough
and so on. And I think both of those
things will actually get solved. So,
we'll probably have at least those will
probably be the two primary sources of
renewable clean almost free or perhaps
free energy.
>> What a time to be alive.
>> If I uh
traveled into the future with you 100
years from now, how much would you be
surprised if we've passed a type one
Kardashev scale civilization?
>> I would not be that surprised if there
was a like a 100-year time scale from
here. I mean, I think it's pretty clear
if we crack the energy problems in one
of the ways we've just discussed, fusion
or or very efficient solar.
then if energy is kind of free and
renewable and clean,
um then that solves a whole bunch of
other problems. So, for example, the
water access problem goes away because
you can just use desalination. We have
the technology, it's just too expensive.
So, only, you know, uh fairly wealthy
countries like Singapore and Israel and
so on like actually use it. But But if
it was uh cheap, then every then, you
know, all countries that have a coast
could. But also, you'd have unlimited
rocket fuel. You could just separate
seawater out into hydrogen and oxygen
using energy, and that's rocket fuel.
So,
uh combined with, you know, Elon's
amazing self-landing rockets, then it
could be like you sort of like a bus
service to to space. So, that opens up,
you know, incredible new resources and
domains. Uh asteroid mining, I think
will become a thing and maximum human
flourishing to the stars. Like, that's
what I uh dream about as well as like
Carl Sagan's sort of idea of bringing
consciousness to the universe, waking up
the universe. And I I think human
civilization will do that in the full
sense of time if we get AI right and uh
and and and crack some of these problems
with it.
>> Yeah, I wonder what it would look like
if you're just a tourist flying through
space.
You would probably notice Earth because
if you solve the energy problem, you
would see a lot of space rockets,
probably. So, it would be like traffic
here in London, [clears throat]
but in space.
>> Yes, [laughter] exactly.
>> a lot of rockets.
>> And then you would probably see floating
in space some kind of source of energy
like solar,
>> Yep.
>> potentially. So, Earth would just look
more on the surface more um
technological.
And then then you would use the power of
that energy then to preserve the natural
>> Like the rainforest and all that kind of
stuff.
>> for the first time in in human history,
we wouldn't be
resource constrained. And I think that's
could be amazing new era for humanity
where it's not zero-sum. Right, I have
this land, you don't have it. Or if we
take, you know, if the tigers have their
forest, then the the local villagers
can't What are they going to use? I I I
think that this will help a lot. No, it
won't solve all the problems because
there's still other human foibles that
will will will still exist, but it will
at least remove one, I think one of the
big vectors, which is scarcity of
resources, you know, including land and
more materials and energy. And
we know we should be as some of us call
it like and others call it about this
kind of radical abundance era where
there's plenty of resources to go
around. Of course, the next big question
is making sure that that's fairly, you
know, shared fairly and everyone in
society benefits from that.
>> So, there is something about human
nature where I go, you know, it's
It's like Borat, like my neighbor, like
I would
like you start trouble. We we we do
start conflicts.
And that's why games throughout, as I'm
learning actually more and more,
even in ancient history, serve the
purpose of pushing people away from war,
actually hot war. So, maybe we can
figure out increasingly sophisticated
video games that pulls they they give us
that
>> the scratch the itch of like conflict,
whatever that is, but but us the human
nature and then avoid the actual hot
wars that would come with increasingly
sophisticated technologies because we're
now have long past the stage where the
weapons we're able to create can
actually just destroy all of human
civilization. So, it's no longer
>> Um
>> that's no longer a great way
to
start with your neighbor. It's
better to play a game of chess
>> or football
>> or football. Yeah.
>> And I think I mean I think that's what
my modern sport is. So and I love
football, watching it and and I just
feel like and I used to play it a lot as
well and it's it's it's it's it's very
visceral and it's tribal and I think it
does channel a lot of those energies
into a which I think is a kind of human
need to belong to some some group and
but into a into a into a fun way healthy
way and and a not a not destructive way
kind of constructive thing. And I think
going back to games again is I think the
originally why they're so great as well
for kids to play things like chess is
they're great little microcosm
simulations of the world. They're
they're simulations of the world too.
They're simplified versions of some real
world situation whether it's poker or or
go or chess different aspects or
diplomacy different aspects of of the
real world and it allows you to practice
at them too. And cuz you know how many
times do you get to practice a massive
decision moment in your life? You know
what job to take what university to go
to? You know you get maybe I don't know
a dozen or so key decisions one has to
make and you got to make those as best
as you can.
And games is a kind of safe environment
repeatable environment where you can get
better at your decision making process
and it maybe has this additional benefit
of channeling some energies into
into more creative and constructive
pursuits.
>> Well, I think it's also really important
to practice
losing and winning.
>> Right.
>> Like losing is a really you know that's
why I love games. That's why I love even
things like Brazilian Jiu Jitsu where
you can get your ass kicked in a safe
environment over and over. It reminds
you about the way about physics about
the way the world works about the
sometimes you lose sometimes you win.
You can still be friends with everybody.
But that that feeling
of losing I mean it's a weird one for
humans to like really like make sense
of. Like that's just part of life. That
is a fundamental part of life is losing.
>> Yeah. And I think in martial arts as I
understand it, but also in things like
like chess is a lot at least the way I
took it, it's a lot to do with
self-improvement, self-knowledge, you
know, that okay, so I did this thing.
It's not about really being the other
person. It's about maximizing your own
potential. If you do it in a healthy
way, you learn to use victory and losses
in a way. Don't get carried away with
victory and and think you're the just
the best in the world. And and and the
losses keep you humble. And always
knowing there's always something more to
learn. There's always a bigger expert
that you can mentor you. You know, I
think you learn that I I I'm pretty sure
in martial arts. And and and I think
that's also the way that at least I was
trained in chess. And so in the same way
and it can be very hardcore and very
important. Of course, you want to win,
but you also need to learn how to deal
with setbacks
in a in a healthy way that and and and
and why are that that feeling that you
have when you lose something into a
constructive thing of next time, I'm
going to improve this, right? Or get
better at this.
>> There is something that's a source of
happiness, a source of meaning that
improvement stuff. It's not about the
winning or losing.
>> Yes, the mastery. There's nothing more
satisfying in a way is like, oh wow,
this thing I couldn't do before, now I
can. And and and again, games and
physical sports and and mental sports,
they're they're ways of measuring.
They're beautiful because you can
measure that that progress.
>> Yeah. I mean, there's something about I
guess why I love role-playing games.
Like the number go up of like
>> on the skill tree. Like literally, that
is a source of meaning for us humans.
Whatever our
>> Yeah, we're quite we're quite addicted
to this sort of yeah, these numbers
going up. And
and and and and maybe that's why we made
games like that because obviously that
is something we're we're we're hill
climbing systems ourselves, right?
>> Yeah, it it would be quite sad if we
didn't have any mechanism to
>> color belts. All the we do we do we do
this everywhere, right? Where we just
have this thing that
>> It's in a I don't want to dismiss that.
There is a source of deep meaning for us
humans.
Um so, one of the incredible stories on
the business on the leadership side is
um what Google has done over the past
year.
So, I
uh I think it's fair to say that Google
was losing on the LLM product side uh a
year ago with Gemini 1.5. And now it's
winning with Gemini 2.5. And you took
the helm and you led this effort. What
did it take to go from let's say {quote}
"losing" to {quote} "winning" in the in
in the span of a year?
>> Yeah, well, firstly, it's absolutely
incredible team that we have, you know,
led by Corrado and Jeff Dean and and
Oriol and the amazing team we have on
Gemini. Absolutely world-class. So, you
can't do it without the best talent. Um
and of course, you have you know, we
have a lot of great compute as well. But
then, it's the research culture we
created, right? And basically, coming
together both different groups in in
Google, you know, there was Google
Brain, world-class team, and and then
the old DeepMind. And pulling together
all the best people and the best ideas
and gathering around to make the
absolute greatest system we could. And
it has been hard, um but we're all very
competitive.
Uh and we, you know, love research. This
is so fun to do.
Um and we, you know, it's great to see
our trajectory. It wasn't a given, but
we're very pleased with um the the where
we are and the rate of progress is the
most important thing. So, if you look at
where we've come to from 2 years ago to
1 year ago to now, you know, I think our
we call it relentless progress along
with relentless shipping of that
progress is um being very successful.
And you know, um it's unbelievably
competitive uh the whole space, the
whole AI space with some of the greatest
entrepreneurs and leaders uh and
companies in the world all competing now
because everyone's realized how
important AI is. Um and it's very, you
know, been pleasing for us to see that
progress.
>> You know, Google's a gigantic company.
Can you speak to the natural things that
happen in that case is the bureaucracy
that emerges? Like you want to be
careful. Like
you know, like that the natural kind of
there's there's meetings and there's
managers and that. Like what what are
some of the challenges from a leadership
perspective breaking through that in
order to, like you said, ship? Like the
the number of products
>> Gemini related products that's been
shipped over the past years is insane.
>> Right. It is. Yeah, exactly. That's
that's what relentlessness looks like.
Um, I think it's it's a question of like
any big company, you know, ends up
having
a lot of layers of management and things
like that. It's sort of the nature of
how it works.
But I still operate and I was always
operating with Old DeepMind as a as a
startup still. Large one, but still as a
startup. And that's what we still act
like today in as with Google DeepMind.
And acting with decisiveness and the
energy that you get from the best
smaller organizations. And we try to get
the best of both worlds where we have
this incredible billions of users
surfaces incredible products that we can
power up with our AI and our and our
research.
And that's amazing and you can, you
know, there's very few places in the
world you can get that. Do incredible
world-class research on the one hand and
then plug it in and and improve billions
of people's lives the next day.
That's a pretty amazing combination. And
we're continually fighting and cutting
away bureaucracy to allow the research
culture and the relentless shipping
culture to flourish. And I think we've
got a pretty good balance whilst being
responsible with it, you know, as you
have to be as a large company and also
with a number of, you know, huge product
surfaces that we have.
>> So funny thing you mentioned about like
the the surface of the billion. I I had
a conversation with a guy named,
brilliant guy,
here at the British Museum called Irving
Finkel. He's a world expert at
cuneiforms, which is a ancient writing
on tablets. And he doesn't know about
chat GPT or Gemini. He doesn't even know
anything about AI. But his first
encounter with this AI is AI mode on the
>> Yes. Yes.
>> He's like, is that what you're talking
about this AI mode?
>> And then, you know, it's just
it's just a reminder that there's a
large part of the world that doesn't
know about this AI thing.
>> Yeah. I know, it's funny cuz if you live
on X and Twitter and I mean, it's sort
of at least my feed, it's all AI and and
there's certain places where, you know,
in the valley and certain pockets where
everyone's just all they're thinking
about is AI. But
a lot of the normal world hasn't hasn't
come across it yet.
>> And that's a great responsibility to
their first interaction.
>> On the the the grand scale of the rural
India or anywhere across the world. You
get to
>> Right. And we want it to be as good as
possible. And in a lot of cases, it's
just under the hood powering making
something like maps or search work
better. And
and it's ideally for a lot of those
people should just be seamless. It's
just new technology that makes their
lives more, you know, productive and and
and helps them.
>> A bunch of folks on the Gemini product
and engineering teams spoken extremely
highly of you on another dimension that
I almost didn't even expect cuz I kind
of think of you as the like deep
scientist and caring about these big
research scientific questions. But they
also said you're a great product guy.
Like how to create a thing that a lot of
people would use and enjoy using. So,
can you maybe speak to what it takes to
create a AI-based product that a lot of
people would enjoy using?
>> Yeah, well, I mean, again, that comes
back from my game design days where I
used to design games for millions of
gamers. People forget about that. I've
had experience with cutting-edge
technology in product. That that that
that is how games was in the '90s. And
so, I love actually the combination of
cutting edge research and then being
applied in a product and to power a new
experience. And so I think it's the same
skill really of of you know imagining
what it would be like to use it
viscerally
and having good taste coming back to
earlier. The same thing that's useful in
science
I think is can also be useful in in
product design and
I've just had a very you know always
been a sort of a multi-disciplinary
person. So I don't see the boundaries
really between you know arts and
sciences or product and research. It's
it's a continuum for me. I mean I only
work on I like working on products that
are cutting edge. I wouldn't be able to
you know have cutting edge technology
under the hood. I wouldn't be excited
about them if they were just run of the
mill products.
So it requires this invention creativity
capability.
>> What are some specific things you kind
of learned about when you
even on the LLM side you're interacting
with Gemini? You know like this doesn't
feel like the layout the the interface
maybe the trade off between the latency
like how
how to present to the user how long to
wait
>> and how that waiting is shown or the
reason capabilities. There's some
interesting things cuz it like you said
it's the very cutting edge. We don't
how to present it how to present it
correctly. So is there some specific
things you've you've learned?
>> I mean it's such a fast evolving space
where evaluating this all the time. But
where we are today is that you want to
continually simplify things
the whether that's the interface or
the interact what you build on top of
the model. You kind of want to get out
of the way of the model. The model train
is coming down the track and it's
improving unbelievably fast this
relentless progress we talked about
earlier. You know you look at 2.5 versus
1.5 and it's just a gigantic improvement
and we expect that again for the future
versions. And so the models are becoming
more capable. So you've got the
interesting thing about the design space
in in in today's world, these AI first
products, is you got to design not for
what the thing can do today, the
technology can do today, but in a year's
time. So, you actually have to be a very
technical product person because you got
to kind of a good intuition for and feel
for okay, that thing that I'm dreaming
about now can't be done today, but is
the research track on schedule to
basically intercept that in 6 months or
a year's time. So, you kind of got to
intercept where this highly changing
technology is going, as well as the
new capabilities are coming online all
the time that you didn't realize before
that can allow like these research to
work. Or now we got video generation,
what do we do with that? This multimodal
stuff, you know, is it one question I
have is is it really going to be the
current UI that we have today, these
text box chats? Seems very unlikely
given once you think about these super
multimodal
systems. Shouldn't it be something more
like Minority Report where you're you're
sort of vibing with it in a
in a kind of collaborative way, right?
It seems very restricted today. I think
we'll look back on today's interfaces
and products and systems as quite
archaic in maybe in a just a couple of
years. So, I think there's a lot of
space actually for innovation to happen
on the product side as well as the the
research side.
>> And then we were offline talking about
this keyboard is the open question is
how, when, and how much will we move to
audio as the primary way of interacting
with the machines around us versus
typing stuff?
>> Yeah, I mean typing is a very low
bandwidth way of doing even if you're
very fast, you know, typer. And I think
we're going to have to start utilizing
other devices, whether that's smart
glasses, you know, audio earbuds,
and eventually maybe some sort of neural
devices where we can increase the the
input and the output bandwidth to
something,
you know, maybe a 100X of what it is
today.
>> I think that, you know, under
appreciated art form is the interface
design. And I think you can not unlock
the power of the intelligence of a
system if you don't have the right
interface. Their interface is really the
way you unlock its power. It's such an
interesting question of how to do that.
So, how how it you would think like
getting out of the way is an real art
form.
>> Yes. You know, it's the sort of thing
that I guess Steve Jobs always talked
about, right? It's simple simplicity,
beauty, and elegance that we want,
right? And we're not there nobody's
there yet, in my opinion. And that's
what I would like us to get to. Again,
it sort of speaks to like Go again,
right? As a game, the most elegant,
beautiful game. Can you, you know, that
can you make an interface as beautiful
as that? And actually, I think we're
going to enter an era of AI generated
interfaces that are probably
personalized to you, so it fits the way
that you your aesthetic, your feel, the
way that your brain works. And and and
and the AI kind of generates that
depending on the task, you know? That
feels like that's probably the direction
we'll end up in.
>> Yeah, cuz some people are power users
and they want every single parameter on
screen, everything everything based like
perhaps me with a keyboard
keyboard-based navigation. I'd like to
have shortcuts for everything. And some
people like the minimalism.
>> Just hide all of that complexity. Yeah,
exactly.
Well, I'm glad you have a Steve Jobs
mode in you as well. This is great.
Einstein mode, Steve Jobs mode. Um all
right, let me try to trick you into
answering a question. When when will
Gemini 3 come out? Is it before after
GTA 6? The world waits for both. And
what does it take
to go from 2.5
to 3.0? Because it seems like there's
been a lot of releases of 2.5 which are
already leaps in performance. So, what
what does it even mean to go to a new
version? Is it about performance? Is it
about a complete different flavor of an
experience?
>> Yeah, well, so the way it works with our
different version numbers is we you
know, we try to collect so maybe it
takes you know, roughly 6 months or
something to to do a new kind of full
run and the full productization of a new
version and during that time lots of new
interesting research iterations and
ideas come up and we sort of collect
them all together that you know, you
could imagine the last 6 months worth of
interesting ideas on the architecture
front. Maybe it's on the data front.
It's like many different possible things
and we collect package that all up test
which ones are likely to be useful for
the next iteration and then bundle that
all together and then we start the new
you know, giant hero training run,
right? And and then
and then of course that gets monitored
and then at the end then there's the of
the pre-training then there's all the
post-training. There's many different
ways of doing that different ways of
patching it. So there's a whole
experimentation phase there which you
can also get a lot of gains out and
that's where you see the version numbers
usually referring to the base model the
pre-trained model and then the interim
versions of 2.5 you know, and the
different sizes and the different little
additions. They're often patches or
post-training ideas that can be done
afterwards off the same basic
architecture and then of course on top
of that we also have different sizes Pro
and flash and flashlight that are often
distilled from the biggest ones you
know, the flash model from the pro model
and that means we have a range of
different choices have you or the
developer of do you want to prioritize
performance or speed right and cost and
we like to think of this Pareto Frontier
well of you know, on the one hand the Y
axis is you know, like performance and
then the the the X axis is you know,
cost or latency and and speed basically
and we we have models that completely
define the frontier. So whatever your
trade-off is that you want as an
individual user or as a as a developer,
you should find one of our models
satisfies that constraint.
>> So, behind the version changes, there is
a big hero run.
>> And then
there's uh just an insane complexity of
productization.
Then there's the distillation of the
different sizes along that Pareto front.
And then as each step you take, you
realize there might be a cool product
and the side quests.
>> Yes, exactly.
>> And then you also don't want to take too
many side quests because then you have a
million versions of million products.
It's very unclear. But you also get
super excited cuz it's super cool. Like
how does even it look at VOs? Very cool.
How does it fit into the bigger thing?
>> exactly. Exactly. And then you're
constantly this process of converging
upstream, we call it, you know, ideas
from the from the product surfaces or or
or from the post training and and even
further downstream than that, you you
kind of upstream that into the the core
model training for the next run. Right?
So, then the main model, the main Gemini
track, becomes more more general. And
eventually, you know, AGI.
>> One hero run at a time.
>> Yes, exactly. Few hero runs later.
Yeah.
>> So, sometimes when you release these new
versions or every version, really,
are benchmarks
productive or counterproductive for
showing the performance of a model?
>> You need them and and but it's important
that you don't overfit to them. Right?
So, they shouldn't be the end and the
be-all and end-all. There's there's
Eleuther AI Arena or it used to be
called Eleuther AI LM-Eval-Harness,
that's one of them that turned out sort
of organically to be one of the the main
ways people like to test these systems,
at least the chatbots.
Obviously, there's loads of academic
benchmarks on from from that test
mathematics and coding ability, general
language ability, science ability, and
so on. And then we have our own internal
benchmarks that we care about. It's a
kind of multi-objective
you know, optimization problem. Right?
You You don't want to be good at just
one thing. We're trying to build general
systems that are good across the board.
And you try and make no regret
improvements. So, where you improve in
like, you know, coding,
but it doesn't reduce your performance
in other areas, right? So, that's the
hard part cuz you you can Of course, you
could put more coding data in or you
could put more
I don't know, gaming data in, but then
does it make worse your language system
or or uh in your translation systems and
other things that you care about. So,
it's you've got to kind of continually
monitor this increasingly larger and
larger suite of of benchmarks. And also
there's uh when you stick them into
products, these models, you also care
about the direct usage and the direct
stats and the signals that you're
getting from the end users, whether
they're coders or or or the average
person using a using the chat
interfaces.
>> Yeah, because ultimately you want to
measure the usefulness, but it's so hard
to convert that into a number.
>> It's It's really vibe-based benchmarks
across a large number of users and it's
hard to know. And I would it would be
just terrifying to me to You know you
have a much smarter model,
but it's just something vibe-based. It's
not not not quite working. That's just
scary cuz and everything you just said,
it has to be smart and useful across so
many domains. So, you you get super
excited cuz it's all of a sudden solving
programming problems it's never been
able to solve before, but now it's
crappy at poetry or something.
>> And it's just I don't know, that's a
stressful. That's so difficult um
>> To balance, yeah.
>> to balance and because you can't really
trust the benchmarks, you really have to
trust the end users.
>> Yeah. And then other things that are
even more esoteric come into play like,
um you know, the style of the persona of
the the the system, you know, how it you
know, is it verbose? Is it succinct? Is
it humorous? You know, and it and
different people like different things.
um you know, it's very interesting. It's
almost like cutting-edge part of
psychology research or personal
personality research. You know, I used
to do that in my PhD like five-factor
personality. What do we actually want
our assistants to be like? And different
people will like different things as
well. So, these are all just sort of new
problems in product space that I don't
think have ever really been tackled
before, but um we're going to sort of
have rapidly have to deal with now.
>> I think it's a super fascinating space
developing the character of the thing.
Yeah. And in so doing, it puts a mirror
to ourselves what are the kind of things
um that we like. Cuz prompt engineering
allows you to control a lot of those
elements, but can the product
uh make it easier for you
to uh control the different flavors of
those experiences, the different
characters that you interact with.
>> Yeah, exactly. So,
>> So, what's the probability of Google
DeepMind winning?
>> Well, I don't see it as sort of winning.
I mean, I think we need to think winning
is the wrong way to look at it given how
important and consequential what it is
we're building. So, funnily enough, I
don't I try not to view it like a game
or competition, even though that's a lot
of my mindset. It's it's about in my
view, all of us have those of us at the
leading edge uh have a responsibility to
um steward this unbelievable technology
that could be used for incredible good,
but also has risks. Um
steward it safely into the world for the
benefit of humanity. That's always um
what I've um
I dreamed about and what we've always
tried to do, and I hope that's what
eventually the community, maybe the
international community will rally
around when it becomes obvious that as
we get closer and closer to to AGI that
um that's what's needed.
>> I agree with you. I think that's
beautifully put. You've said that um you
talk to and are on good terms with
the leads of some of these uh labs. As
the competition heats up,
how hard is it to maintain sort of those
relationships?
>> It's been okay so far. I try to pride
myself in being
collaborative. I'm a collaborative
person. Research is a collaborative
endeavor. Science is a collaborative
endeavor, right? It's all good for
humanity in the end if you cure
incredible, you know, terrible diseases
and you come with an incredible cure.
This is net win for humanity. And the
same with energy, all of the things I'm
interested in in in helping solve with
AI. So, I just want that technology to
exist in the world and be used for the
right things. And and and the the kind
of the benefits of that, the
productivity benefits of that being
shared for every the benefit of
everyone. So, I try to maintain good
relations with all the leading lab uh
people. They've very interesting
characters, many of them, as you might
expect. Um but yeah, I'm on good terms,
I I hope, with pretty much all of them
and uh I I think that's going to be
important when when things get even more
serious than they are now. Uh that there
are those communication channels and uh
that's what will facilitate
uh co- operation or collaboration if
that's what was required, especially on
things like safety.
>> Yeah, I hope there's some collaboration
on stuff that's uh sort of less
high-stakes and in so doing serves as a
mechanism for maintaining friendships
and relationships. So, for example, I
think the internet would love it if you
and Elon somehow collaborate on creating
a video game, that kind of thing. That I
think that enables camaraderie and good
terms. And also, you two are legit
gamers, so it's just fun to
>> fun to play some stuff.
>> would be awesome. We've talked about
that in the past and it may be a cool
thing that that, you know, we can do.
And I agree with you, it'd be nice to
have um kind of side projects in a way
where where we one can just lean into
the collaboration aspect of it and it's
a sort of
uh win-win for both sides and it's um
and it kind of builds up that that that
uh collaborative muscle.
>> I see the scientific endeavor as that
kind of side project for humanity. And I
I think Google DeepMind has been really
pushing that. Uh I would love it if to
see other labs do more scientific stuff
and then collaborate cuz it just seems
like easier to collaborate on the big
scientific questions.
>> I agree and I would love to see a lot of
people A lot of the other labs talk
about science, but I think we're really
the only ones using it for science and
doing that and that's why projects like
AlphaFold are so important to me and I
think to our mission is to show uh how
AI can this you know be clearly used in
a very concrete way for the benefit of
humanity and and also we spun out
companies like Isomorphic off the back
of AlphaFold to do drug discovery and
it's going really well and build sort of
you know you can think of build
additional AlphaFold type type systems
to go into chemistry space to help
accelerate drug design and the examples
I think we need to show uh and society
needs to understand of what AI can bring
these huge benefits.
>> Well, from the bottom of my heart thank
you for pushing the scientific efforts
forward with with rigor, with fun, with
humility, all of that. I just love to
see it and still talking about P equals
NP. I mean it's just incredible. So I
love it. Uh there are there there's been
uh seemingly a war for talent. Some of
it is meme. I don't know.
Um what do you think about Meta buying
up talent with huge salaries and and the
heating up of this battle for talent.
And I should say that I think a lot of
people see DeepMind as a really great
place to do
uh cutting edge work for the reasons
that you've outlined is like there's
this vibrant scientific culture.
>> Yeah, well look I I of course
um you know there's a strategy that that
Meta is taking right now. I think that
um from my perspective at least I think
the people that are real uh believers in
the mission of AGI and what it can do
and understand the real consequences
both good and bad from that and what's
what that responsibility entails. I
think they're mostly doing it to be like
myself to be on the frontier of that
research. So you know they can help
influence the way that goes and steward
that technology safely into the world.
And you know, Meta right now are not at
the frontier, maybe they'll they'll
manage to get back on there. And you
know, it's probably rational what
they're doing from their perspective
because they're behind and they need to
do something. But I think
there's more important things than than
just money. Of course, one has to pay,
you know, people their market rates and
all of these things, and that continues
to go up.
But as and and and I was expecting this
because more and more people are finally
realizing leaders of companies what I've
always known for 30 plus years now,
which is that AGI is the most important
technology probably that's ever going to
be invented. So in some senses it's it's
rational to be doing that. But I also
think there's a much bigger question. I
mean, people in AI these days are very
well paid. You know, I remember when we
were starting out back in 2010, you
know, I didn't even pay myself a couple
of years cuz there was wasn't enough
money. We couldn't raise any money. And
these days interns are being paid, you
know, the amount that we raised as our
first entire seed round. So it's pretty
funny. And I remember the days where we
used I used to have to to work for free
and and almost pay my own way to do an
internship, right? Now it's all the
other way around. But that's just how it
is. It's the new world. And
but I think that you know, we've been
discussing that what happens post AGI
and energy systems are solved and so on.
What does even money going to mean? So I
you know, in the economy and and we're
going to have much bigger issues to work
through and how does the economy
function in that world and companies? So
I think you know, it's a little bit of a
side issue about
salaries and things of like that today.
>> Yeah, when you're facing such gigantic
consequences and and gigantic
fascinating scientific questions.
>> Which may be only a few years away. So
>> So on a practical sort of programmatic
sense, if we zoom in on jobs, we can
look at programmers because it seems
like AI systems are currently doing
incredibly well at programming and
increasingly so. So a lot of people that
uh,
program for a living, and programming,
are worried they will lose their jobs.
How worried should they be, do you
think? And what's the right way to
uh, sort of adjust to the new reality
and ensure that you survive and thrive
as a human in the programming world?
>> Well, it's interesting that programming,
and it's again counterintuitive to what
we thought years ago, maybe, that some
of the skills that we think of as harder
skills are turned out maybe to be the
easier ones for various reasons, but you
know, coding and math because you can
create a lot of synthetic data and
verify if that data's correct. So,
because of that nature of that, it's
easier to make things like synthetic
data to train from.
Um, it's also an area, of course, where
we're all interested in, cuz we're as
programmers, right, to help us and get
faster at it and more productive. So, I
think the for the next era, like the
next 5-10 years, I think what we're
going to find is people who are kind of
embrace these technologies, become
almost at one with them,
um, whether that's in the creative
industries or the technical industries,
will become sort of super-humanly
productive, I think. So, the great
programmers will be even better, but
they'll be even 10x even what they are
today. And because they you'll be able
to use their skills to utilize the the
tools to the maximum,
uh, you know, exploit them to the
maximum. And, um, so I think that's what
we're going to see in the next domain.
Um, so that's going to cause quite a lot
of change, right? And so, that's coming.
A lot of people benefit from that. So, I
think one example of that is if coding
becomes easier, um, it becomes available
to many more creatives to do more,
uh, and, uh, but I think the top
programmers will still have huge
advantages as terms of specifying, going
back to specifying what the architecture
should be, the questions should be, how
to guide these, um,
coding assistants in a way that's
useful, and, you know, check whether the
code they produce is good. So, I think
there's plenty of, um,
uh, head room there for the foreseeable,
you know, next few years. So, I think
there's there's several interesting
things there. One is there's a a lot of
imperative to
just get better and better consistently
of using these tools so that you're
riding the wave of the improvement
improving models versus like competing
against them.
But sadly, but that's the the nature of
of life on Earth.
Um there could be a huge amount of value
to certain kinds of programming at the
cutting edge and less value to other
kinds. For example, it could be like you
know, front end web design might uh be
more amenable to to to as as you
mentioned to
generation
uh by AI systems and maybe for example,
game engine design or something like
this or back end designers or guiding
systems in high-performance situations,
high-performance programming type of
design decisions that might be extremely
valuable. But it will shift where the
humans are needed most and that's scary
for people to adjust.
>> I think that's right. That that anytime
where there's a lot of disruption and
change, you know, we've had this is not
just this time. We've had this in many
times in human history with the
internet, uh mobile, but before that was
the Industrial Revolution.
Um and it's going to be one of those
eras where there will be a lot of
change. I think there'll be new jobs we
can't even imagine today just like the
internet created and then those people
with the right skill sets to ride that
wave will become incredibly uh valuable.
Right, those skills. But maybe people
will have to relearn or adapt a bit uh
their current skills. And it's the the
thing that's going to be harder to deal
with this time around is the I think
what we're going to see is something
like probably 10 times the impact the
Industrial Revolution had and but 10
times faster as well.
Right? So, instead of 100 years, it
takes 10 years. And so, that's going to
make it, you know, it's like a 100x uh
the impact and the speed combined. So,
that's what I think going to make it
more difficult for society to to to deal
with and it's good there's a lot to
think through and I think we need to be
discussing that right now and I I you
know, encourage top economists in the
world and philosophers to start thinking
about um uh how should is society going
to be affected by this and what should
we do including things like um you know,
universal basic provision or something
like that where a lot of the um
increased productivity uh get shared out
and distributed uh to society um and
maybe in the form of surface services
and other things where if you want more
than that, you still go and get some
incredibly rare skills and things like
that um and and make yourself unique um
but uh but there's a basic provision
that is provided.
>> And if you think of government as a
technology, there's also interesting
questions not just in the economics but
just politics.
How do you design a system that's
responding to the rapidly changing times
such that you can represent
the different pain that people feel from
the different groups and how do you
reallocate resources
in a way that um addresses that pain and
represents the hope and the pain and the
fears of different people
uh in a way that doesn't lead to
division cuz
uh politicians are often really good at
sort of fueling the division and using
that to get elected. The other mi
defining the other and then saying
that's bad and sort of based on that I
think that's often counterproductive to
leveraging a rapidly changing technology
how to
help the world flourish. So, we almost
uh need to improve our political systems
as well rapidly. If you think of them as
a technology.
>> Definitely and I think I think we'll
need new governance structures,
institutions probably to help with this
transition. So, I think political
philosophy and political science is
going to be key
to that. But, I think the number one
thing first of all that is to create
more abundance of resources, right? Then
there's the So, that's the number one
thing. Increase productivity, get more
resources, maybe eventually get out of
the zero-sum situation. Then the second
question is how to use those resources
and distribute those resources. But,
yeah, you can't do that without having
that abundance first.
>> Uh you mentioned to me the book The
Maniac by
Benjamin Labatut, a book on uh
first of all, about you. There's a bio
about you.
>> It's strange, yeah.
>> It's unclear Yes, sir.
>> It's unclear how much is fiction, how
much is reality.
Um but, I think the central figure that
is John von Neumann. I would say it's a
haunting and beautiful exploration of
madness and genius and
let's say the double-edged sword of
discovery.
And
you know, for
people don't know, John von Neumann is a
kind of legendary mind. He contributed
to quantum mechanics. He was
on the Manhattan Project. He is widely
considered to be the father of or
pioneered the modern computer and AI and
so on. So, as many people say, he's like
one of the smartest humans ever. So,
it's just fascinating. And what's also
fascinating as as a person who saw
nuclear science and physics become the
atomic bomb. So, you you got to see
ideas become a thing that has a huge
amount of impact on the world. He also
foresaw the same thing for computing.
>> He's He And that's the a little bit
again, beautiful and haunting aspect of
the book.
then taking a leap forward and looking
at the at least at all AlphaZero AlphaGo
AlphaZero big moment that maybe John von
Neumann's
thinking was brought to to to to
reality. So, I I I guess the question is
what do you think if you got to hang out
with John von Neumann now? What what
would he say
about what's going on?
>> Well, that would be an amazing
experience. You know, he's a fantastic
mind and and I also love the way he he
spent a lot of his time at Princeton at
the Institute of Advanced Studies, a
very special place for thinking. And um
it's amazing how much of a polymath he
was and the the spread of things he
helped invent, including of course the
von Neumann architecture that all the
modern computers are based on. And um he
had amazing foresight. I think he would
have loved where we are today and he
would have um I think he would have
really enjoyed AlphaGo being a you know,
games that he also did game theory. I
think he foresaw a lot of what would
happen with learning machines, systems
that that that are kind of grown, I
think he called it rather than
programmed. I'm not sure how even maybe
he wouldn't even be that surprised. This
the fruition of what I think he already
foresaw in the 1950s.
>> I wonder what advice he would give. He
got to see the building of the atomic
bomb with the Manhattan Project. I'm
sure there's interesting stuff that
maybe is not talked about enough, maybe
some bureaucratic aspect, maybe the
influence of politicians, maybe
maybe not enough of picking up the phone
and talking to
people that are called enemies by the
said politicians. There might be some
like deep wisdom that we just may have
lost from that time, actually.
>> Yeah, I'm sure I'm sure there is. I
mean, after we we you know, study I read
a lot of books from that time as well,
chronicle time
and some brilliant people involved. I I
agree with you. I think maybe there
needs to be more dialogue and
understanding.
Um I hope we can learn from those those
times. I think the difference here is
that the AI has so many it's a multi-use
technology. Obviously, we're trying to
do things I like that like solve, you
know, all diseases,
help with energy
and scarcity. These incredible things.
This is why all of us and and myself,
you know, I worked to start the on this
journey 30 plus years ago. And um, but
of course there are risks, too. And
probably von Neumann, my guess is he
foresaw both. And um, and I think he
sort of said, I think it's to his wife
that that that it would be a this is a
computers will be even more impactful in
the world. And as we just discussed, you
know, I think that's right. I think it's
going to be 10 times at least of the
industrial revolution. So, I think he's
right. So, I think he would have been, I
imagine, fascinated by
uh, where we are now.
>> And I think one of the, maybe you can
correct me, but one of the takeaways
from the book
is that reason,
as uh, said in the book, Mad Dreams of
Reason, is not enough for guiding
humanity as we build these super
powerful technology. That there's
something else. I mean, that there's
also like a religious component.
Whatever God, whatever religion gives,
it gives it pulls at something in the
human spirit that raw, cold
reason doesn't give us.
>> And I I agree with that. I think we need
to approach it with, whatever you want
to call it, the a spiritual dimension or
humanist dimension. It doesn't have to
be to do with religion, right? But this
idea of of a soul, what makes us human,
the spark that we have, perhaps it's to
do with consciousness when we finally
understand that.
Um,
I think that has to be at the heart of
the endeavor.
Um, and technology, I've always seen
technology as the enabler, right? The
tools that that that enable us to to
flourish and to understand more about
the the world. And I'm sort of with
Feynman on this, and he used to always
talk about science and art being
companions, right? You can understand it
from both sides. The beauty of a flower,
how beautiful it is, and also understand
why the colors of the flower evolved
like that, right? That just makes it
more beautiful than than than just the
intrinsic beauty of the flower. And and
I've always sort of seen it like that.
And maybe, you know, in the Renaissance
times, the great discoverers then, like
people like Da Vinci, you know, they
were I don't think he saw any difference
between science and art and perhaps
religion, right? They were everything
was it's just part of being human and
being inspired about the world around
us. And that's what I the philosophy I
try to take and one of my favorite
philosophers is Spinoza and I think he
combined that all very well, you know,
this idea of trying to understand the
universe and understanding our place in
it and that was his kind of way of
understanding religion. And I think
that's quite beautiful. And for me every
all of these things are related,
interrelated, the technology and what it
means to be human. And I think it's a
very important though that we remember
that as when we're immersed in the
technology and the the research. I think
a lot of researchers that I see in in
our field are a little bit too narrow
and only understand the technology. And
I think also that's why it's important
for
this to be debated by society at large
and I'm very supportive of things like
this the AI summits that will happen and
governments understanding it. And I
think that's one good thing about the
chatbot era and the product era of AI is
that everyday person can actually feel
and and interact with cutting-edge AI
and and and feel feel it for themselves.
>> Yeah, because they they force the
technologists to have the human
conversation. Yeah, for sure. That's the
hopeful aspect of it like you said it's
a dual-use technology that we're
forcefully integrating the entire
humanity into it by into the discussion
about AI because ultimately AGI will be
used for things that states use
technologies for, which is conflict and
so on. And the more
we
integrate humans into this picture by
having chats with them, the more it will
guide
>> Yeah, be able to adapt society will be
able to adapt to these technologies.
Like we've always done in the past with
with the incredible technologies we've
invented in the past.
>> Do you think there will be something
like a
Manhattan Project
where
um there will be an escalation of the
power of this technology and states in
their old way of thinking will try to
use it as weapons technologies and there
will be this kind of escalation?
>> I hope not.
Um I think that would be
uh very dangerous to do and I think also
you know, not the right use of the
technology. I I hope we'll end up with
more something more collaborative if
needed like more like a like a CERN
project, you know, where
um it's research-focused and the best
minds in the world come together to
carefully complete the final steps and
make sure it's responsibly done before
you know, like uh deploying it to the
world. We'll see. I mean, it's difficult
with the current geopolitical climate, I
think, uh to to see cooperation, but
things can change. And um I think at
least on the scientific level, it's
important for the researchers to to to
to keep in touch and and and keep close
to each other on at least on those kinds
of topics.
>> Yeah, that I personally believe on the
education side and um
immigration side, it would be great if
both directions uh people from the West
immigrated China and China back. I mean,
there is some like family human aspect
of people just intermixing.
>> And thereby those ties grow strong so
you can't sort of divide against each
other this kind of old school way of
thinking. And so multi uh multi cultural
multi disciplinary research teams
working on scientific questions, that's
like the hope. Don't don't let the the
warm leaders that are warmongers because
they divide us. I think science is the
ultimately really beautiful connector.
>> Yeah, science has always been uh I think
quite a a very collaborative endeavor
and you know, scientists know that it's
it's a it's a collective endeavor as
well and we can all learn from each
other. So, perhaps it could be a vector
to get a bit of cooperation.
>> What's your uh ridiculous question?
What's your P doom? Probability the
human civilization destroys itself?
>> Well, look, I I don't have a
>> it's a you know, I don't have a P doom
number. The reason I don't is because I
think it's would imply a level of
precision that is not there. So, like I
don't know how people are getting their
P doom numbers. I think it's a kind of a
little bit of a ridiculous notion
because um what I would say is it's
definitely non-zero, and it's probably
non-negligible. So, that in itself is
pretty sobering. And my my view is it's
just hugely uncertain, right? What these
technologies are going to be able to do,
how fast are they going to take off, how
controllable they're going to be. Some
things may turn out to be, and
hopefully, like way easier than we
thought, right? Um but it may be there's
some really hard um
problems that are harder than we guess
today. And I think uh we don't know that
for sure. And so, in under those
conditions of a lot of uncertainty, but
huge stakes both ways, you know, on the
one hand, we we could solve all
diseases, energy problems, the not the
the the the scarcity problem, and then
travel to the stars and conquer the
stars, and maximum human flourishing. On
the other hand, is this sort of P doom
scenarios. So, given the uncertainty
around it and the importance of it, it's
clear to me the only rational, sensible
approach is to proceed with cautious
optimism. So, we want the outcome we
want the um
uh the benefits, of course, uh and uh
all of the the amazing things that AI
can bring. And actually, I would be
really worried for humanity if I if
given the other challenges that we have,
climate, disease, you know, aging, uh
resources, all of that, if I didn't know
something like AI was coming down the
line, right? How would we solve all
those other problems? I think it's hard.
Um so, I think we've you know, it could
be amazingly transformative for good.
but on the other hand, you know, there
are these risks that we know are there,
but we can't quite quantify. So, the the
best thing to do is to use the
scientific method to do more research to
try and uh more precisely define those
risks and of course address them.
Um and I think that's what we're doing.
I think there probably needs to be uh 10
times more effort on that than there is
now as we're getting closer and closer
to the to the to the AGI line.
>> What would be the source of worry for
you more? Would it be human caused or AI
AGI caused?
Yeah, humans abusing the technology
versus AGI itself through mechanism that
you spoken about which is fascinating
deception or this kind of stuff getting
better and better and better secretly
and then
>> I think they they they operate over
different time scales and they're
equally important to address. So,
there's just the the the the common
gardenal variety of like, you know, bad
actors using new technology uh in this
case general purpose technology and
repurposing it for harmful ends. And
that's a huge uh risk. And I think that
has a lot of complications because
generally, you know, I'm in huge favor
of open science and open source and in
fact we did it with all our science
projects like AlphaFold and all of those
things uh for the benefit of of of the
scientific community. Um but how does
one restrict bad actors access to these
powerful systems whether they're
individuals or even rogue states uh and
but enable access at the same time to
good actors to to maximally build on top
of. It's pretty tricky problem that
there's I've not heard a clear solution
to. So, there's the bad actor use case
problem and then there's obviously uh as
the systems become more agentic and and
closer to AGI um
and more autonomous, how do we ensure
the guardrails and they stick to what we
want them to do uh and under our
control?
>> Yeah, I tend to maybe on my mind is
limited worry more about the humans the
bad actors. And there it could be
in part how do you not put destructive
technology in the hands of bad actors,
but in another part from again
geopolitical technology perspective how
do you reduce the number of bad actors
in the world? That's that's also an
interesting human problem.
>> Yeah, it's a hard problem. I mean look
we we we can
maybe also use the technology itself to
help early
warning on some of the bad actor use
cases, right? Whether that's bio or
nuclear or whatever it is. Like AI could
be potentially helpful there as long as
the AI that you're using is itself
reliable, right? So it's a sort of
interlocking problem and that's what
makes it very tricky and and again it
may require some agreement
internationally at least between China
and the U and and the US of of of some
basic standards, right?
>> I have to ask you about the the book the
maniac. There there's this the the hand
of God moment Lisa Doll's move 78.
That perhaps the last time a human did a
move of sort of pure human genius and
beat AlphaGo or like broke its brain if
sorry to anthropomorphize, but it's an
interesting moment cuz I think in so
many domains it will keep happening.
>> Yeah, it's a special moment and you know
it was great for Lisa Doll and you know
I think it's in a way they were sort of
inspiring each other. We as a team were
inspired by Lisa Doll's brilliance and
nobleness and then maybe he got inspired
by you know what AlphaGo was doing to
then conjure this incredible
inspirational moment. It's all you know
captured very well in the in the
documentary about it and
I think that will continue in many
domains where there's this at least for
the for the again for the foreseeable
future of like the humans bringing in
their ingenuity
um, and asking the right question, let's
say,
uh, and then utilizing these tools,
uh, in a way that, um, then cracks a
problem.
>> Yeah, what as the AI becomes smarter and
smarter,
one of the interesting questions we can
ask ourselves is what makes humans
special.
It does feel, um, perhaps biased that we
humans are
deeply special.
I don't know if it's our intelligence.
It could be something else, that that
other thing that's outside the mad
dreams of reason.
>> I think that's what I've always
imagined, uh, when I was a kid and
starting on this journey of like, um, I
was, of course, fascinated by things
like consciousness, did did a
neuroscience PhD to look at how the
brain works, especially imagination and
memory. I focused on the hippocampus.
And it's sort of going to be
interesting. I always thought the best
way, of course, one can can philosophize
about it and have thought experiments
and maybe even do actual experiments
like you do in neuroscience on on real
brains. But in the end, I always
imagined that building AI, a kind of
intelligent artifact, and then comparing
that to the human mind and seeing what
the differences were,
uh, would be the best way to uncover
what's special about the human mind, if
indeed there is anything special. And I
suspect there probably is, but it's
going to be hard to you know, I think
this journey we're on will help us, uh,
understand that and define that. And you
know, there may be a difference between
carbon-based substrates that we are and
silicon ones when they process
information. You know, one of the best
definitions I like of of of
consciousness is it's the way
information feels when we process it,
right?
>> Um, it could be. I mean, it doesn't have
it's not a very helpful scientific
explanation, but I think it's kind of
interesting intuitive one. And, um, and
so, you know, on this this this journey,
this scientific journey we're on will, I
think, um, help uncover that mystery.
>> Yeah, what I cannot create I do not
understand. That's uh, somebody you
deeply admire, Richard Feynman, like you
mentioned. You also reach
for the the Wigner's dreams of
universality that he saw in constraint
domains, but also broadly, generally in
in mathematics and so on. So, so many
aspects on which you're pushing towards
not to start trouble at the end, but uh
Roger Penrose.
uh you know, do you do you think
consciousness does this hard problem of
consciousness, how information feels?
do you think consciousness first of all
is a computation?
And if it is, if it's information
processing, like you said, everything
is, is it something that could be
modeled by a classical computer?
>> Or is it a quantum mechanical nature?
>> Well, look, I Penrose is an amazing
thinker, one of the greatest of the
modern era, and he we've had a lot of
discussions about this. Of course, we
cordially disagree, which is, you know,
I I I feel like um I mean, he
collaborates with a lot of good
neuroscientists to see if he could find
mechanisms for quantum mechanics
behavior in the brain, and they to my
knowledge, they haven't found anything
um convincing yet. So, my betting is
there is is that that that it's mostly,
you know, it is just classical computing
that's going on in the brain, which
suggests that all the phenomena uh are
modelable
or mimicable by uh classical computer.
But, we'll see, you know, there there
may be this final mysterious things of
the feeling of consciousness, the
qualia, these kinds of things that
philosophers debate where it's unique to
the substrate. We may even come towards
understanding that when if we do things
like neural link and and uh have neural
interfaces to the AI systems, which I
think we probably will eventually, um
maybe to keep up with the AI systems, uh
we might actually be able to feel for
ourselves what it's like to compute on
silicon, right? So, um and maybe that
will tell us.
Uh so, I think it's it's going to be
interesting. And I had a debate once
with the late Daniel Dennett about why
do we think each other are conscious.
Okay, so it's for two reasons. One is
you're exhibiting the same behavior that
I am. So, that's one thing. Behaviorally
you seem like a conscious being if I am.
But, the second thing which is often
overlooked is that we're all running on
the same substrate. So, if you're
behaving in the same way and we're
running on the same substrate, it's most
parsimonious to assume you're feeling
the same experience that I'm feeling.
But, with a AI
uh that's on silicon, we won't be able
to rely on the second part. Even if it
exhibits the first part, that behavior
looks like a behavior of a conscious
being. It might even claim it is.
Um but, we but but we wouldn't know how
it actually felt.
Um and it probably couldn't know we what
we felt. At least in the first stages,
maybe when we get to superintelligence
and the technologies that builds,
perhaps we'll we'll be able to um bridge
that.
>> Yeah, I mean that's a huge test for
radical empathy. It's to empathize with
a different substrate.
>> Exactly. We never had to confront that
before.
>> Yeah, to maybe maybe through
brain-computer interfaces be able to
truly empathize what it feels like to be
a computer.
>> Well, for information to be computed not
on a carbon system.
>> I mean that's deeply exciting. I mean,
some people kind of think about that
with plants, with other life forms,
which is different.
Similar substrate, but
sufficiently far enough on the
evolutionary tree that it's requires a
radical empathy. But, to do that with a
computer
>> I mean, we sort of there are animal
studies on this of like of course higher
animals like, you know, killer whales
and dolphins and dogs and and monkeys,
you know, they have some and elephants,
you know, they have some aspects
certainly of consciousness, right? Even
though they're not might not be that
that that smart on an IQ sense. So, so
we can already empathize with that and
maybe even some of our systems one day
like we built this thing called dolphin
gemma, you know, which can one of our
version of our system was trained on
dolphin and whale sounds and maybe we'll
be able to build a an interpreter or
translator at some point. Should be
pretty cool.
>> What gives you hope for the future of
human civilization?
>> Well, what gives me hope is that I think
our almost limitless ingenuity, first of
all. I think the best of us and the best
human minds are incredible.
Um and you know, I love
you know, meeting and watching
any human that's the top of their game,
whether that's sport or science or art.
You know, it's it's it's just nothing
more wonderful than that, seeing them in
their element in flow. Um I think it's
almost limitless. You know, our brains
are
general systems, intelligent systems.
So, I I think it's almost limitless what
we can potentially do with them. And
then the other thing is our extreme
adaptability. I think it's going to be
okay in terms of there's going to be a
lot of change, but that but look where
we are now without effectively our
hunter-gatherer brains. How is it we
can, you know, we can cope with the
modern world, right? Flying on planes,
doing podcasts, you know, playing
computer games in virtual simulations. I
mean, it's already mind-blowing given
that our mind was was developed for, you
know, hunting buffaloes on the on the
tundra. And and so,
I think this is just the next step and
and and it's actually kind of
interesting to see how
societies already adapted to this
mind-blowing AI technology we have today
already. It's sort of like, "Oh, I talk
to chatbots. Totally fine."
>> And it's very possible that this very
podcast activity which I'm here for will
be completely replaced by AI.
>> I'm very replaceable and I'm waiting for
it.
>> that you can do it, Lex. I don't think.
>> I thank you. That's
>> That's what we humans do to each other.
We compliment.
All right. And I'm deeply grateful for
us humans to have this infinite capacity
for curiosity, adaptability, like you
said, and also compassion and ability to
love.
>> All of those human things.
>> things that are deeply human.
>> Well, this is a huge honor, Demis.
You're one of the truly special humans
in the world. Thank you so much for
doing what you do and for talking today.
>> Well, thank you very much, Lex.
>> Thanks for listening to this
conversation with Demis Hassabis. To
support this podcast, please check out
our sponsors in the description and
consider subscribing to this channel.
And now, let me answer some questions
and try to articulate some things I've
been thinking about. If you'd like to
submit questions, including in audio and
video form, go to lexfridman.com/ama.
I got a lot of amazing questions,
thoughts, and requests from folks. I'll
keep trying to pick some randomly and
comment on it at the end of every
episode.
I got a note on May 21st this year that
said, "Hi, Lex. 20 years ago today,
David Foster Wallace delivered his
famous This Is Water speech at uh Kenyon
College.
What do you think of this speech?"
Well, first, I think this is probably
one of the greatest and most unique
commencement speeches ever given.
But of course, I have many favorites,
including the one by Steve Jobs.
And David Foster Wallace is one of my
favorite
writers and one of my favorite humans.
There's um
tragic honesty to his work, and it
always felt as if he was engaging in a
constant battle with his own mind.
And the writing, his writing,
were kind of his notes from the front
lines of that battle.
Now, onto the speech. Let me quote some
parts. There's of course the parable of
the fish and the water that goes,
"There are these two young fish
swimming along and they happen to meet
an older fish
swimming the other way
who nods at them and says, 'Morning,
boys. How's the water?'
And the two young fish swim on for a
bit, and then eventually, one of them
looks over at the other and goes,
"What the hell is water?"
In the speech, David Foster Wallace goes
on to say, "The point of the fish story
is merely that the most obvious,
important realities are often the ones
that are hardest to see and talk about."
Stated as an English sentence, of
course, this is just a banal platitude.
But the fact is that in the day-to-day
trenches of adult existence, banal
platitudes can have a life or death
importance.
Or so I wish to suggest to you in this
dry and lovely morning.
I have several takeaways from this
parable and the speech that follows.
First,
I think we must question everything, and
in particular the most basic assumptions
about our reality, our life, and the
very nature of existence.
And that this project is a deeply
personal one.
In some fundamental sense, nobody can
really help you in this process of
The call to action here, I think, from
uh David Foster Wallace, as he puts it,
is to quote, "To be just a little less
arrogant. To have just a little more
critical awareness about myself and my
certainties."
Because a huge percentage of the stuff
that I tend to be automatically certain
of is, it turns out, totally wrong and
deluded.
All right, back to me. Lex speaking.
Second takeaway is that the central
spiritual battles of our life are not
fought on a
mountaintop somewhere at a meditation
retreat,
but it is fought in the mundane moments
of daily life.
Third takeaway is that we too easily
give away our time and attention to the
multitude of distractions that the world
feeds us.
The insatiable black holes
of attention.
David Foster Wallace's call to action in
this case is to be deeply aware of the
beauty in each moment
and to find meaning in the mundane.
I often quote David Foster Wallace in
his advice that the key to life is to be
unborable.
And I think this is exactly right.
Every moment, every object, every
experience
when looked at closely enough contains
within it infinite richness to explore.
And since Demis Hassabis of this very
podcast episode and I are such fans of
Richard Feynman, allow me to also quote
Mr. Feynman on this topic as well.
Quote,
I have a friend who's an artist
and has sometimes taken a view which I
don't agree with very well.
He'll hold up a flower and say, "Look
how beautiful it is."
And I'll agree.
Then he says,
"I as an artist can see how beautiful
this is, but you as a scientist take
this all apart and it becomes a dull
thing."
And I think that's kind of nutty.
First of all, the beauty that he sees is
available to other people and to me too,
I believe.
Although I may not be quite as refined
aesthetically as he is, I can appreciate
the beauty of a flower.
At the same time, I see much more about
the flower than he sees.
I can imagine the cells in there,
the complicated actions inside which
also have beauty.
I mean, it's not just beauty at this
dimension, at 1 cm. There's also beauty
at the smaller dimensions. Their inner
structure, also the processes. The fact
that the colors in the flower evolved in
order to attract insects to pollinate it
is interesting. It means that the
insects can see the color.
It adds the question, does this
aesthetic sense also exist in lower
forms? Why is it aesthetic? All kinds of
interesting questions which the science
knowledge only adds to the excitement,
the mystery, and the awe of a flower.
It only adds.
All right, back to David Foster
Wallace's speech. He has a great story
in there
that I particularly
enjoy.
It goes,
there are these two guys sitting
together in a bar in the remote Alaskan
wilderness.
One of the guys is religious, the other
is an atheist. And the two are arguing
about the existence of God with that
special intensity that comes after about
the fourth beer. And the atheist says,
"Look, it's not like I don't have actual
reasons for not believing in God. It's
not like I haven't ever experimented
with the whole God and prayer thing.
Just last month I got caught away from
the camp in that terrible blizzard.
And I was totally lost, and I couldn't
see a thing, and it was 50 below. So, I
tried it. I fell to my knees in the snow
and cried out, 'Oh God, if there is a
God, I'm lost in this blizzard and I'm
going to die if you don't help me.'"
And now back in the bar, the religious
guy looks at the atheist all puzzled.
"Well, then you must believe now," he
says. "After all, there you are, alive."
The atheist just rolls his eyes. "No,
man. All that happened was a couple of
Eskimos happened to be wondering by and
showed me the way back to the camp."
All this I think teaches us that
everything is a matter of perspective,
and that wisdom may arrive if we have
the humility to keep shifting and
expanding our perspective on the world.
Thank you for allowing me to talk a bit
about David Foster Wallace.
He's one of my favorite writers and
he's a beautiful soul.
If I may, one more thing I wanted to
briefly comment on.
I found myself to be in this strange
position
of getting attacked online often from
all sides,
including being lied about, sometimes
through selective misrepresentation,
but often through downright lies.
I don't know how else to put it.
This all breaks my heart, frankly.
But, I've come to understand that it's
the way of the internet and the cost of
the path I've chosen.
There's been days when it's been rough
on me mentally.
It's not fun being lied about,
especially when it's about things that
are usually
for a long time have been a source of
happiness and joy for me. But again,
that's life.
I'll continue exploring the world of
people and ideas with empathy and rigor,
wearing my heart on my sleeve
as much as I can.
For me, that's the only way to live.
Anyway, a common attack on me is about
my time at MIT and Drexel, two great
universities I love and have tremendous
respect for.
Since a bunch of lies have accumulated
online about me on these topics to a sad
and at times hilarious degree, I thought
I would once more state the obvious
facts about my bio for the small number
of you who may care.
TLDR, two things. First, as I say often,
including in a recent podcast episode
that somehow was listened to by many
millions of people,
I proudly went to Drexel University for
my bachelor's, master's, and doctorate
degrees.
Second, I am a research scientist at MIT
and have been there in a paid research
position for the last 10 years.
Allow me to elaborate a bit more on
these two things now, but please skip if
this is not at all interesting.
So, like I said, a common attack on me
is that I have no real affiliation with
MIT.
The accusation, I guess, is that I'm
falsely claiming an MIT affiliation
because I taught a lecture there once.
Nope. That accusation against me is a
complete lie.
I have been at MIT for over 10 years in
a paid research position from 2015
to today.
To be extra clear, I'm a research
scientist at MIT working in LIDS, the
Laboratory for Information and Decision
Systems in the College of Computing.
For now, since I'm still at MIT, you can
uh see me in the directory and on the
various lab pages.
I have indeed given many lectures at MIT
over the years, a small fraction of
which I posted online.
Teaching for me always has been just for
fun and not part of my research work.
I personally think I suck at it, but I
have always learned and grown from the
experience. It's like Feynman spoke
about, if you want to understand
something deeply,
it's good to try to teach it.
But, like I said, my main focus has
always been on research. I published
many peer-reviewed papers that you can
see in my Google Scholar profile.
For my first 4 years at MIT, I worked
extremely intensively. Most weeks were
After that, in 2019, I still kept my
research scientist position, but I split
my time taking a leave to pursue
projects in AI and robotics outside MIT
and to dedicate a lot of focus to the
As I've said, I've been continuously
surprised just how many hours preparing
for an episode takes. There are many
episodes of the podcast for which I have
to read, write, and think for 100, 200,
or more hours across multiple weeks and
months.
Since 2020,
I have not actively published research
papers.
Just like the podcast, I think it's
something that's a serious full-time
effort.
But, not publishing and doing full-time
research has been eating at me
because I love research and I love
programming and building systems that
test out interesting technical ideas,
especially in the context of human AI or
human-robot interaction.
I hope to change this in the coming
months and years.
What I've come to realize about myself
is if I don't publish or if I don't
launch systems that people use, I
definitely feel like a piece of me is
missing.
It legitimately is a source of happiness
Anyway, I'm proud of my time at MIT. I
was and am constantly surrounded by
people much smarter than me, many of
whom have become lifelong colleagues and
friends.
MIT is a place I go to escape the world,
to focus on exploring fascinating
questions at the cutting edge of science
and engineering.
This again makes me truly happy.
And it does hit pretty hard on a
psychological level when I'm getting
attacked over this.
Perhaps I'm doing something wrong.
If I am, I will try to do better.
In all this discussion of academic work,
I hope you know that I don't ever mean
to say that I'm an expert at anything.
In the podcast and in my private life,
I don't claim to be smart. In fact, I
often call myself
an idiot and mean it.
I try to make fun of myself as much as
possible and in general to celebrate
others instead.
Now, to talk about Drexel University,
which I also love, am proud of, and am
deeply grateful for my time there.
As I said, I went to Drexel for my
bachelor's, master's, and doctor degrees
in computer science and electrical
engineering.
I've talked about Drexel many times,
including, as I mentioned, at the end of
a recent podcast,
the Donald Trump episode, funny enough,
that was listened to by many millions of
people,
where I answered a question about
graduate school and explained my own
journey at Drexel and how grateful I am
for it.
If it's at all interesting to you,
please go listen to the end of that
episode or watch the related clip.
At Drexel, I met and worked with many
brilliant researchers and mentors from
whom I've learned a lot about
engineering, science, and life.
There are many valuable things I gained
from my time at Drexel.
First, I took a large number of very
difficult math and theoretical computer
science courses.
They taught me how to think deeply and
rigorously and also how to work hard and
not give up even if it feels like I'm
too dumb to find a solution to a
technical problem.
Second, I programmed a lot during that
time, mostly C, C++.
I programmed robots,
optimization algorithms, computer vision
systems, wireless network protocols,
multimodal machine learning systems, and
all kinds of simulations of physical
systems.
This is where I really developed a love
for programming,
including,
yes, Emacs and the Kinesis keyboard.
Uh I also during that time read a lot. I
played a lot of guitar, wrote a lot of
crappy poetry, and uh trained a lot of
uh in Judo and Jiu-Jitsu,
which I cannot sing enough praises to.
Jiu-Jitsu humbled me on a daily basis
throughout my 20s, and it still does to
this very day whenever I get a chance to
train.
Anyway, I hope that the folks who
occasionally get swept up in the
chanting online crowds that want to tear
down others don't lose themselves in it
too much.
In the end, I still think there's more
good than bad in people.
But, we're all, each of us, a mixed bag.
I know I am very much flawed. I speak
awkwardly. I sometimes say stupid
I can get irrationally emotional. I can
be too much of a dick when I should be
kind.
I can lose myself in a biased rabbit
hole before I wake up to the bigger,
more accurate picture of reality.
I'm human,
and so are you,
for better or for worse.
And I do still believe we're in this
whole beautiful mess together.
I love you all.
[music]
