What this covers

Demis Hassabis and Lex Fridman explore how natural systems become learnable by artificial intelligence, moving from the specifics of recent DeepMind models to the foundations of AGI and deep physics. The conversation centers on a counterintuitive claim: that evolution and physical selection impose structure on nature, placing complex systems on low-dimensional manifolds that classical neural networks can efficiently discover. Hassabis walks through concrete examples—AlphaFold modeling protein structures, Veo 3 learning fluid dynamics and lighting from passive video observation, WeatherNet outperforming traditional simulation—to show that understanding can emerge without embodied interaction or explicit physics engines. He positions this efficiency not as a limitation but as a window into how AGI might work and what it could reveal about fundamental questions in mathematics and biology.

The discussion spans the practical and the metaphysical. On technique, Hassabis distinguishes between scaling current approaches and the harder problem of research taste—the capacity to pose the right question, which he argues remains the frontier that separates breakthrough discovery from incremental improvement. He examines Model Plus Search as a mechanism for genuine novelty, using the famed AlphaGo "move 37" as an exemplar of how search can push beyond training data. On risk, he outlines near-term and long-term threats separately: bad actors repurposing general tools versus increasingly agentic autonomous systems. Throughout, Hassabis argues that classical computation on classical hardware can accomplish far more than previously thought, reframing the P vs NP question itself as a physics question about the universe as an informational system. The conversation touches on consciousness, the boundary between living and non-living matter, and how human productivity might transform over the next decade as tools become tighter extensions of human intent.

Sharpest takeaway

Hassabis argues that most natural systems are efficiently learnable by classical neural networks because evolution and physical selection imbue them with discoverable structure, and that this insight—rather than brute force or quantum computation—is the path toward AGI and answering deep questions about reality.

  • Nature's systems have structure shaped by selection pressures, so they sit on low-dimensional manifolds a classical learner can recover
  • Veo 3 modeling fluids and lighting from passive video shows understanding can emerge without embodiment or explicit physics engines
  • AGI built on neural networks atop classical computers could be the ultimate tool to answer questions like P vs NP and the origin of life

The argument · threads11 threads · 45 claims
0.85

AGI is probable within five years and verifiable through rigorous testing and lighthouse moments of genuine invention.

8 pointscentrality 5/5
  • Current AI systems can do incremental hill-climbing along an S-curve but no one has unequivocally shown they can make a big leap like inventing the Transformer architecture; whether scaling existing approaches is sufficient for AGI or one or two more breakthroughs are needed is roughly a 50/50 open question, which is why DeepMind devotes about half its resources to blue-sky ideas and half to scaling current capabilities.

    “we don’t know, I would say it’s kind of 50/50 whether new things are needed or whether the scaling the existing stuff is going to be enough.”

  • To verify true AGI, run two complementary tests: a brute-force battery of tens of thousands of cognitive tasks plus a month or two of probing by a few hundred world-leading experts to check for obvious flaws and consistency; and look for 'lighthouse moments'—inventing a new conjecture like Einstein's relativity (testable via a pre-1900 knowledge cutoff back-test) or inventing a game as deep and elegant as Go, with a true AGI able to do several such things across domains.

    “make the system available to a few hundred of the world’s top experts, the Terence Taos of each subject area and 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 can be pretty confident we have a fully general system.”

  • There is roughly a 50% chance of achieving AGI within the next five years (by 2030), where AGI is defined by a high bar of matching the brain's general cognitive functions consistently across all domains—not today's jagged, inconsistent intelligence—plus possessing genuine invention and creativity capabilities that current systems lack.

    “My estimate is sort of 50% chance by in the next five years, so by 2030 let’s say.”

  • Research taste or judgment—picking the right question, the right experiment, the right hypothesis—is what separates great scientists from merely good ones and is the hardest thing for AI to mimic; coming up with a deep, falsifiable conjecture is far harder than solving one, and today's systems cannot do it because that creative leap of imagination (like Einstein's relativity) is not achievable by naive search on top of a model.

    “I think that’s going to be one of the hardest things to mimic or model is this idea of taste or judgment. I think that’s what separates the great scientists from the good scientists.”

  • Fully autonomous end-to-end recursive self-improvement is technically conceivable but may not be desirable because it implies a hard-takeoff scenario; current systems like AlphaEvolve keep a human in the loop and only do incremental hill-climbing on narrowly specified objectives, and cannot yet make architectural leaps like inventing the Transformer from a vague high-level instruction such as 'make a better version of yourself.'

    “I would say I’m not sure it’s even desirable because that’s a kind of hard takeoff scenario. But these current systems like Alpha Evolve, they have human in the loop deciding on various things”

  • Counterintuitively, coding and mathematics—skills humans consider hard—turn out to be among the easier domains for AI because you can generate large amounts of synthetic data and verify whether that data is correct, making it easier to train these systems than for fuzzier domains.

    “coding and maths, 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.”

  • Benchmarks are necessary but should not be the be-all-and-end-all; building general systems is a multi-objective optimization where you seek no-regret improvements (gains in one area like coding that don't degrade others like translation or language), and ultimately usefulness must be measured by hard-to-quantify end-user signals and vibes rather than benchmark numbers alone.

    “You need them, but it’s important that you don’t overfit to them. So they shouldn’t be the be all and end all.”

  • There are three forms of scaling happening concurrently—pre-training, post-training, and inference/test-time compute (the new 'thinking' paradigm where systems get smarter the more inference time they're given)—and there is substantial room left in all of them, with inference compute now potentially becoming a larger demand than training because AI is used by billions in products.

    “there’s sort of three scalings that are happening concurrently”

0.74

Neural networks on classical computers can exceed prior belief in what they model, from protein structures to game mastery.

5 pointscentrality 5/5
  • Contrary to longstanding neuroscience 'action in perception' theories that you must act in the world (via embodiment, robotics, or simulated action) to truly perceive intuitive physics, systems like Veo 3 show intuitive physics can be learned through passive observation alone.

    “there’s a lot of theories in neuroscience, it’s called action in perception where you need to act in the world to really, truly perceive it in a deep way... But it seems like you can understand it through passive observation, which is pretty surprising to me.”

  • Veo 3 demonstrates a form of understanding—not deep philosophical understanding but intuitive physics akin to a human child's—by accurately predicting coherent next video frames including liquids, materials, and specular lighting, reverse-engineered purely from watching videos, which challenges the theory that physical understanding requires embodied interaction with the world.

    “there’s a lot of theories in neuroscience, it’s called action in perception where you need to act in the world to really, truly perceive it in a deep way... But it seems like you can understand it through passive observation, which is pretty surprising to me.”

  • DeepMind's WeatherNet system produces the best weather predictions in the world, outperforming traditional fluid-dynamics models that require massive supercomputers and days of calculation, and has recently extended to cyclone/hurricane path prediction—demonstrating that even near-chaotic real-world systems can be efficiently modeled by neural networks.

    “we’ve created 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.”

  • Classical systems (Turing machines running neural networks) can go much further than previously believed—modeling protein structures and beating world champions at Go, feats many thought decades away or requiring quantum computers—and AGI built as neural networks on classical computers would be the ultimate expression of this, directly informing the P vs NP question.

    “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. They can do things like model the structures of proteins and play go to better than world champion level.”

  • Any pattern generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm, because natural systems have structure imposed by evolutionary and selection processes, which creates a low-dimensional manifold a neural network can learn rather than requiring intractable brute-force enumeration.

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

0.66

Search methods combined with foundation models generate genuine novelty beyond passive pattern recognition.

3 pointscentrality 5/5
  • The AlphaGo 'move 37'—a never-before-seen strategy found by Monte Carlo Tree Search—exemplifies how a model that captures all known data can still go beyond it through a search process, which is the core mechanism needed for genuine creativity and scientific discovery, distinct from merely modeling existing knowledge.

    “With AlphaGo, we just use Monte Carlo Tree Search and that’s what found move 37, the new never seen before strategy in Go. And so that’s how you can go beyond potentially what is already known.”

  • AlphaEvolve is a promising hybrid system that uses LLMs to propose candidate solutions and evolutionary computing on top to find novel regions of the search space; combining foundation models with search/reasoning algorithms (evolutionary methods, Monte Carlo tree search) is a fertile direction because the model captures known dynamics while search pushes beyond them toward genuine novelty.

    “LLMs are kind of proposing some possible solutions and then you use evolutionary computing on top to find some novel part of the search space.”

  • Traditional evolutionary computing of the 1990s and early 2000s, despite promising results, could never evolve genuinely new emergent properties—you only ever got a subset of the properties you put in—but combining these methods with modern foundation models may overcome that limitation, mirroring how natural evolution clearly did generate new capabilities from bacteria onward.

    “LLMs are kind of proposing some possible solutions and then you use evolutionary computing on top to find some novel part of the search space.”

0.54

Human-AI collaboration will produce mutual inspiration and superhuman productivity, with humans guiding high-level architecture and creativity.

7 pointscentrality 3/5
  • For the foreseeable future, breakthroughs will come from humans bringing ingenuity and asking the right questions while utilizing AI tools, exemplified by Lee Sedol's move 78 (the 'hand of God') inspired by competing with AlphaGo—a mutual inspiration pattern likely to recur across many domains as humans and AI inspire each other.

    “that’ll continue in many domains where there’s this, at least again for the foreseeable future, of the humans bringing in their ingenuity and asking the right question, let’s say, and then utilizing these tools in a way that then cracks a problem.”

  • Designing AI-first products requires designing not for what the technology can do today but for what it will be able to do in a year, so a product person must be deeply technical—having good intuition for whether the research track is on schedule to intercept the desired capability in six months to a year—and should continually simplify the interface to get out of the way of the rapidly improving model.

    “you’ve 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”

  • Over the next five to ten years, people who embrace AI tools and become 'at one with them' will become superhumanly productive—the best programmers becoming 10x what they are today—while value shifts toward high-level skills like specifying architecture and guiding/checking coding assistants, with some domains (e.g. front-end design) more automatable than others (e.g. high-performance or game-engine design).

    “people who embrace these technologies become almost at one with them... will become superhumanly productive, I think. So the great programmers will be even better, but there’ll be even 10X even what they are today.”

  • Future interfaces will be AI-generated and personalized to each user's aesthetic and cognitive style, dynamically adapting to the task, because text-box chat is a low-bandwidth interaction; new input/output devices like smart glasses, earbuds, and eventually neural interfaces could increase bandwidth roughly 100x over today's typing.

    “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”

  • Google DeepMind's success in shipping relentlessly despite Google's size comes from continuing to operate with startup decisiveness and energy while continually cutting away the bureaucracy and management layers natural to big companies, combining world-class research with the rare ability to deploy to billions of users overnight.

    “I still operate and I was always operating with old DeepMind as a start-up still. A large one, but still as a start-up.”

  • Within five to ten years, AI systems will be able to dynamically generate game content and narrative on the fly around a player's imagination, overcoming the historic fragility and cost limits of hand-coded open-world games and cellular-automata emergence, creating the ultimate personalized choose-your-own-adventure where the player genuinely co-creates the story rather than receiving an illusion of choice.

    “Now we are maybe on the cusp in the next few years, five, 10 years of having AI systems that can truly create around your imagination, can dynamically change the story and storytell the narrative around and make it dramatic no matter what you end up choosing.”

  • AI is the most important technology probably ever going to be invented, which Hassabis says he has always known for 30-plus years.

    “AGI is the most important technology probably that’s ever going to be invented.”

0.53

AI will arrive ten times faster than the Industrial Revolution with ten times greater impact, requiring new governance and abundance-first economics.

3 pointscentrality 4/5
  • If the energy problem is solved via fusion or very efficient solar with better batteries and transmission, a cascade of other problems dissolves: water access becomes solvable through cheap desalination, seawater can be split into hydrogen and oxygen for unlimited rocket fuel enabling bus-service access to space and asteroid mining, ushering in a non-zero-sum era of radical abundance where humanity is no longer resource-constrained.

    “if energy is kind of free and renewable and clean, then that solves a whole bunch of other problems... the water access problem goes away because you can just use desalination.”

  • AI will likely have about ten times the impact of the Industrial Revolution and arrive roughly ten times faster—compressing what took a century into about a decade—making it harder for society to adapt and necessitating that economists and philosophers begin now to plan structures like universal basic provision to share the productivity gains.

    “I think what we’re going to see is something like probably 10 times the impact the Industrial Revolution had, but 10 times faster as well. So instead of a 100 years, it takes 10 years”

  • Managing the AI transition will require new governance structures and institutions informed by political philosophy and science, but the number-one priority is first creating abundance of resources by increasing productivity to escape the zero-sum situation; only then can the second question—how to fairly distribute those resources—be addressed.

    “I think the number one thing, first of all is to create more abundance of resources... 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.”

0.53

International collaboration under a CERN-like model is preferable to competitive state weaponization of AI.

3 pointscentrality 4/5
  • An escalating Manhattan-Project-style weaponization of AI by states would be very dangerous and the wrong use of the technology; a more collaborative, research-focused, CERN-like model—where the best minds come together to carefully complete the final steps and ensure responsible deployment—is preferable, and even amid a difficult geopolitical climate researchers should maintain cross-border contact, ideally with basic standards agreed between China and the US.

    “I hope we’ll end up with something more collaborative if needed, more like a CERN project where 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 deploying it to the world.”

  • Framing AI development as 'winning' a competition is the wrong way to look at it; those at the leading edge have a responsibility to steward this consequential technology safely into the world for the benefit of humanity, and the international community should rally around this as AGI approaches.

    “I think winning is the wrong way to look at it given how important and consequential what it is we’re building.”

  • Meta's strategy of buying talent with huge salaries is rational because they are behind and not currently at the research frontier, but true believers in the AGI mission are mostly motivated by being on the frontier to influence and steward the technology rather than by money, and post-AGI the very meaning of money is an open question.

    “Meta right now are not at the frontier. Maybe they’ll manage to get back on there and it’s probably rational what they’re doing from their perspective because they’re behind and they need to do something.”

0.53

AI poses non-negligible existential risk and near-term dual-use dangers, requiring ten times more safety research.

3 pointscentrality 4/5
  • Assigning a precise P-Doom number is a somewhat ridiculous false-precision exercise, but the probability that AI poses an existential risk is definitely non-zero and probably non-negligible, which combined with huge uncertainty about capabilities, takeoff speed, and controllability—and enormous upside in solving disease, energy, and scarcity—makes cautious optimism the only rational approach, requiring roughly ten times more safety research as AGI nears.

    “I don’t have a P-Doom number. The reason I don’t is because I think it would imply a level of precision that is not there.”

  • AI poses two equally important risks on different timescales: the near-term problem of bad actors (individuals or rogue states) repurposing general-purpose technology for harm, which creates a hard tension with open science since you must restrict bad actors while enabling good actors; and the longer-term problem of increasingly agentic, autonomous systems requiring guardrails to keep them under control.

    “how does one restrict bad actors access to these powerful systems, whether they’re individuals or even rogue states, but enable access at the same time to good actors to maximally build on top of? It’s pretty tricky problem that I’ve not heard a clear solution to.”

  • The great filters that make complex life rare are most likely in the past rather than ahead of us; having any life at all is highly unlikely, and the jump from single-cell to multi-cell life took roughly a billion years on Earth (only after bacteria captured mitochondria), showing how improbably difficult that transition was.

    “I think they’re most likely in the past, if you read that book of how unlikely to go have any life at all. And then single cell to multi-cell seems an unbelievably big jump that took a billion years”

0.35

Consciousness is likely classical computation and could be mimicked by machines, though substrate verification remains uniquely hard.

2 pointscentrality 3/5
  • Consciousness is probably mostly classical computation in the brain (no convincing quantum mechanism has been found, contra Penrose), implying it could be mimicked by classical computers; but verifying machine consciousness is uniquely hard because we infer each other's consciousness from both shared behavior and shared substrate, and with silicon AI we lose the substrate cue—demanding a radical empathy for a different substrate that humanity has never had to confront, possibly bridgeable via neural interfaces.

    “my betting is that it’s mostly it is just classical computing that’s going on in the brain, which suggests that all the phenomena are modelable or mimicable by a classical computer.”

  • Information is primary and more fundamental than energy and matter; energy, matter, and information can all be converted into each other, but the universe is best understood as an informational system, which reframes the P vs NP question as a physics question.

    “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.”

0.33

Humanity's adaptability, ingenuity, and capacity for meaning through mastery and games will enable survival of AI transition.

5 pointscentrality 2/5
  • It is remarkable and troubling that the average person does not worry more about fundamental unanswered questions—the definition of life versus non-living, the nature of time, consciousness, gravity, and quantum weirdness—given that we can inspect living things down to the atomic level yet still cannot define life in a simple clear way; humanity has developed coping mechanisms and distractions to get on with daily life despite these mysteries.

    “I don’t really understand why the average person doesn’t worry about this stuff more. How can we not have a good definition of life and not living a non-living and the nature of time and let alone consciousness and gravity”

  • Cold reason alone is not enough to guide humanity as it builds powerful technologies; the endeavor must include a spiritual or humanist dimension—a sense of soul, the human spark, possibly tied to consciousness—and science, art, and even religion are interrelated rather than opposed, as seen in figures like Da Vinci, Feynman, and Spinoza.

    “I think we need to approach it with whatever you want to call it, a spiritual dimension or humanist dimension... this idea of a soul, what makes us human, this spark that we have... I think that has to be at the heart of the endeavor.”

  • Games like chess, poker, Go, and diplomacy are simplified microcosm simulations of real-world situations that let people safely practice high-stakes decision-making, which is valuable because in real life one only gets about a dozen key decisions to make; games also healthily channel tribal and conflict energies into constructive rather than destructive outlets.

    “They’re simplified versions of some real world situation... it allows you to practice at them too, because how many times do you get to practice a massive decision moment in your life?”

  • Hope for human civilization rests on two near-limitless human traits: ingenuity—the general intelligence that lets the best minds reach the top of any field—and extreme adaptability, evidenced by hunter-gatherer brains evolved for the tundra now coping seamlessly with planes, podcasts, and even mind-blowing AI chatbots.

    “what gives me hope is that I think our almost limitless ingenuity, first of all... And then the other thing is our extreme adaptability.”

  • Beyond winning or losing, the self-improvement and mastery enabled by games and sports—measurably going from being unable to do something to being able to do it—is a deep source of human happiness and meaning, and learning to handle losses constructively (channeling the feeling into improvement) keeps one humble and motivated.

    “There’s nothing more satisfying in a way. It’s like, oh wow, this thing I couldn’t do before. Now I can. And again, games and physical sports and mental sports, their ways of measuring their beautiful, because you can measure that progress.”

0.31

Research judgment—posing deep falsifiable questions—is harder for AI than solving them and separates great from good scientists.

1 pointcentrality 3/5
  • In real blue-sky research there is no such thing as failure: as long as you pick experiments and hypotheses that meaningfully split the hypothesis space roughly in half—a kind of binary search—a failed experiment is equally valuable to a successful one because it tells you where to go next, provided the hypothesis is interesting, falsifiable, and within current technological reach.

    “when you do real Blue Sky research, there’s no such thing as failure really. As long as you are picking experiments and hypotheses that meaningfully split the hypothesis space and you learn something.”

0.29

Modeling natural systems requires choosing granularity level to capture relevant dynamics without overcomplication.

4 pointscentrality 2/5
  • Most emergent systems like cellular automata are modelable by classical systems through forward simulation, but chaotic systems where initial conditions matter and lead to uncorrelated end states could be the difficult boundary cases that resist efficient classical modeling.

    “Of course there’s the question of things like chaotic systems where the initial conditions really matter and then you get to some uncorrelated end state. Now those could be difficult to model.”

  • When modeling any natural system you must choose a cutoff level of granularity that captures the dynamics you care about without over-modeling: for a virtual cell, the protein level (using AlphaFold) likely suffices rather than going down to the atomic or quantum-mechanical level, with higher-level simulations built on those building blocks to produce emergent behavior.

    “you got 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? And then it captures the dynamics that you’re interested in.”

  • The learnability principle extends beyond living things: mountains shaped by weathering, planetary orbits, and asteroid shapes have all survived repeated physical processes, so even non-biological natural systems carry a reverse-learnable pattern—'survival of the stablest'—whereas abstract man-made things like large-number factorization may have no such structure and require brute force.

    “I sometimes call it survival of the stablest or something like that because of course there’s evolution for life, living things, but there’s also, if you think about geological times, so the shape of mountains, that’s been shaped by weathering processes over thousands of years”

  • A 'Virtual Cell'—a full in-silico simulation of a cell's internals, ideally starting with a yeast cell as the simplest single-cell organism—could let researchers run experiments and predictions in silico that save enormous wet-lab time, potentially achieving a hundredfold speedup by doing most search computationally and only validating in the wet lab, modeled at the protein level rather than the atomic level using AlphaFold as a building block.

    “Maybe you could a hundred x 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’s the dream.”