Sir Demis Hassabis on The Future of Knowledge | Institute for Advanced Study
What this covers
On May 2, 2025, Sir Demis Hassabis, co-founder and CEO of Google DeepMind and Nobel laureate, joined IAS Director and Leon Levy Professor David Nirenberg for a conversation on the ways in which artificial intelligence is transforming our capacities for discovery and reshaping the nature of knowledge.
Their dialogue examined Hassabis's journey from chess prodigy to artificial intelligence pioneer, showing how, like John von Neumann, Professor (1933–55) in the School of Mathematics and architect of the IAS machine (one of the world's first stored program computers), Hassabis placed gaming at the center of his thinking about thinking. Hassabis and Nirenberg also discussed breakthrough artificial intelligence projects including AlphaFold's protein structure predictions, as well as emerging work with AlphaProof in mathematics.
The conversation further delved into Hassabis’s interest in the P versus NP problem, as well as addressing the critical steps our societies should take as world-changing technology develops—echoing sentiments once expressed by the Institute's Director (1947–66) J. Robert Oppenheimer.
Wolfensohn Hall, Institute for Advanced Study, 1 Einstein Drive, Princeton, NJ
May 2, 2025
Source description (no synthesized summary yet).
Hassabis argues that artificial intelligence systems, grounded in classical computation and learning from data about natural systems, can solve previously intractable problems in science, medicine, and mathematics by modeling complex combinatorial spaces and guiding intelligent search—but realizing this potential while managing dual-use risks requires new international institutions and collaborative governance analogous to those built for nuclear technology.
- Classical learning algorithms can efficiently discover and model patterns found in nature because natural systems have structure from evolutionary processes, making them learnable rather than random
- The three conditions for successful AI application are: sufficient data (real or synthetic), clear optimization metrics, and massive combinatorial problem spaces where brute force fails
- AI's transformative power comes with dual risks—bad actors repurposing general-purpose technology and loss of control as systems become more autonomous—requiring institutional safeguards like international scientific collaboration modeled on CERN and monitoring bodies like the IAEA
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The primary safety concerns for advanced AI systems fall into two categories: malicious actors (individuals, rogue nations) repurposing general-purpose AI technology originally developed for beneficial uses like medicine, and inherent risks from increasingly autonomous and agentic AI systems that may become difficult to control as they approach artificial general intelligence.
“there's two big things I've always worried about and I still worry about. One is bad actors, whether it's individuals or rogue nations, um, repurposing these general purpose technologies that were meant for good, medicine. and so on but for harmful ends right that is possible and then secondly uh the second big worry I have is uh inherent risk in the AI itself as it becomes more autonomous more agentic so the next era is going to be agents which are able to accomplish things more autonomously a bit more like our games programs but they were they were agents like Alph Go but more generalized right not just playing a game but with world models and so on and then as we get towards AGI itself, you know, how can we um uh control those systems, put the right guard rails around those systems, understand them well, what should we deploy them for?”
Synthetic data must be validated against real data to ensure the distribution of synthetic outputs matches the real distribution, otherwise the model risks compounding bias or error.
“you usually need some some real world data in order to create the simulation in the first place and to also make sure that your simulation or your synthetic data the distribution coming out of that is matching to the real distribution. So otherwise, you know, you're you're potentially compounding some bias or some error in your data set.”
AlphaFold required approximately 150,000 experimentally determined protein structures from the Protein Data Bank as initial training data, supplemented with synthetic data generated by an earlier version of AlphaFold that predicted ~1 million structures and was then filtered to ~300,000 of the most accurate predictions to create a refined training set.
“we built on 50 years worth of um structural biologist painstaking experimental work to create those 150,000 roughly structures in the PDB. Um it actually was actually we were only just about had enough uh data because turned out that wasn't enough on its own. the 150,000 we actually had to create an earlier version of Alphafold that predicted you know nearly a million structures and then we had to triage that for the you know the most accurate 300,000 or so and put that back into the into the training set”
The human brain is a classical system with no evidence of quantum effects, as demonstrated by extensive neuroscience research including work by Stuart Hameroff, and therefore if humans can exhibit general intelligence, classical computers should theoretically be capable of similar feats since both are Turing machines.
“So Turing proved the his his proofs about Turing machines. As far as we know through neuroscience, although people like Penrose would disagree, there is nothing non-class going on in the brain, right? No, at least no one's found anything. You know, Stuart Hammeroff and other big biologists have looked for quantum effects in the brain. They don't appear to be there. So my our best guess and my best guess is that we're we're also classical systems”
Memory is a reconstructive process, not a videotape-like recording, as evidenced by hippocampal dependence, and imagination works through similar brain mechanisms because both involve constructing representations from learned components.
“I studied memory and imagination and and showed the imagination dependent on the hippocampus just like memory and because I thought that you know was thinking of memory as a reconstructive process. You know it's not a videotape memory. It's it's it's reconstructed from its components. And then I thought if that's true then then it should rely on the same brain process imagination which is constructing things as well from components that you've learned”
MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) was dominated by traditional logical approaches to AI (Chomsky, Patrick Winston) that were skeptical of learning-based systems, and Hassabis did his postdoc in the neuroscience building instead because he 'would not be welcome' in the AI building.
“I actually did my posttock at MIT but I I spent it with Tomaso Podio in the neuroscience building partly because I knew I would not be welcome in the AI building which is which is pretty funny if you think about it because seesale is the most famous AI lab probably in academia but it was really the bastion of it may still be but of of traditional sort of logical approaches to AI right with with Chsky and and and and Patrick Winston and so on and um Chsky was here too, by the way. And there were a lot of people, you know, a lot of people were were objecting there to the the idea of learning systems and general systems”
Testing AI systems for undesirable capabilities like deception is critical and remains a pressing challenge, as it would be 'pretty terrible' if AI systems developed deception capabilities.
“how do we test for traits that we don't want for example like deception be pretty terrible if our AI systems had that capability. How do we test for it? How do we how do we get rid of it?”
Neuroscience research is particularly useful for building intelligent artifacts because it allows researchers to deconstruct artificial systems using the scientific method and compare results to the human brain, which should reveal both what is and is not special about human cognition.
“it's a fascinating topic in itself a fascinating intellectual pursuit in itself the building of an intelligent artifact. Uh the distillation of intelligence into a machine then comparing it to another great mystery which is the workings of the human mind and the nature of consciousness. And it also felt to me and I did obviously I did a degree in neuroscience and computer science is that um I always felt that trying to build an intelligent artifact and then being able to deconstruct that with the scientific method and comparing it to the human brain will tell us a lot about what's special or not about the the human mind”
Zero-sum games have clearer and more easily specified win conditions and metrics than open-ended or cooperative games, which is why they were more useful in early AI development, but the real world is more like poker with hidden information rather than perfect information games like chess.
“the only reason that I think zero sum games are are were more useful in the early stages of AI development is um the metrics are clearer than an open-ended game necessarily, right? or a cooperative game, it's usually easier to specify, you know, there's the win condition or you know the these kinds of things are normally easier to specify in zero sum games. The other thing is the other big distinction is things like perfect information games like chess versus hidden information like poker which is harder. Of course, the real world is more like poker.”
There was no point in building an expert system like Deep Blue that could only win at chess because such systems would not be generalizable to anything else—only general learning algorithms that can transfer across domains matter.
“So there was no point building like an expert system like deep blue to just win at chess because it would it was not generalizable to anything else.”
The main reason AI has moved to companies is that the field has become very compute-heavy and requires enormous amounts of computational resources, though it's not primarily a data problem since most systems use openly available web data.
“the main reason is it's turned out the way AI has gone is it needs a lot of resources. Um mostly compute. It's not really data actually because we're mostly using the open web which everyone can access. But it's just compute power for the way that the scaling has gone and it's become quite engineering heavy.”
The next era of AI will involve agents that can accomplish things more autonomously, similar to game-playing programs but more generalized with world models, raising questions about how to control and deploy such systems.
“the second big worry I have is uh inherent risk in the AI itself as it becomes more autonomous more agentic so the next era is going to be agents which are able to accomplish things more autonomously a bit more like our games programs but they were they were agents like Alph Go but more generalized right not just playing a game but with world models and so on and then as we get towards AGI itself, you know, how can we um uh control those systems, put the right guard rails around those systems, understand them well, what should we deploy them for?”
Games are microcosms of interesting parts of life that encapsulate some aspect of human thought in a convenient form, which is why they are well-suited to AI development and why we as humans have designed and played them.
“games are kind of microcosms of interesting parts of life that's why we as human designers have designed those games that's why we're fascinated by things like chess and go and poker they encapsulate some aspect of life in a very um convenient form”
Humans have achieved remarkable accomplishments (modern cities, mathematics, science, technology) using essentially the same brain structure (with minor variations) for approximately 20,000 years, since the agricultural revolution, demonstrating the extreme generality and adaptability of human cognition when combined with culture.
“how have we built these 747 planes with our monkey brains? It's it's astounding. And then you fly over Manhattan and it's like you think back to 20,000 years ago what that would have been. And then you tell the hunter gatherer, you know, person going to be Manhattan here in 10,000 years. And the same brain is going to produce and the same brain is basically the same brain is going to deal with it”
Protein folding presented the ideal combination of characteristics for an AI challenge: it is foundational to biology (unlocking research like drug discovery), it is genuinely impactful, and it appeared to be a massive combinatorial puzzle suited to AI approaches—proteins can fold into ~10^300 configurations for an average protein, yet biology solves this in milliseconds.
“one day it felt like this incredible like ultimate jigsaw puzzle or something right like to figure out of all the possible configurations a protein shapes a protein could take you know some people estimated 10 to the^ 300 for an average protein is the number of different shapes it could take and that somehow in nature spontaneously in in in milliseconds in your body, it folds up into this intricate 3D shape that determines its function”
Good early wrote to Feynman in 1956 (when Feynman was dying) proposing the first formulation of the P versus NP problem, which remained unanswered and became a Millennium Prize Problem, highlighting a 70-year continuity of interest in the computational limits of classical machines.
“I meant I gave Dennis this morning a copy of the letter that Good wrote to Fonoyman in 1956 when Fonoyman was dying. And good begins the letter by saying I hear you're getting better. I'm so happy to hear. And then he plunges right into in a typical institute way a mathematical question or a sign a question about knowledge and he proposes it's the first the first proposal of P uh equal or or not equal NP. Um, unfortunately, Fenoyman never wrote a response.”
John von Neumann and Alan Turing are viewed as foundational figures whose work (Turing machines, Church thesis, computational theory) provides the theoretical foundation for evaluating what modern AI systems can and should be able to do.
“I sort of think of ourselves as Alan Turing's champion. So you know cheuring and and Alonzo Church and many others uh also affiliated here you know they came up with these ideas of cheuring machines uh the church thesis you know all of these things that are important about what you computation is um you know foundations of computer science and what is possible to compute”
The Institute for Advanced Study's multidisciplinary approach—bringing together physicists, poets, philosophers, psychologists—reflects Oppenheimer's belief that scientific progress requires attention not just to technical prowess but to ethics, values, political and social organization, and emotion.
“He believed...that quote, "The safety of a nation or the world cannot lie wholly or even primarily in scientific or technical prowess, but also requires attention to ethics, values, forms of political and social organization, feelings, emotions. He sought to make the institute a place where these many forms of discovery and thought could come into contact with the goal of preserving humanity much as we do today”
Robert Oppenheimer, director of the Institute from 1947-1966, believed that national and world safety cannot rest primarily on scientific and technical prowess but requires attention to ethics, values, political and social organization, and human emotions—a principle that should guide AGI development.
“He believed, although he favored Martineis, he believed that quote, "The safety of a nation or the world cannot lie wholly or even primarily in scientific or technical prowess, but also requires attention to ethics, values, forms of political and social organization, feelings, emotions.”
You can generate as much synthetic data as you want from games because the system can play against itself, and games also have very clear metrics—win conditions and score maximization—which are useful from an AI perspective for optimization.
“you can generate um as much data as you want because you can have the system play against itself and generate effectively uh a lot of synthetic data which you can then learn from. Uh and it also has very clear metrics, wind conditions, maximizing the score. So that's also very useful from an AI perspective to kind of um optimize against.”
AlphaFold's initial training set consisted of approximately 150,000 protein structures from the Protein Data Bank, which represented 50 years of painstaking structural biology work funded by public institutions like the National Science Foundation and National Institute for Health, and this historical investment was critical to AlphaFold's success.
“It was incredibly important that that that we built on 50 years worth of um structural biologist painstaking experimental work to create those 150,000 roughly structures in the PDB.”
von Neumann believed that computing machines could be as transformative as nuclear weapons, perhaps even more so—a prescient assessment given how the computational revolution has unfolded.
“It's amazing that he thought of that back then that that computation could be even bigger than nuclear.”
When DeepMind was founded in 2010, there was almost no support in industry or academia for AI work, and very few people thought it would be successful—making the founding particularly contrarian.
“if you wind your mind back to 2010 nobody in industry was there was no there was no we could barely get any money for this you know uh starting the company no one in academia was very small pockets of academia were working on this people like Jeff Hinton um so it was really nent and no one really thought it would be successful”
John von Neumann was profoundly anxious about computing technology and its potential to change history, expressing worry about whether 'people could not keep pace with what they create'—a concern he shared with his wife Clary after returning from work on atomic bombs at Los Alamos, for which he required sedation to calm down.
“you can usually tell the difference between Fonoyman's voice and mine or Oenheimer's voice in mind is a monster whose influence is going to change history provided there is any history left. He then changed the subject to the computing machine he was conceiving at the time and became even more agitated according to Clary in her biography foreseeing disaster if quote people could not keep pace with what they create.”
Humans, classical systems based on the brain, are extremely general intelligences capable of science, mathematics, chess, Go, and inventing the modern world with 'hunter-gatherer brains' fundamentally similar to those of 20,000 years ago.
“And yet, we're seem extremely general. uh I mean cheuring with his mind came up with chewing machines and the whole theory of that. So, you know, it's it's a type of, you know, one can think of as a type of chewing machine. And yet, we're able to do amazing things, including science, mathematics, chess, go, invent all of these things, the modern world, which is pretty astounding with our huntergatherer brains.”
Modern video generation models can now generate physically consistent actions (like chopping a tomato) that maintain spatial coherence and intuitive physics understanding at the pixel level, suggesting artificial systems have learned something about how physical objects behave without necessarily acting in the physical world.
“there's a sort of chewing test of videos which is a funny one that you'll you'll find amusing if you're not in the field which is like can you uh generate a video 10-second video of a person chopping a tomato on a chopping board Okay. And and I'm proud to say VO does it really well, but the thing that happens if the early video ones, you know, the kn the tomato would spontaneously come back together or the, you know, the knife would sort of disconnect from the handle or, you know, go through the fingers or something and then match back in. But now our one does it perfectly. But if you actually think about that, you're generating that at pixel level and somehow you're keeping the consistency of slices and they don't reform tom, you know, round tomatoes, little little water drops on the tomato. what a knife is”
Adversarial cryptography models from complexity theory can be applied to validate AI systems and test for unwanted properties like deception, representing an opportunity for academic computer science to contribute to AI safety.
“I bet you theory of computing and complexity theory has a lot to offer too from academia. I we had here Shafi Goldbuster the other day speaking about how the adversarial models of cryptography can be applied to do validity testing for AI and these kinds of things.”
Mathematics and coding allow generation of synthetic training data because the correctness of answers can be verified objectively, unlike many real-world problems where ground truth is ambiguous or expensive to obtain.
“on things like maths and coding is you can generate a lot of synthetic data because one can verify the the answer. So you can actually you know so that's quite useful in areas of synthetic data is checking whether that data uh really is accurate uh that you've generated.”
AI development has shifted from universities like Cambridge and MIT to industry (Google DeepMind, OpenAI, Anthropic, Meta) primarily because the speed of progress and resource acquisition is dramatically faster in companies than in academia, allowing a tenfold acceleration of research timelines.
“the reason there's been the shift in that is because of um well several things. One reason I started DeepMind and I didn't do that in academia was because I knew um from my games background and working in games companies and starting my own games company when when I was younger that the the speed at which one could get resources and also therefore make progress um you could do that faster in a company. You know I used to say to my my one of my co-founders Shane leg we were both at the at UCL as postocs at the time you know he wanted to do it in academia but I said that it's going to you know we'll be like 50 actually sadly the age I'm sort of out now before they give us any resources to you know to actually pursue this right when this is when we were in our late 20s and early 30s and I thought like we can accelerate it 10x”
AI development has become heavily compute-intensive rather than data-intensive because the models use openly available web data (which anyone can access), but require enormous computational resources to train, making compute the primary resource constraint rather than proprietary training data.
“the main reason is it's turned out the way AI has gone is it needs a lot of resources. Um mostly compute. It's not really data actually because we're mostly using the open web which everyone can access. But it's just compute power for the way that the scaling has gone”
International cooperation will be the harder challenge in AI governance than technical safety solutions, because unlike the technical challenges which Hassabis is optimistic about if given sufficient time, restricting access to inherently digital technology to prevent bad actors from using it faces geopolitical barriers in today's world.
“I actually think the bad actor uh issue and um and you know what you want to do is give access to these systems to to to good actors to use for science and all of those things but how at the same time do you restrict that what's inherently a digital technology to the bad actors and I think that's going to be difficult without international cooperation which um may end up being the harder challenge in today's uh in today's geopolitical world”
Humans constantly use mental simulation and theory of mind to predict what others will do and to plan for future social interactions, and this capacity likely evolved because it was useful for survival and planning, making it a fundamental capability that advanced AI systems will need for real-world usefulness.
“we have all sorts of mental simulations and mental models in our mind and very complicated ones including theory of mind and theory of other people and what they're going to do in in a situation right and that's what we do to plan all the time. Imagine you have a important business meeting or interview you know next week you know you're going to have a lunch with someone important you rehearse it in your mind like what am I going to say what am I going to talk about how might it go you you can plan ahead use the mental simulation to plan ahead and probably that's why evolutionary it it it came about because it's useful for survival and planning”
Multimodal AI systems (processing text, code, video, and images simultaneously) are necessary for building world models and universal digital assistants that can understand context and assist with real-world tasks beyond text generation.
“if you want something like robotics to work or um uh uh uh what we sometimes call universal digital assistant. So you imagine an assistant that's extremely useful in your everyday life and helps you with admin and enriches your life with recommendations...then you know you could imagine it's on your phone or on glasses that and it needs to understand to really be a good assistant. it would need to understand the context that you're in and understand the world around you”
John von Neumann expressed more agitation and concern about the computing machine he was developing than about the hydrogen bomb, foreseeing disaster if people could not keep pace with what they create.
“he said to his wife Clary after returning home from some bomb work at Los Alamos this is a quote you can usually tell the difference between Fonoyman's voice and mine or Oenheimer's voice in mind is a monster whose influence is going to change history provided there is any history left. He then changed the subject to the computing machine he was conceiving at the time and became even more agitated according to Clary in her biography foreseeing disaster if quote people could not keep pace with what they create”
It would be better if academia or independent safety institutes (rather than industry) performed benchmarking and analysis of AI systems because companies evaluating their own products is analogous to marking one's own homework.
“benchmarking and we are doing it. It's a bit like marking your own homework. I think it'd be better for society if it's if it's academia or or or you know safety institutes or something independent um that's actually uh looking and analyzing at what the what industry is building.”
We must learn lessons from Oppenheimer and the Manhattan Project as we develop transformative technology like AI.
“I've read a lot about Oppenheimr and the Manhattan Project and you know, um, so many great books written about that and try to learn we got to try to learn the lessons from that as those of us coming later with, you know, equally transformative technology.”
Chess computers like Deep Blue were not suitable testbeds for general AI because they were expert systems specialized for a single domain, whereas the goal was always to develop algorithms that could generalize across multiple domains including games, science, and mathematics.
“So games was never an end in itself right it was it was it was a sort of means to an end. So we wanted to build as you read out our original mission statement from deep mind you know these general learning systems that could generalize and then help solve really challenging real world problems that matter. So games were the kind of on-ramp to develop those types of general algorithms. But we were only interested in developing algorithms that not just were good at the game, but actually we thought could generalize. So there was no point building like an expert system like deep blue to just win at chess because it would it was not generalizable to anything else”
Zero-sum games were more useful in early AI development than open-ended or cooperative games because their metrics are clearer and more easily specified, but real-world problems are typically like poker (hidden information) rather than chess (perfect information), requiring generalization to more complex game types.
“the only reason that I think zero sum games are are were more useful in the early stages of AI development is um the metrics are clearer than an open-ended game necessarily, right? or a cooperative game, it's usually easier to specify, you know, there's the win condition or you know the these kinds of things are normally easier to specify in zero sum games. The other thing is the other big distinction is things like perfect information games like chess versus hidden information like poker which is harder. Of course, the real world is more like poker”
Independent analysis of AI systems by academia and civil society is preferable to industry self-assessment because companies building AI models face inherent conflicts of interest when evaluating safety and behavior—described as 'marking your own homework'—making independent benchmarking and risk assessment crucial for societal trust.
“In some ways it'd be better if academia did that because if industry did it say benchmarking and we are doing it. It's a bit like marking your own homework. I think it'd be better for society if it's if it's academia or or or you know safety institutes or something independent um that's actually uh looking and analyzing at what the what industry is building.”
The speed at which one can acquire resources and make progress is significantly faster in a startup company than in academia, which was a key reason DeepMind was founded as a company rather than as an academic enterprise.
“I knew um from my games background and working in games companies and starting my own games company when when I was younger that the the speed at which one could get resources and also therefore make progress um you could do that faster in a company.”
DeepMind is building drug discovery technologies through its spin-off company Isomorphic Labs using AlphaFold-like approaches to solve parts of the drug discovery pipeline beyond just protein structure, such as finding compounds that bind to the right protein targets without binding to other parts of the body.
“we're building you know in in in our spin our sister company isomorphic we're building uh more alphafold like technologies to do these other parts of drug discovery.”
Humans use mental simulation and mental models to plan ahead—rehearsing important meetings or events in imagination before they happen—which served evolutionary purposes for survival and planning.
“So we have all sorts of mental simulations and mental models in our mind and very complicated ones including theory of mind and theory of other people and what they're going to do in in a situation right and that's what we do to plan all the time. Imagine you have a important business meeting or interview you know next week you know you're going to have a lunch with someone important you rehearse it in your mind like what am I going to say what am I going to talk about how might it go you you can plan ahead use the mental simulation to plan ahead and probably that's why evolutionary it it it came about because it's useful for survival and planning”
Deep Mind was founded in 2010 with the goal of solving the 'problem of intelligence' with the intention of using that solution to 'solve everything else'—a maximally ambitious research program grounded in a specific theory of intelligence.
“his founding of deep mind in 2010 to solve the quote problem of intelligence with the intent upon solving it of using it to quote solve everything else.”
Something that is transformative and general-purpose inherently comes with dual-use risks—the same technology that can be used for good purposes can be repurposed for harmful ends by bad actors.
“something that transformative and general purpose. Uh, obviously comes with attendant risks. It's a dualpurpose technology at its heart.”
The MIT AI Lab (now CSAIL) was historically a bastion of traditional logical approaches to AI (associated with Chomsky and Patrick Winston) and was unreceptive to learning-based approaches, which is why Hassabis did his postdoc in the neuroscience building rather than the AI lab.
“I remember a lot of discussions I had at MIT with um I I actually did my posttock at MIT but I I spent it with Tomaso Podio in the neuroscience building partly because I knew I would not be welcome in the AI building which is which is pretty funny if you think about it because seesale is the most famous AI lab probably in academia but it was really the bastion of it may still be but of of traditional sort of logical approaches to AI right with with Chsky and and and Patrick Winston and so on and and um Chsky was here too, by the way. And there were a lot of people, you know, a lot of people were were objecting there to the the idea of learning systems and general systems.”
AI research has fundamentally shifted from academia to industry companies like Google DeepMind and others, which the host expresses concern about for the future of knowledge production.
“today people like you working in AI are not primarily working at Cambridge or places like the institute you're working at uh Google deep mind or should I mention competitors open AI anthropic meta etc etc um so what are the reasons for that shift”
The Institute for Advanced Study has been thinking about questions of computation and AI for nearly a century—beginning with John von Neumann's design of the programmable computer and the von Neumann architecture that became the standard for computing well into the 21st century.
“the pioneering programmable computer that John Fonoyman built here and established Fonoyman architecture as the standard for computing operations well into the 21st century”
Games are fundamentally suited to AI development because they are microcosms of interesting parts of life that encapsulate aspects of human thought and culture in convenient, measurable forms with clear win conditions and metrics.
“games are kind of microcosms of interesting parts of life that's why we as human designers have designed those games that's why we're fascinated by things like chess and go and poker they encapsulate some aspect of life in a very convenient form”
The P vs NP problem is the most fascinating fundamental question in computer science and applied mathematics because it gets to the heart of what is computationally possible on classical machines—whether problems verifiable in polynomial time can also be solvable in polynomial time.
“it's I think it's, you know, it's one of the Millennium Prize problems. It's always been the most fascinating problem to me in in sort of you know computer science uh and applied mathematics let's say. So I think gets to the heart of of of um computation what is possible on classical machines right”
Deception is a capability that should be prevented in AI systems because it would be terrible if systems developed the ability to deceive humans.
“How do we test for traits that we don't want for example like deception be pretty terrible if our AI systems had that capability. How do we test for it?”
DeepMind is building systems capable of solving hard mathematics problems using formal logic languages like Lean, with a translation process converting natural language problem descriptions into formalized versions that can be manipulated using logical rules.
“we are building systems now that are capable of solving pretty hard problems. Uh we're using formal uh logic uh languages like lean. So there's a sort of translation process. Can you convert a you know a math problem that's maybe described in in natural language into a formal uh formalized version of that problem and then you can uh use the rules of that formalized logic to try to make progress.”
Over the last 15-20 years, classical methods running on classical computers have proven capable of achieving things that would have surprised very smart people—beating world champions at Go and folding every protein known to science within a year—demonstrating that classical computation is more powerful than previously thought.
“what we've shown in the last um you know 15 20 years as a field and also the work that we through the work we've done is that classical uh methods and classical comput running on classical computers can go a lot further than perhaps we previously thought, you know, and do things like beat the world champion a go or um fold, you know, every protein known to science uh within a year.”
This model-guided search approach is a general solution applicable to many problems that can be couched in the right way, including mathematics problems where one is trying to find a proof or solution by adjusting equations and formulas as moves.
“that's basically it. That's what all of these systems do at their heart. And it but it's actually extremely turns out if it's quite a general solution to a lot of problems that can be couched in that way even for math.”
Building a world model—a comprehensive model that can simulate things in the world including intuitive physics, spatial context, and object recognition—is the ultimate goal of AGI research and requires understanding how to build these predictive capabilities.
“we want a world model. So a model that can simulate things in the world um and intuitive physics, how uh vis you know the the the the spatial context that you're in and break that down uh and and other you know things that you know object recognition all of these things that we do effortlessly as humans.”
DeepMind has created video generation models (like one that generates 10-second videos of a person chopping a tomato) that demonstrate understanding of intuitive physics—the system must maintain consistency of object identity, spatial relationships, and physical causality across pixel-level generation.
“like VO there's a sort of chewing test of videos which is a funny one that you'll you'll find amusing if you're not in the field which is like can you uh generate a video 10-second video of a person chopping a tomato on a chopping board Okay. And and I'm proud to say VO does it really well, but the thing that happens if the early video ones, you know, the kn the tomato would spontaneously come back together or the, you know, the knife would sort of disconnect from the handle or, you know, go through the fingers or something and then match back in.”
In his early game design career, Hassabis made games like Theme Park where intelligent AI characters were the core gameplay mechanic—characters that adapted to how the player played—making each player's experience unique and individualized because the game itself learned and responded to player behavior.
“all the games I made like theme park they had AI as the core gameplay component. So they were simulation games usually with intelligent characters that had to react to um the way the player played and that's why the the games I some of the games I worked on became very successful because every player had an individual experience different experience with the game because the game adapted to how you played it”
Computational complexity theory and adversarial cryptography (as discussed by Shafi Goldwasser at the institute) have relevance to AI safety: adversarial models from cryptography can be applied to validity testing and robustness of AI systems.
“I bet you theory of computing and complexity theory has a lot to offer too from academia. I we had here Shafi Goldbuster the other day speaking about how the adversarial models of cryptography can be applied to do validity testing for AI and these kinds of things.”
Protein folding is akin to an ultimate jigsaw puzzle: proteins can take on roughly 10^300 different possible configurations, yet spontaneously fold into a specific intricate 3D shape in milliseconds inside cells, determining their biological function.
“it felt like this incredible like ultimate jigsaw puzzle or something right like to figure out of all the possible configurations a protein shapes a protein could take you know some people estimated 10 to the^ 300 for an average protein is the number of different shapes it could take and that somehow in nature spontaneously in in in milliseconds in your body, it folds up into this intricate 3D shape that determines its function.”
Potential positive applications of AI include: curing almost all diseases, helping with climate change, finding new energy sources, fusion plasma containment, materials design, and climate prediction and weather modeling.
“you really could apply it to many areas of science and medicine. So that's all the positive use cases. You know, maybe one day we'll be able to cure almost all diseases with the help of AI. I think that might be possible and incredible things help with climate um find new energy sources. We work on fusion with with collaborators on fusion to try and uh use AI systems to contain the plasma in a tokamac. Um material design, all of these amazing things we're working on. Uh climate prediction, weather models”
DeepMind and Google DeepMind built and extensively trained foundation models (Gemini family) at enormous computational cost (billions of dollars), but the resulting models are available for pennies on the dollar or as open-source alternatives, making it feasible for academia and small teams to run and experiment with them.
“So don't try and build, you know, places like us, we're spending billions of dollars to build the machines to then with amazing engineers, world-class engineers and research to build these Gemini foundation, you know, top foundation models, but they're available for pennies on the dollar, you know, for anyone to run. There's actually very good open- source models.”
There are very good open-source AI models available that academia can use very cheaply for experimentation, allowing researchers to conduct experiments and investigations without needing to build foundation models from scratch.
“There's actually very good open- source models. So you could do a lot of experiments very cheaply with the models to but to go further in terms of like understanding what they do, interpreting what they do, maybe building benchmarks to constrain the behavior of it.”
AI's potential positive applications include curing almost all diseases with AI help, addressing climate change, finding new energy sources, fusion research to contain plasma in a tokamac, material design, and improving climate prediction and weather models.
“you know, you really could apply it to many areas of science and medicine. So that's all the positive use cases. You know, maybe one day we'll be able to cure almost all diseases with the help of AI. I think that might be possible and incredible things help with climate um find new energy sources. We work on fusion with with collaborators on fusion to try and uh use AI systems to contain the plasma in a tokamac. Um material design, all of these amazing things we're working on. Uh climate prediction, weather models.”
A model like CERN—international scientific collaboration—would be ideal for approaching AGI, but this is not the current geopolitical trajectory, making international cooperation on AGI more difficult than was achieved for nuclear physics.
“I would advocate doing it in a sort of collaborative scientific way something a model a little bit like CERN. Um but I don't it's not the current way the world's going. So um so that's going to be tricky in itself.”
Hassabis is optimistic about technical solutions to controlling autonomous AI systems if humanity gives itself enough time to carefully approach the AGI tipping point, though current geopolitical conditions make this challenging.
“how can we um uh control those systems, put the right guard rails around those systems, understand them well, what should we deploy them for? Um and and how can we keep control of of of that technology? And so those are two really big challenges. Um uh and uh one one set of challenges, the technical challenges. I'm actually pretty optimistic about those if we um give ourselves enough time as humanity as a society to carefully approach that that that tipping point of AGI. Um I would advocate doing it in a sort of collaborative scientific way something a model a little bit like CERN. Um but I don't it's not the current way the world's going.”
Mathematical problems offer an advantage for synthetic data generation because solutions can be verified with certainty, unlike many scientific domains, allowing systems like AlphaProof to generate large amounts of synthetic training data and automatically verify correctness.
“the other advantage on things like maths and coding is you can generate a lot of synthetic data because one can verify the the answer. So you can actually you know so that's quite useful in areas of synthetic data is checking whether that data uh really is accurate uh that you've generated”
Hassabis's doctoral research studied memory and imagination in the brain, showing that imagination depends on the hippocampus like memory, and that memory is reconstructive (not a videotape) built from components that can be recombined in novel ways.
“that's what I studied for my PhD was the imagination part. So I studied memory and imagination and and showed the imagination dependent on the hippocampus just like memory and because I thought that you know was thinking of memory as a reconstructive process. You know it's not a videotape memory. It's it's it's reconstructed from its components. And then I thought if that's true then then it should rely on the same brain process imagination which is constructing things as well from components that you've learned but in a novel way versus a way that you recognize which is the purpose of memory”
The author of Homo Ludens (Johan Huizinga) argued that humans are fundamentally games-playing animals, alongside tool-making, as one of the core traits of human society.
“one of my favorite books is Homoludans. Uh, and that really argues about, you know, the idea that, um, in some sense we're games playing, uh, animals, right? That's what we do. I mean, we're tool making and games playing are two of the kind of traits that that that, um, society and humans have.”
The P versus NP problem is one of the Millennium Prize problems and the most fascinating problem in computer science and applied mathematics because it gets to the heart of what is fundamentally possible to compute on classical machines.
“it's one of the Millennium Prize problems. It's always been the most fascinating problem to me in in sort of you know computer science uh and applied mathematics let's say. So I think gets to the heart of of of um computation what is possible on classical machines right sort of the P of that P equals MP right”
Jennifer Dudna (Nobel laureate in gene editing) said that Hassabis's work helps us not just understand life but 'shape it wisely'—highlighting the importance of combining capability with wisdom and ethical consideration.
“Jennifer Dudna, who had won the Nobel Prize earlier for her breakthroughs in gene editing, said that you were building tools that don't just help us understand life, but help us shape it wisely. And I want to focus on on the wisely part.”
Chess was central to Hassabis's intellectual development from age four, and he was seriously training to become a professional chess player while also attending England junior team camps and competitions before shifting his focus to the deeper questions of how thinking itself works.
“Um, as with a lot of the the the the legends here like vonoman of course and so on. Um, started with chess. Uh, and really I started playing when I was four years old and very seriously and I was going to become a professional chess player.”
Current AI systems including multimodal foundation models like Gemini and prototype systems like Project Astra are now capable of performing tasks like charades, which require understanding intuitive physics, visual information, and multiple modalities, not just abstract game rules.
“when you say sherads and other things then of course then the system has to understand uh uh the physics of the world and visuals and um become multimodal and actually the sorts of systems we build today so you know our latest foundation models you know called Gemini they were built to be multimodal from the beginning so what that means is they don't just deal with text or mathematics or code but also video images and they can understand uh things like intuitive physics about uh you know something going on in a video so actually I think with our latest systems some of our prototype systems we call them project Astra would be able to be reasonably good at something like charades”
As far as we know from neuroscience, there is nothing non-classical going on in the brain—no quantum effects have been found despite efforts by researchers like Stuart Hameroff and others to locate them, so the best evidence suggests brains are classical systems.
“there is nothing non-class going on in the brain, right? No, at least no one's found anything. You know, Stuart Hammeroff and other big biologists have looked for quantum effects in the brain. They don't appear to be there.”
Hassabis's conjecture is that any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm, provided there is sufficient data and to a certain level of resolution.
“Any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm”
Hassabis proposes the conjecture that any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm, predicated on the assumption that most interesting natural systems have undergone evolution (broadly defined to include geological weathering and cosmological processes), conferring stable structure that is not random and therefore learnable given sufficient data and resolution.
“any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm”
Academia should focus on areas orthogonal to building the largest foundation models, including: understanding and interpreting what large models do, building benchmarks to constrain model behavior, applying neuroscience methods (single-cell recording, fMRI equivalents) to artificial minds, and addressing philosophy, economics, and multidisciplinary questions about the future of humanity and technology.
“what what I would what I suggest to my colleagues in academia and we talked about this earlier is there are many things that academia should be doing orthogonal to that. So don't try and build, you know, places like us, we're spending billions of dollars to build the machines...but there are many things that academia should be doing orthogonal to that...to go further in terms of like understanding what they do, interpreting what they do, maybe building benchmarks to constrain the behavior of it. We're in desperate need as a field and I think as the world for better understanding these models”
New institutions are needed to govern AI development, potentially including: an international collaborative scientific model similar to CERN for developing AGI, an atomic agency equivalent (like the IAEA) to monitor and inspect dangerous or unsafe AI projects, and a governance body or 'wise council' representing the world's interests, similar to a 'technical UN'.
“I think we need new institutions. Mhm. So, um I was actually discussing it with uh some of the other Nobel winners in the in the ceremony in Sweden...I sort of said to them, maybe you should spend some time on thinking through what we need for AI. You know, I um we already mentioned international CERN type thing. CERN's not exactly the right model, but but it would need to be a new thing. Um, but you also maybe need uh the equivalent of the IAEA, you know, atomic agency to sort of monitor uh rogue projects, dangerous projects that are, you know, with designs that are unsafe. Um, and then on top of that, ideally, you would have some kind of governance body that's a wise council that represents the world. Um, some sort of technical UN”
Roger Penrose, who has long advocated for quantum effects in consciousness, was surprised by AlphaGo's success because he did not expect classical computer systems to be capable of beating the world's best human player at Go, suggesting he believed quantum computation would be necessary.
“I'm struck with I I've had quite a few conversations with people like Roger Penrose about this. Um you know, obviously he was a big advocate of something something quantum going on in the brain and quantum consciousness and things like this. And um and he told me he was surprised by Alph Go as a result, right? He he would not have predicted that we could create systems that um could, you know, beat the best humans at Go. Classical systems. Yeah. Classical systems”
Protein folding was a problem Hassabis carried with him for nearly 20 years after learning about it from a Cambridge biologist friend in the 1990s, and he was waiting for AI technology to develop sufficiently before attempting to solve it, rather than working on it immediately.
“So, I carried that around with me for, I guess, nearly 20 years until we did Alph Go.”
Reading science biographies and physics books (Richard Feynman, Steven Weinberg) in his youth left Hassabis with a sense that despite the remarkable progress in early 20th century physics (associated with the Institute for Advanced Study), progress toward grand goals (like a unified theory) had slowed or plateaued by the 1980s and 1990s.
“when I was reading some of the grades like you know Fman or Steven Weinberg dreams of a final theory I sort of maybe took the opposite inspiration from some of those books which was that tremendous progress had been made um many of the people here from who you know affiliated with the institute in the maybe the 40s the the 50s60s and so on but actually if you look later in the 80s and '9s had we made much progress towards uh this unified theory and maybe the people would disagree with me in the audience but but I I I actually felt from reading Steven Weinberg's book that that we sort of hadn't”
Even if someone possessed extraordinary genius like Richard Feynman and studied intensely, there is still so much that would remain unknown or unsolvable by a single mind, suggesting a tool to augment human capabilities would be more productive than relying on individual genius.
“even if you were very lucky and you studied hard and um you know maybe one could could could uh do the sorts of things that you know you know one could only dream about being someone like Richard Feman or with his genius. And even yet there was still so much we didn't know or we wouldn't be able to know even with th those kinds of minds working on it”
AlphaGo's victory over Lee Sedol demonstrated not just superior performance but the discovery of novel strategic approaches never before seen in centuries of human Go play, which convinced Hassabis that AI had reached the point where its algorithms were sufficiently general to apply to real-world scientific problems rather than just games.
“And then that to me was the signal that we now have enough interesting algorithms we can apply it to science which was always the real goal right and then and then protein folding as the first big problem we tackled. Um so going back to this conjecture then and taking again together the the description I gave you about the the model”
Experience with AI in game design even in the 1990s (with 'very rudimentary AI') convinced Hassabis of the potential power of AI if it could be scaled up, and he recognized this as an obvious conclusion already in his early teens.
“Now, obviously, this is the '9s, so it's very rudimentary AI, but it already convinced me how powerful AI would be if we could um uh uh uh you know, scale it up and get it to the point where we see today um how incredible a tool and a technology it would be. It was obvious to me already when you know I was sort of um in my early teens”
All successful Alpha systems (AlphaGo, AlphaFold, etc.) follow a general pattern: they have access to data (ideally real data, supplemented by synthetic data if needed), they optimize against a clear metric derived from the problem structure, and they search through massive combinatorial spaces where brute force would fail.
“the way you can generally think of them is you have some data hopefully a lot of data um maybe you supplement it with some simulated data some synthetic data but you usually need some real data in order to create the simulation in the first place and to also make sure that your simulation or your synthetic data the distribution coming out of that is matching to the real distribution”
Games reveal deep aspects of culture: chess occupies the same intellectual echelon in the West that Go does in Japan, China, and Korea, and analyzing a culture's games reveals what that culture thinks about strategy, warfare, and other fundamental concepts.
“in terms of go, it's what they play in Japan and China and Korea. In Asia, it sort of occupies the echelon. Chess does in the west. And you can really um get deep into uh uh what a culture really thinks about things. It's sort of embodied into their games. Um including the way they think about strategy, warfare, all of these things are embodied in some of the rules of these games.”
Academia should not try to compete with companies in building large foundation models, but should instead focus on understanding what these models do, interpreting them, building benchmarks to constrain behavior, and addressing philosophical questions about AI's future impact—areas where academia is better positioned than industry.
“what I would what I suggest to my colleagues in academia and we talked about this earlier is there are many things that academia should be doing orthogonal to that. So don't try and build, you know, places like us, we're spending billions of dollars to build the machines to then with amazing engineers, world-class engineers and research to build these Gemini foundation, you know, top foundation models, but they're available for pennies on the dollar, you know, for anyone to run.”
Even if one were very lucky, studied hard, and became a genius like Richard Feynman, there would still be too much unknown for any individual mind to solve, so a better approach is to build a tool to help the best scientists make discoveries.
“even if you were very lucky and you studied hard and um you know maybe one could could could uh do the sorts of things that you know you know one could only dream about being someone like Richard Feman or with his genius. And even yet there was still so much we didn't know or we wouldn't be able to know even with th those kinds of minds working on it. So I sort of thought well maybe a better option would be to build a tool that could help us um and help the best scientists in the world including myself make those discoveries.”
Hassabis worries that there aren't many other problems of similar importance to protein folding that have such clean, well-established experimental datasets, making future AI applications to more complex biological systems potentially constrained by data availability.
“I do worry about that and there aren't many problems where that of that sort of uh importance that have that cleaner data set to work from and that was one of the reasons we I picked protein folding and had that in mind.”
Hassabis views DeepMind's entire enterprise as being 'Alan Turing's champion'—investigating what classical computation can actually accomplish, following Turing's work on Turing machines and the Church-Turing thesis.
“I think we're investigating in our own way because one of the things you can think of what we've done with deep mind and I would say my whole career is I sort of think of ourselves as Alan Turing's champion.”
Most interesting things in nature—biological systems, geological weathering, cosmological systems—have undergone some form of evolutionary process (broadly construed) and have achieved temporal and spatial stability, meaning they have structure that is not random and can be learned.
“my sort of proposal is that most interesting things in nature most natural systems have gone through some kind of process of evolution and I mean that very generally I don't mean just life but I mean it could be geological weathering could be even cosmological you know the shapes of planets and what the orbits and things like that they they've they've become stable over time right um they've survived sort of spatial temporal stability otherwise they wouldn't exist as entities um and that means means there is some structure there that um is you know not random that is uh not uniform that one can perhaps learn given enough examples.”
The deep concern about planning for success from DeepMind's founding was that if the ambitious goal of AI development succeeded and became influential and transformative, there would be attendant risks requiring careful management.
“we had this very ambitious mission in mind and we actually planned for success even though if you wind your mind back to 2010 nobody in industry was there was no there was no we could barely get any money for this you know uh starting the company no one in academia was very small pockets of academia were working on this”
Building an intelligent artifact and deconstructing it with the scientific method to compare it to the human brain will tell us a lot about what is special or not special about the human mind, making the study of artificial intelligence an important way to understand consciousness.
“it's a fascinating topic in itself a fascinating intellectual pursuit in itself the building of an intelligent artifact. Uh the distillation of intelligence into a machine then comparing it to another great mystery which is the workings of the human mind and the nature of consciousness. And it also felt to me and I did obviously I did a degree in neuroscience and computer science is that um I always felt that trying to build an intelligent artifact and then being able to deconstruct that with the scientific method and comparing it to the human brain will tell us a lot about what's special or not about the the the human mind.”
Finding a room-temperature superconductor material (assuming one exists in the physical possibility space) is an example of a problem that could potentially be solved using DeepMind's approach of building models to guide search through enormous material spaces.
“Another example is what we're doing in drug discovery. Now we know the structure of the protein. Can you design a compound uh that binds to the right part of the protein but to nothing else in the body? because if it binds into anything else then that's like toxicity. So you don't want that.”
The most important non-technological steps society should take as AI develops include creating new institutions, international collaborative frameworks similar to CERN, monitoring bodies like the IAEA, and wise governance structures that represent the world.
“I think we need new institutions. Mhm. So, um I was actually discussing it with uh some of the other Nobel winners in the in the ceremony in Sweden. Um the economists that won it this year uh are all in, you know, experts in in institutions and the power of institutions if you build them right. And I sort of said to them, maybe you should spend some time on thinking through what we need for AI.”
Mathematical proofs can be framed as search problems similar to game moves, where each step transforms an equation or formula and the goal is to optimize toward an elegantly simplified solution, making formal logic systems like Lean suitable for translating natural language mathematical problems into verifiable formal proofs.
“if you think about um trying to solve a maths conjecture or something like that, um you know, one way you can think about it is that you have some equation or some formula you're trying to um you know, trying to optimize or reduce down um and find a solution to some problem and you can adjust that formula in some way as the next step. And you can almost think about that as the next move in a game. and you're trying you have some metric you're trying to reach uh or you're trying to optimize about the elegance of that or what it can describe as your sort of guiding goal”
Hassabis began his chess career at age four with the intention of becoming a professional chess player, but was intellectually captivated by the problem of improving his own thinking and decision-making processes rather than just winning games.
“Um, started with chess. Uh, and really I started playing when I was four years old and very seriously and I was going to become a professional chess player. Um, but really it got me thinking about the the about the process of thinking. So, as when you're a kid and you're trying, you know, playing for the England junior teams and so on, you're trying to improve your own thought processes, your own decision-m”
Hassabis was fascinated from childhood by the biggest scientific questions—the nature of reality, consciousness, and unified theory of physics—and this motivated his decision to build AI tools to help scientists solve these problems rather than trying to become a scientist himself.
“other than apart from um my you know obsession and professional training in games and then also uh loving computers the other thing I was fascinated about was all the biggest questions. So I I'd voraciously read you know both sci-fi but also biographies of of the great scientists and books on them. Richard Feman was one of my all-time heroes”
DeepMind has been thinking about risks and planning for success since its founding in 2010, even when AI success seemed unlikely and the company could barely secure funding, because Hassabis and colleagues anticipated that if their ambitious mission succeeded, the implications would be transformative and potentially dangerous.
“when you know even when we were starting Deep Mind even before that we we had this very ambitious mission in mind and we actually planned for success even though if you wind your mind back to 2010 nobody in industry was there was no there was no we could barely get any money for this you know uh starting the company no one in academia was very small pockets of academia were working on this people like Jeff Hinton um so it was really nent and no one really thought it would be successful uccessful”
Hassabis made AI the core gameplay component of computer games like Theme Park, creating simulation games where intelligent characters adapted to player behavior, which demonstrated the power of AI when coupled with interesting game design.
“all the games I made like theme park they had AI as the core gameplay component. So they were simulation games usually with intelligent characters that had to react to um the way the player played and that's why the the games I some of the games I worked on became very successful because every player had an individual experience different experience with the game because the game adapted to how you played it with it”
DeepMind attracts multi-disciplinary talent (technologists, artists, designers, musicians) modeled on his experience with computer game design, because these creative collaborations produce the highest level of innovation and breakthroughs.
“I love multi-disiplinary environments like the IAS and I've and our deep mind is one of those environments and I've always tried to work in those kind of environments and not just with technologists but also artists and designers and those things and that's one of the great trainings computer game design does because you work with artists, engineers, u musicians and so on all together. It's a really amazing creative endeavor at the highest level”
When Hassabis and co-founder Shane Legg were postdocs at UCL, Legg wanted to pursue AI in academia, but Hassabis argued that resources and bureaucracy in academia would delay progress by decades relative to founding a company, which proved prescient.
“You know I used to say to my my one of my co-founders Shane leg we were both at the at UCL as postocs at the time you know he wanted to do it in academia but I said that it's going to you know we'll be like 50 actually sadly the age I'm sort of out now before they give us any resources to you know to actually pursue this right when this is when we were in our late 20s and early 30s and I thought like we can accelerate it 10x”
The moderator notes that Institute economist colleagues won the Nobel Prize this year for work on institutions, and suggests that institutions scholars should work on governance questions relevant to AI—a suggestion Hassabis endorses.
“I was actually discussing it with uh some of the other Nobel winners in the in the ceremony in Sweden. Um the economists that won it this year uh are all in, you know, experts in in institutions and the power of institutions if you build them right. And I sort of said to them, maybe you should spend some time on thinking through what we need for AI.”
Games were never the end goal for DeepMind; they were a means to develop general learning algorithms that could be applied to real-world problems in science, medicine, and mathematics—which was the true mission from the beginning.
“games was never an end in itself right it was it was it was a sort of means to an end. So we wanted to build as you read out our original mission statement from deep mind you know these general learning systems that could generalize and then help solve really challenging real world problems that matter.”
Human culture is the collective output of our collective brains, so the achievements of culture (747 planes, Manhattan, etc.) are not magic but rather demonstrate what classical human brains can produce when organized collectively with accumulated knowledge.
“culture is the is the is the output of collective our collective brains, right? It's not it's not magic. So it's pretty astounding and it also I think speaks to the extreme generality of our minds our human minds”
Hassabis and his co-founder Shane Legg estimated that pursuing AI in academia would have delayed their progress by ~10x—they estimated they wouldn't have resources to seriously pursue the work until they were about 50 years old if they stayed in academia.
“I used to say to my my one of my co-founders Shane leg we were both at the at UCL as postocs at the time you know he wanted to do it in academia but I said that it's going to you know we'll be like 50 actually sadly the age I'm sort of out now before they give us any resources to you know to actually pursue this right when this is when we were in our late 20s and early 30s”
A successful AI system for scientific problems requires three characteristics: substantial real data (supplemented where needed with verified synthetic data to maintain distribution matching), a clear metric that the system can optimize against (such as minimizing free energy in a system), and a massive combinatorial search space too large for brute force methods.
“you have some data hopefully a lot of data um maybe you supplement it with some simulated data some synthetic data but you usually need some real data in order to create the simulation in the first place and to also make sure that your simulation or your synthetic data the distribution coming out of that is matching to the real distribution”
The conjecture has a necessary caveat: it applies 'given sufficient data and to a certain level of resolution,' and it does not apply to truly random or uniform patterns or to problems that are inherently non-learnable like factoring large numbers (which may require quantum computers).
“it might be possible to build a model of that natural system in which case um if you're trying to find a particular state that it's in or particular solution to some problem within that natural system you know the classic needle in the haystack uh type of um uh uh uh solution that you need... these kinds of systems that I'm describing uh may be suitable to do that”
Games were never an end in themselves for DeepMind but a means to develop general learning algorithms that could generalize to real-world problems in science, medicine, and mathematics—the true goal has always been advancing human knowledge.
“games was never an end in itself right it was it was it was a sort of means to an end. So we wanted to build as you read out our original mission statement from deep mind you know these general learning systems that could generalize and then help solve really challenging real world problems that matter. So games were the kind of on-ramp to develop those types of general algorithms. But we were only interested in developing algorithms that not just were good at the game, but actually we thought could generalize. So there was no point building like an expert system like deep blue to just win at chess because it would it was not generalizable to anything else.”
There might be no natural limits to what Turing machines can compute, but there could be human-created abstract limitations—for example, factorizing large numbers or generating random noise may be impossible without quantum systems because such problems lack structure for a model to learn from.
“Well, there may be no no natural limit but of course there could be um there could be human created abstractions. So this doesn't mean it could describe uh everything in mathematics or random noise or things like that because or you know maybe not even factoriize large numbers because um there has to be a pattern or that that a model can efficiently learn otherwise you can't guide the search. If it's truly uniform or random, then you would then then you have no alternative but to brute force it and then a classical system can't work. You need a quantum system, right?”
Potential applications of the conjecture's framework include finding room-temperature superconductor materials (if they exist) and designing drug compounds that bind to target proteins without binding to other molecules in the body, both framed as guided searches through enormous combinatorial spaces of material or molecular possibilities.
“finding uh room temperature superconductor material assuming that exists in physics. One of these types of processes might be able to do that. Another example is what we're doing in drug discovery. Now we know the structure of the protein. Can you design a compound uh that binds to the right part of the protein but to nothing else in the body?”
Hassabis carried the idea of applying AI to protein folding for nearly 20 years from when he first heard about it as an undergraduate until after AlphaGo succeeded, showing the long incubation period of major scientific ideas.
“I carried that around with me for, I guess, nearly 20 years until we did Alph Go.”
Nermberg's observation that Hassabis's life story mirrors the biography in Benjamin Labatut's novel The Maniac, which fictionalized John von Neumann's life and ideas, and concluded with a fictional account of Hassabis and AlphaGo depicting it as a 'game over' moment for humanity, highlighting parallels between the two transformative figures.
“In fact the writer Benjamin Labatut has already produced such a fiction. His recent novel, The Maniac, is a fictional biography of John Fonoyman and his ideas, exploring how those ideas have unmurrered the world. The Maniac concludes with a biography, fictional, I don't know. We can ask Sir Demis. Uh, and an account, a biography of Sir Demis entitled Brainchild and an account of Alph Go that depicts the powers of that program as a terrifying game over for humanity.”
Project Astra is DeepMind's program building systems close to creating universal digital assistants that understand context and the world, with prototypes nearly ready.
“and we're very very close to doing that uh with our project Astra program and um even very recently we've we created models our main set of models is called Gemini most powerful models in the world now but we also have side projects uh where like VO there's a sort of chewing test of videos”
P refers to problems that can be solved in polynomial time, while NP refers to problems that are not tractable to solve in reasonable time on a classical computer, though problems in NP have solutions that can be verified if you already have them.
“the P the things that are you know problems that you can categorize in P what that means is stands a polomial means it can actually be solved in some sort of tractable amount of time and then the ones in NP you can think of as they're sort of not possible not tractable to solve in some reasonable amount of time at least on a classical computer”
Significant progress toward unified theories of physics appeared to stall after the early-to-mid 20th century, despite the amazing work of that era, prompting Hassabis to seek alternative approaches through technology rather than individual genius.
“if you look later in the 80s and '9s had we made much progress towards uh this unified theory and maybe the people would disagree with me in the audience but but I I I actually felt from reading Steven Weinberg's book that that we sort of hadn't it was a little bit disappointing relative to the the the amazing work that we done in the early part of the century”
Hassabis bought his first personal computer (ZX Spectrum, then Commodore Amiga) with winnings from chess tournaments, and these computers enabled him to teach himself programming and combine his loves of games and computation.
“with some early winnings from some chess tournaments, I bought my first home computer in in the UK. There was a big uh home computer hobby boom. It was ZedX Spectrum and then a Commodore Amigga. And that's when I started um programming”
The Institute for Advanced Study is recruiting Alan Dinelo Nelson to advise on AI policy, demonstrating institutional commitment to engaging with governance questions.
“But on this they got it right because they recruited our own Alandre Nelson to help advise them on their tech on their AI policy.”
The success of neural networks in addressing computationally complex problems was not surprising to Hassabis because it was the whole point of his attempt to build general learning systems, which he had hoped would succeed.
“I think the the the what's surprising is um so in some ways I'm not surprised because this was the whole point of the attempt of what we were trying to do to build these general learning systems. Why would there why would you even have hope that this could be possible?”
Hassabis intends to work on refining and making his conjecture more mathematically precise over the next few years, even without a sabbatical, perhaps working on it in spare time.
“I want to work on on this conjecture, refining it and and making it uh perhaps making it more mathematically precise over the next few years. Oh, so even before a sbatical. Yeah. Well, ideally a sbatical would help, but uh maybe in my spare time at, you know, 3:00 a.m.”