Folding Science / DAVOS 2025

Summary

This discussion focused on the intersection of artificial intelligence (AI) and biology, particularly in the context of scientific research and drug discovery. The conversation featured Demis Hassabis, co-founder of Google DeepMind, and Ardem Patapoutian, a neuroscientist at Scripps Research Institute.

The discussion highlighted the revolutionary impact of AI tools like AlphaFold in predicting protein structures, which has significantly accelerated research in structural biology and drug discovery. Patapoutian described how AlphaFold has transformed his work, allowing rapid analysis of protein structures and mutations that previously took years to study.

Hassabis outlined DeepMind’s vision for expanding AI’s role in biology, including simulating entire cells and revolutionizing the drug discovery process. He emphasized the potential for AI to dramatically reduce the time and cost of developing new drugs, potentially bringing the process down from years to months.

Both speakers discussed the challenges and opportunities in applying AI to neuroscience, particularly in understanding complex brain functions and behaviors. They noted the potential for AI to help analyze large datasets of neuronal activity and connectivity.

The conversation also touched on philosophical questions about intelligence and cognition. Hassabis suggested that some long-held beliefs about language acquisition and embodied cognition might need to be reconsidered in light of recent AI advancements.

While acknowledging the rapid progress and excitement surrounding AI, Hassabis cautioned against overhyping near-term capabilities. He suggested that while AI is delivering impressive results now, its full transformative potential may take 5-10 years to be fully realized. Both speakers expressed optimism about AI’s future impact on scientific research and our understanding of biology and the brain.

Keypoints

Major discussion points:

– The current and potential future impacts of AI on biology and drug discovery

– How AI tools like AlphaFold are revolutionizing structural biology and accelerating research

– The intersection of AI and neuroscience, including efforts to understand the brain

– Debates around embodied cognition and how AI is challenging some assumptions about intelligence

– Predictions and timelines for achieving artificial general intelligence (AGI)

Overall purpose:

The goal of this discussion was to explore the convergence of AI and biology, examining how AI tools are transforming biological research and drug discovery, while also considering how developments in AI may shape our understanding of intelligence and the brain.

Tone:

The overall tone was enthusiastic and optimistic about the potential of AI in biology, while still maintaining a measured, scientific perspective. The speakers were excited about recent advances but also acknowledged limitations and areas needing further breakthroughs. The tone remained consistent throughout, balancing excitement for AI’s potential with realistic assessments of current capabilities and challenges.

Speakers

– Alison Snyder

Role: Moderator

– Demis Hassabis

Role: Co-founder and CEO of Google DeepMind

Expertise: Artificial Intelligence, AlphaFold

– Ardem Patapoutian

Role: Professor of Neuroscience at Scripps Research Institute

Expertise: Neuroscience, sense of touch, molecular basis of pain and inflammation

Additional speakers:

None identified

The Convergence of Artificial Intelligence and Biology: A Revolutionary Frontier

This discussion, moderated by Alison Snyder, featured Demis Hassabis, co-founder and CEO of Google DeepMind, and Ardem Patapoutian, Professor of Neuroscience at Scripps Research Institute and Nobel Laureate in Physiology or Medicine. The conversation explored the transformative impact of artificial intelligence (AI) on biology, drug discovery, and our understanding of the brain.

Background of the Speakers

Demis Hassabis, introduced as a pioneer in AI, founded DeepMind with the goal of solving intelligence and using it to accelerate scientific discovery. His background spans neuroscience, computer science, and gaming. Ardem Patapoutian, known for his groundbreaking work on touch and pain sensation, brings a perspective as a “hobbyist” in AI, offering insights from the biological sciences.

AI’s Revolutionary Impact on Structural Biology and Drug Discovery

Both speakers enthusiastically agreed on the revolutionary impact of AI tools like AlphaFold in predicting protein structures. Hassabis highlighted how AI has accelerated protein structure prediction by orders of magnitude, with AlphaFold’s performance validated through the CASP competition. Patapoutian described how AlphaFold has transformed his work, allowing rapid analysis of protein structures and mutations that previously took years to study.

The potential of AI to dramatically speed up the drug discovery process was a key point of discussion. Hassabis proposed an ambitious vision of reducing the timeline from target identification to clinical trials from years to months or even weeks. Patapoutian concurred, sharing his experience with virtual screening to find drug candidates more efficiently. This shared optimism about AI’s potential to accelerate drug discovery underscores the transformative power of AI in pharmaceutical research.

Future Directions for AI in Biology

The conversation then turned to the future applications of AI in biology. Hassabis outlined DeepMind’s vision for expanding AI’s role, including simulating entire cells and revolutionizing the drug discovery process. He spoke about moving towards simulating dynamic biological interactions and pathways, with the ultimate goal of creating a “virtual cell” simulation.

Patapoutian complemented this vision by discussing the potential for AI to understand protein localization and function in cellular contexts. Both speakers expressed excitement about applying AI to analyze large datasets of neuronal activity, highlighting the synergy between AI and neuroscience research.

The Intersection of AI and Neuroscience

The discussion delved into the relationship between AI and neuroscience, revealing some thought-provoking insights. Hassabis noted the surprising effectiveness of simple algorithmic ideas in scaling AI capabilities, challenging traditional theories in neuroscience and linguistics. He suggested that some long-held beliefs about language acquisition and embodied cognition might need to be reconsidered in light of recent AI advancements, particularly the success of large language models.

Patapoutian emphasized the potential of AI to help understand complex brain functions and consciousness. He highlighted the current limitations in predicting behavior from neuronal activity in complex animals and suggested that AI could be instrumental in bridging this gap. The potential of AI in developing brain-computer interfaces for paralyzed patients was also discussed, showcasing the practical applications of this technology in neuroscience.

Artificial General Intelligence (AGI) Development

The conversation touched upon the development of Artificial General Intelligence (AGI). Hassabis described human intelligence as the initial benchmark for AGI development but cautioned against overhyping near-term AGI predictions. He suggested that while AI is delivering impressive results now, its full transformative potential may take 5-10 years to be fully realized. Hassabis emphasized the need for breakthroughs in reasoning and creativity for true AGI, referencing Turing machines as a foundational concept.

Patapoutian, while not directly addressing AGI timelines, expressed excitement about AI’s unpredictable future impacts on science. This subtle difference in perspective highlighted the speakers’ different backgrounds, with Hassabis providing a more cautious, AI-developer view and Patapoutian focusing on the potential scientific applications.

Challenges and Opportunities in Data Generation and Utilization

Both speakers acknowledged the challenges in applying AI to biology and neuroscience, particularly regarding data acquisition and quality. Patapoutian raised the question of data availability, given that much of the relevant data in drug discovery is proprietary. Hassabis responded by discussing creative solutions, such as using a combination of public data, specially generated data, and synthetic data to train AI models for biology applications. He also highlighted DeepMind’s efforts to improve data efficiency in AI models.

Patapoutian introduced the exciting possibility of AI helping scientists generate new research ideas, potentially accelerating the pace of scientific discovery. This concept resonated with both speakers as a promising direction for AI in science.

Conclusion

This discussion highlighted the exciting convergence of AI and biology, demonstrating how AI tools are transforming biological research and drug discovery. It also explored how developments in AI may shape our understanding of intelligence and the brain, challenging some traditional views in neuroscience and linguistics. While acknowledging the rapid progress and excitement surrounding AI, the speakers maintained a balanced view, recognizing both the current limitations and the vast potential for future breakthroughs. As AI continues to evolve, its impact on biology and neuroscience promises to yield profound insights and advancements in the years to come, with the potential to accelerate scientific discovery and address complex biological challenges.

Alison Snyder: Thank you all for being here this morning. Thank you to those of you watching online. In industry the buzz around AI has largely centered on the technology’s capacity to streamline business along with the possibility that it might advance towards a super intelligence. But AI is also showing its power as a tool for science in particular for tackling the complex realm of biology. And this morning we’re going to talk about what’s coming for the convergence of AI and biology and how different the future of therapeutics might look. And to do that I’m joined by Demis Hassabis who is the co-founder and CEO of Google DeepMind which is developing AI technologies including tools to solve complex sort of hard to crack math science and engineering problems with the ultimate goal of building artificial general intelligence. He received the Nobel Prize in chemistry just a few months ago for his work on one of those tools called AlphaFold which solved one of biology’s longstanding problems which is how to predict three-dimensional structure of proteins and other molecules. And next to him is Ardem Patapoutian. He is a professor of neuroscience at the Scripps Research Institute. He also received the Nobel Prize in medicine in 2021 for his research which focuses on the sense of touch and how we sort of perceive our bodies in space by looking at the molecules that mediate those processes including the molecular basis of pain and inflammation and he’s trying to discover potential new medicines for treating pain. So thank you both for being here. I don’t think there’s a statute of limitations on wishing someone congratulations for the Nobel Prize so congratulations to you both. And I wanted to start actually just big picture for a second because there have been some very public predictions recently that far more sophisticated AI will be coming to us this year and so my first question to you is should we expect something big at the intersection of AI and biology?

Demis Hassabis: Well maybe I’ll kick off. It’s great to be here and I think this year is going to be very exciting. I mean the last few years have been very exciting and each year seems to seem to beat the previous one and I’m pretty sure that will also happen this year. The big advances people are expecting in the general AI space are agent based systems systems that are able to kind of complete tasks on their own on behalf of the user. I think it’s going to be interesting bringing that towards science but I think AI applied to science is a lot richer than just the language models and things like AlphaFold. These are kind of bespoke models that are built using the same principles as the general models but are then applied specifically to particular domains. And I hope I mean we work we and others are working on trying to design drugs with AI and I think with our little spin out company Isomorphic I think we’ll hopefully have some AI design drugs in the clinical trials by the end of the year. That’s the plan.

Alison Snyder: I want to ask you both. You’re both Nobel Prize winners so you know at this point like what about the brain or intelligence has sort of framed your work? What question is it that sort of you know maybe not every day but sort of frames your work that you do and drives you so maybe I’ll start with you Ardem.

Ardem Patapoutian: I guess first I should say thanks for having me here. I looked at this session and I felt like it’s a very contrasting one. You have the world’s expert on AI and using it for biology and you’ve got this hobbyist who’s a neuroscientist that’s kind of fascinated by using it in his research and but almost an outsider so I think it’s a good perspective. I think with respect to the intersection of neuroscience and AI both very exciting just the alpha fold and I’m sure we’ll talk more about it has made such a huge impact just not in basic neuroscience but drug discovery. We are also doing that and happy to talk about it but overall neuroscience is very excited about AI. I think specifically from the perspective of despite decades of research we still really don’t understand how the brain works. There’s a lot of efforts on making sure we understand all the connections. We can look at which neurons fire at the same time but despite all of this work if you look at pattern of neurons firing in any animal model and say can we predict what behavior is going to come next? Maybe we can do it in C. elegans the worm that have only 300 neurons but anything complex like mammals we have absolutely no idea and that’s been one of the holy grails of neuroscience not just to predict behavior but as we talk about more complex thoughts, intelligence, consciousness where is that where does this come from and I think AI really has the potential to answer this because that’s what it’s very good at generate lots of data to let it come up with hypotheses and I know a lot of neuroscientists are very excited about this for the future.

Alison Snyder: Can I ask you the same question what is sort of the big question about either the brain or intelligence that sort of drives you?

Demis Hassabis: Well it’s interesting that you know that the early days of this sort of general AI development certainly the early days of DeepMind back in 2010 when we started we were inspired a lot by neuroscience sort of architectures that we knew about the brain algorithms that we knew the brain was using things like reinforcement learning so we knew that in the in the limit some of this would scale if we could get that right to a general intelligence obviously the human mind being the example but then in obviously in the last sort of five six years it’s it’s flipped over in the AI world to be more engineering heavy and less it less reliant on neuroscience ideas but so now I think that and the idea was always to reverse that at some point and then have AI systems that was sophisticated enough and we’re probably getting to that point now where we could apply it back to neuroscience and understanding the brain further and I think Ardem’s right like we you know I’ve always felt you need a tool like AI to understand a lot of phenomena in biology whether that’s a virtual cell or the brain because it’s so complex I just think it’s difficult to you know describe it just with mathematical equations and I think AI is exactly the right kind of tool and right approach to understand very complicated systems with weak interactions dynamic emergent systems that you know have incredible enormous complexity and obviously a ton of data and then you’ve got to find the structure and insight in that data and I think AI is the perfect tool for that. Would you say biology is like uniquely suited for AI? I think in in in some sense yes I mean there are other fields too that we’re very interested in chemistry mathematics and other things but I think biology is specifically difficult for other approaches to crack and and and maybe potentially specifically suitable for AI especially things like the amount of data you can generate and and then and then the test and hypothesis test.

Alison Snyder: You’ve described yourself as a self-described as a an avid AlphaFold user can you talk a little bit about how you’ve used it for your work in particular?

Ardem Patapoutian: Yeah I think it’s one of the most amazing quick advance advancements in science I’ve ever experienced just to lay the groundwork when I started as when I was PhD students and postdoc 25 years ago what we call structural biology which is you know for a while we know the sequence of human proteins but reading the sequence doesn’t tell you what it does and doesn’t tell you what it looks like how it’s folded in space so to find out the structure of a protein it was five years of work of a of a graduate student for their PhD to do this and even when you talk about membrane proteins which is what I studied is ion channels and if you’re interested in this I plug my session on Thursday that in this room I think I’ll talk about how you sense touch so we found a sensor of how it senses pressure and for membrane proteins you couldn’t even no one knew how to find this the structure of these proteins it was a complete mystery and then you know a few years back there was this new technology called cryo EM which came about which experimentally you can use microscopy electron microscopy to look at and identify the structure of this protein and became a boom in biology everybody was hiring people who know how to use this technology and so within my lifetime first it went from you can’t do it to maybe you can do it but it will take five years to do it and all of a sudden alpha fold comes in where you just type in the sequence and it tells you the structure so I mean I just you know I had to say all this background to give you the idea of how unique of a moment it is in biology where where where where you can do this not only that right now for example there are mutations in my protein pieces that cause we call gain of function which means it causes too much activity it has one mutation that changes the structure and the function of the protein and we can now without doing the experiment take that mutation and ask alpha fold tree how does the structure change because of this one change and we get to know immediately otherwise it would have taken at least a year or two of research so it’s really mind-boggling and this is not this is the basic biology part what we also do is we know targeting this large molecule with a small drug like molecule could be beneficial for pain and we can use what we call virtual screening to find molecules that would block this channel that could be developed as a drug discovery and once again ten years ago we tried to do this in actual screening, I was associated with Novartis, the pharma company, and we screened three million compounds to find small molecules that would block this channel, and we couldn’t. And now by virtual screening against alpha fold structures, we’re starting to find them. So truly revolutionizing both basic understanding of biology as well as drug discovery. And it’s probably just the tip of the iceberg of how it could help the drug discovery enterprise.

Alison Snyder: So alpha fold’s been used by at least thousands of scientists, I think millions of citations that involve it. So I’m curious for you, what are some of the next problems in biology that you sort of want to tackle or the scientific tools that you want to sort of pursue?

Demis Hassabis: Yeah, I mean, it’s really amazing. And I never get tired of hearing how amazing scientists like R.M. use alpha fold. It just continues to amaze us, and it’s so gratifying to, that’s what we were hoping with building a tool like that and putting out there into the world. And it’s been phenomenally fast, actually, how it’s been taken up. There’s over two million researchers around the world. And the way we sort of see it, and just sort of building what Arden was saying, is like, I’ve always thought, can you do the search part of biology or the experimental sciences in silico? That’s the thing that takes all the time. And then the final step, of course, you still need to validate it in the wet lab, or do a clinical trial or whatever is to make sure that your model prediction is correct, because it’s not always correct. But if that works in general, that could save you 10x, 100x of the time and cost, clearly, with virtual screening, things like that. And the number I always like to give is that not only did we, you can fold, obviously, a protein and get the structure in a few seconds, but so we ended up folding all of them, 200 million known to science. So that would have taken like a rule of thumb of like a billion years of PhD time, because five years per protein. And so it’s kind of mind-blowing, really, how much faster that could be. And then what I’m thinking is that’s not the only, of course, for drug discovery or something like that, that’s just one piece of the puzzle, like knowing the 3D structure of the protein. And so then I was sort of thinking, why can’t we revolutionize the whole of the drug discovery process, bring it down from many, many years, a decade or something, to do the average time it takes to go from target to a candidate and put it into the clinic? Why can’t we get that down to maybe months or weeks? The same kind of acceleration we saw with the structures part. And so that’s what we’ve been working on now. You’ve got the structure. Now can you design a chemical compound? So we’re sort of moving into chemistry that binds to the right part of the protein surface, but also, importantly, doesn’t bind to anything else. So you can do a virtual screen for not only, I want the most potent thing to bind to the target in question, but I want it to be really clean drug that doesn’t bind to anything else in the body. And I think eventually you could imagine personalized medicine where it’s optimized maybe overnight by an AI system for your personal metabolism to be perfect for you. And so then in more general fundamental biology, I think the way we’re thinking about it, and you see that with AlphaFold3, our work there is moving up the interaction stack. So if you think of AlphaFold2 is essentially cracked the picture of a static protein. But biology is not static. All the interesting things are when there’s dynamics and interactions happen. And AlphaFold3 is the first sort of next step on that where it’s got pairwise interactions. A protein interacts with another protein, or a protein interacts with a ligand, or a protein interacts with RNA, DNA. And then you might think about a pathway. And then eventually, my dream would be to simulate a virtual cell.

Alison Snyder: What could you do with a tool like that, with a virtual cell?

Ardem Patapoutian: You know, there’s this linking back to structural biology. The way each protein is, the structure is found, you pull it out of, this is the experimental part, you pull it out of the cell in a very artificial circumstance, and you look at what it looks like. There’s a whole field of cryo-EM tomography now, which the whole idea is to look at ultra structure of proteins, et cetera, in the native cellular environment. And so it’s remarkably very, very interesting And so it’s remarkably very little is known about that. And a lot of the analysis is biochemical, taken out of context. So I think seeing the cell as is, and where things are, both with respect to the architecture and 3D space is very, very interesting. So I’ll give you an example. Let’s say a certain protein is expressed very highly in a cell. But when you actually look at it and see it, all of it is localized at the tip of the neuron, where something very specific’s happening, that’s gonna give you a very different understanding than just levels of expression, for example. So it’d be very useful for biology. But just to, I think, continue what Demis was saying about the drug discovery process and AI, there’s so much money wasted in clinical trials for safety data, for example. And I think that’s what he was talking about, is the specificity. If you make sure it doesn’t bind other proteins, that’s what we call off-target effects. And so if you can eliminate that before you start, huge, huge positive. But if you don’t mind, I was gonna ask you, how do you make sure you have the data? Because most of this data exists in closed pharmaceutical company that they don’t really share, or one company doesn’t trust other company’s data, and large language models want a lot of data. So how do you deal with this?

Demis Hassabis: Yeah, well, look, I mean, the data question is a tricky one, and especially as you start looking at interactions, because then the amount of data to cover, you know, even an end-by-end space is sort of exponentially gets larger than the static picture. So first of all, you can generate some key data to fill in the gaps of where the public data doesn’t have it so with CROs, and so you can sort of pay money effectively to generate some specialized data. We’re also generating a lot of synthetic data, so that’s the key question. And in fact, AlphaFold needed that. We obviously started with the PDB with 150,000 or so known structures that would carefully accumulated over the last 50 years of painstaking experimental work. And we needed that as the basis, but that wasn’t enough to train AlphaFold 2. We had to generate, you know, nearly a million structural predictions and then kind of triage them to what the system thought were the most accurate, maybe 200,000, 300,000, 350,000 in the end, and put them back in to the training for the final AlphaFold 2. And it’s obviously very tricky. You’ve got to be very careful if you’re using synthetic data that it’s actually correctly representing the distribution. It’s not, you’re not, you’re not somehow training on your own errors, and you obviously need to still anchor on some held out test set that is real data. And obviously the PDB experimenters are adding to that every month, you know, maybe a few dozens of structures. So you can keep testing yourself against these unseen structures. And in fact, there was a, there’s a whole competition arranged around that called CASP, which accumulates sort of around a hundred structures every couple of years that haven’t been published yet, known. And so you can really test a very, in a sort of very careful, careful way, whether your predictions are any good. And then there’s even ideas of using sort of molecular dynamics and other computational systems that may be quite expensive to generate the initial data, but then can be fed into an AI system and we should be a lot more efficient downstream. So there’s, it’s complicated. And also the other thing is obviously developing new algorithms that are data efficient. And that’s one of the holy grails of general intelligence research. And of course the human mind does that very well by generalizing from examples. And so it’s actually a great domain for general AI research too, to be able to sort of be efficient with data.

Alison Snyder: You mentioned it before, you’re talking about speed. And I think a lot of times when people talk about AI, they talk about speed, but is it just like a scoop machine? Is it just going to take scientists to where they were going faster or is there something very fundamentally different? Should we expect more from these tools?

Demis Hassabis: Well, look, I think the first thing is acceleration of by 10 X, 100 X of many of a lot of the painstaking experimental work potentially. But I think there’s also the potential for these systems to discover new areas. And of course true invention is not possible yet with AI. Like it can’t come up with a new hypothesis or a new conjecture. It can maybe solve a complicated, say, conjecture in math. I think we’re very close to some big breakthroughs in that front. I think we’ll see that this year. But that’s different from actually coming up with the theory or the hypothesis as the best human scientists do. But there might be more brute force ways of doing that. You could imagine a system that maps out the current knowledge, but also understands when it’s at the leaf nodes, I call it, of the existing knowledge, right? And then you could do a sort of search process from there and get to a new part of the search space that has not been explored yet. And examples of that are with our program AlphaGo, which beat the world champion sort of famously at Go, Lisa Doll, in 2016 and kind of was part of creating the modern boom that we see in AI today, sort of coming of age, if you like. Not only did it win that match, but it came up with creative moves that had never been seen before, even though we played Go for thousands of years, hundreds of years and a couple of thousand years now. And I can imagine that was incredible to me because that would be so useful in science to be able to discover sort of new parts of the search space. So I can see kind of techniques like that being able to be extended into some areas of science, at least.

Alison Snyder: I want to zoom out a little bit. And AI and neuroscience have sort of always been intertwined to some degree, sometimes more so than other times. And I’m just sort of curious from both of you, what have you learned about the brain from AI, and in particular from deep learning? And then I’ll reverse that question.

Ardem Patapoutian: I think it’s, to me, again, not being an expert need to qualify that. It’s the promise of what it can do for neuroscience specifically. I mean, I think that was the first thing I talked about, is starting to make sense of these large datas of neuronal firing and connectivity that we’re gathering, and not understanding how that’s translated into action, whether it’s behavior or thought. I don’t think it’s done a major improvement there. But I think so many people are working on this, getting so close to it. But I wanted to add that I think although they will not replace scientists, also just scientists like me, I think it’s going to be very exciting to see how it can help us think of new projects and what to do. And I actually do this exercise myself every few months. I go in and ask, tell them, it already knows me that what I study, and ask what other things we should study given we study pressure sensing. And among the 10 ideas, one of them actually were something that we thought of and we started to. So it’s not that far off, I think, from being an ally to scientists and help us decide even conceptually how to study and what to study in addition to the actual application of it.

Demis Hassabis: Yeah, I think we’re early days about what it’s telling us about the brain, and specifically neuroscience data. But I’ll tell you a few surprising things, I think, on our journey over the last 15 years in doing AI research is that these, I would say, relatively simple type of algorithmic ideas, in the end, like back propagation, reinforcement learning, these sorts of things, ended up scaling to something that’s pretty impressive, the kind of multimodal foundational models you see today, without too many specialized systems. Maybe it had to be like that if you think about it. Biology’s evolved, and it’s from simple components, simple and elegant components. So perhaps that’s how it had to be. I mean, the biggest surprise to me is the way that language is being cracked, without explicit need for concepts or other things that a lot of neuroscientists, perhaps, thought would be necessary. Even if you go to some of the famous people at MIT, like Chomsky, and the kind of old school of thinking about language systems as sort of like logic systems, innate to the brain, it doesn’t seem, I feel like that needs to all be updated in a sense. Linguistics, neurolinguistics, and how the brain acquires language and concepts. It seems that from experience, it kind of just happens naturally from being immersed in it, perhaps like a child is, or the way that we’re training these large models, language models.

Alison Snyder: And one of the places I was intrigued by where your work intersects is around this idea, it used to be an idea of cognition had to be sort of embodied, right, that your intelligence was, in part, how you sort of move through the world, you get your sense information. Are you saying that that’s maybe changing?

Demis Hassabis: I think that was another hypothesis was that there’s this whole branch of neuroscience called action in perception. So with this idea is you can’t actually, in some sense, really perceive the world fully and model the world, the physics of the world, let’s say, without acting in it. And the canonical examples were like the understanding of weight. You kind of need to really act in the world to sort of get a true understanding of things like weight and the physicalness of the world. But it turns out, it seems like you can potentially learn that just from language and videos, so passive observation. Of course, I think the systems will bend. And of course, that’s what agent-based systems are going to be about this year, which is acting in the world as well. And then I think we will get even more intelligent systems. And then they will be useful for things like robotics. But I don’t think it’s necessary. It doesn’t seem like it’s necessary to acquire intelligence.

Alison Snyder: Yeah, that’s really surprising. What do you think about that as a neuroscientist?

Ardem Patapoutian: Not much to add. But just to go to the robotics side, there’s so much exciting research in neuroscience in a way of, for example, people who are paralyzed and can’t sense anything in their arms, people who’ve put in these devices like arms that have all these sensors and then connect it to parts of the brain where we know that touch perception occurs. And AI is heavily involved in this programming, these devices as well. And with some experience, these folks now start sensing this touch, which is your half robotic, half brain. And again, very exciting new improvements that AI is going to play a huge role in advancing those as well.

Alison Snyder: I have two more questions. And we’ve got four minutes. So we’re going to do it. The yardstick oftentimes for AI is human intelligence. Is that sort of the, I don’t know, is that a flawed yardstick? If we obsess about sort of these human-centric things, are we missing out on more promising or maybe more specific ideas, ways, machines that could help us?

Demis Hassabis: Well, the reason that we and my co-founder, Shane Legge, our chief scientist, are co-ing the term artificial general intelligence actually back in around 2000, the reason we sort of use human intelligence as a yardstick is the only example we have of general intelligence potentially in the universe. So it’s a very special data point, let’s call it, in the search space of intelligence. Probably not the only way to build a general intelligence. Eventually, when we start understanding intelligence at a more fundamental level, there’ll be probably other architectures that could deliver that. And maybe AI will go down a different route eventually. But initially, that’s the gold standard. And it’s the only evidence we have of that. What I actually think about more is more the theoretical Turing machines. My all-time hero is Alan Turing, kind of proved that the Turing machine, which is the basis of all modern computers, can in theory model anything that’s computable in the world. And the brain is probably a type of Turing machine. And so if we can simulate the cognitive capabilities that humans have in its totality, then you know you have a Turing-powerful system, which means, in theory, it could compute anything.

Alison Snyder: OK. My last question is, so again, circling back to where we were, some tech leaders have talked a lot about and compressed the timeline for AGI. I’ve heard you talk about how you think there’s still a lot of breakthroughs that are needed. Why do you believe that? Has that changed? Because a lot of these people, this compression has happened in the past eight weeks, even.

Demis Hassabis: Yeah, well, I mean, look, I think you need to look at the, I mean, there’s a lot of hype in the area, of course. I mean, there’s a lot of real, but the hype sort of run away with itself, as often happens in these very hot areas. It’s a bit unfortunate, I would say. So what I would say is, like, if anything, it’s AI, although it’s delivering really impressive things today, hopefully things like AlphaFold is an example, it’s over-hyped in the near term. But I think it’s still under-appreciated in the medium to long term. I still think people understand how revolutionary and transformative this technology is going to be in the sort of five, 10-year timescale. I do think one or two other things, only a handful, small handful of probably big breakthroughs are needed. It might be none. It might be just scaling from here. But I suspect that there may be one or two things that are missing, which will take more of a five-year timescale. I think those people are predicting sort of this year, next year, maybe have slight ulterior motives as to why they’re predicting it that soon, like raising a lot of money and things like this. What are the one or two other things? Well, I think we need some things in reasoning and planning. We don’t really understand how to do full creativity, like we were discussing earlier. Could you come up with general relativity, like Einstein did, based on the knowledge that he had at the time in the 1900s? Like, I don’t think any of our systems could do that, anywhere close to that. Or could you invent Go, rather than just play a great move in Go? So those are my two example, my go-to examples on where we’re still missing something.

Alison Snyder: OK, one more question. Sorry. Looking ahead to the next decade, how do you envision sort of AI shaping our understanding of the human brain, thinking about some of those advances that he’s?

Ardem Patapoutian: I’m letting Demis talk, because he’s the expert on this. I think, to me, I’m a reductionist neuroscientist, so I get nervous about definitions. I don’t even know how to define intelligence, let alone AGI. All I can say is it’s super exciting. I’m going for the ride. I feel like, as long as we’re a little bit careful, it’s a wonderful. And you know, one of the big things I say in science is that you don’t know what’s going to happen, and you can’t predict it. You do this exciting research, and people can use it in different ways, find different things. So I’m not much into predicting, but I think it’s going to be very exciting. Thank you.

Alison Snyder: Thank you both for doing this. I really appreciate it. Thank you to everyone watching online, and thank you to everyone here in the audience. Thank you.

Agreement Points

AI revolutionizing protein structure prediction and drug discovery

speakers

– Demis Hassabis
– Ardem Patapoutian

arguments

AI accelerating protein structure prediction by orders of magnitude

AlphaFold revolutionizing basic biology understanding and drug discovery

summary

Both speakers agree that AI, particularly AlphaFold, has dramatically accelerated protein structure prediction and is transforming drug discovery processes.

Potential of AI to simulate complex biological systems

speakers

– Demis Hassabis
– Ardem Patapoutian

arguments

Moving towards simulating dynamic biological interactions and pathways

Potential for AI to create a “virtual cell” simulation

Using AI to understand protein localization and function in cellular context

summary

Both speakers discuss the potential of AI to simulate and analyze complex biological systems, from protein interactions to entire cell simulations.

Similar Viewpoints

Both speakers believe that AI has the potential to significantly accelerate the drug discovery process through virtual screening and efficient compound design.

speakers

– Demis Hassabis
– Ardem Patapoutian

arguments

Potential to dramatically speed up drug discovery process from years to months

AI enabling virtual screening to find drug candidates more efficiently

Both speakers express excitement about AI’s potential to analyze large neuronal datasets and advance our understanding of complex brain functions.

speakers

– Demis Hassabis
– Ardem Patapoutian

arguments

Applying AI to analyze large datasets of neuronal activity

Potential of AI to help understand complex brain functions and consciousness

Unexpected Consensus

AI challenging traditional theories in neuroscience and linguistics

speakers

– Demis Hassabis
– Ardem Patapoutian

arguments

Challenging traditional linguistic theories through language model achievements

Surprising effectiveness of simple algorithmic ideas in scaling AI capabilities

explanation

Both speakers, despite their different backgrounds, acknowledge that AI is challenging established theories in neuroscience and linguistics, suggesting a need to update our understanding of these fields.

Overall Assessment

Summary

The speakers show strong agreement on AI’s transformative impact on biology, drug discovery, and neuroscience. They share optimism about AI’s potential to accelerate research and provide new insights into complex biological systems.

Consensus level

High level of consensus, with both speakers from different backgrounds (AI development and neuroscience) agreeing on the significant potential and current impact of AI in biological sciences. This consensus implies a strong interdisciplinary recognition of AI’s value in advancing scientific research and understanding.

Different Viewpoints

Timeline for AGI development

speakers

– Demis Hassabis
– Ardem Patapoutian

arguments

Caution against overhyping near-term AGI predictions

Excitement about AI’s unpredictable future impacts on science

summary

Hassabis expresses caution about near-term AGI predictions and suggests a longer timeline, while Patapoutian shows more openness to unpredictable rapid advancements.

Unexpected Differences

Overall Assessment

summary

The main areas of disagreement were subtle and primarily related to the timeline and specific applications of AI in biology and AGI development.

difference_level

The level of disagreement between the speakers was relatively low. Both generally agreed on the potential and impact of AI in biology and drug discovery. The main differences were in their emphasis and perspective, with Hassabis providing a broader, more cautious view of AGI development, while Patapoutian focused more on specific applications and expressed more openness to unpredictable advancements. These differences do not significantly impact the overall discussion on AI’s role in biology and drug discovery.

Partial Agreements

Both speakers agree on AI’s potential to revolutionize drug discovery, but Hassabis focuses more on the overall process acceleration, while Patapoutian emphasizes specific applications like virtual screening.

speakers

– Demis Hassabis
– Ardem Patapoutian

arguments

Potential to dramatically speed up drug discovery process from years to months

AlphaFold revolutionizing basic biology understanding and drug discovery

Similar Viewpoints

Both speakers believe that AI has the potential to significantly accelerate the drug discovery process through virtual screening and efficient compound design.

speakers

– Demis Hassabis
– Ardem Patapoutian

arguments

Potential to dramatically speed up drug discovery process from years to months

AI enabling virtual screening to find drug candidates more efficiently

Both speakers express excitement about AI’s potential to analyze large neuronal datasets and advance our understanding of complex brain functions.

speakers

– Demis Hassabis
– Ardem Patapoutian

arguments

Applying AI to analyze large datasets of neuronal activity

Potential of AI to help understand complex brain functions and consciousness

Key Takeaways

AI is revolutionizing biology, particularly in areas like protein structure prediction and drug discovery

Tools like AlphaFold have dramatically accelerated research timelines in structural biology

Future AI applications in biology may include simulating dynamic cellular processes and a ‘virtual cell’

AI is challenging some traditional theories in neuroscience and linguistics

There is significant potential for AI to advance understanding of complex brain functions

While AI capabilities are advancing rapidly, true artificial general intelligence likely still requires breakthroughs in areas like reasoning and creativity

Experts caution against overhyping near-term AGI predictions while acknowledging AI’s transformative long-term potential

Resolutions and Action Items

DeepMind aims to have AI-designed drugs in clinical trials by end of the year through Isomorphic Labs

Unresolved Issues

How to acquire sufficient high-quality data for AI models in drug discovery, given much data is proprietary

Whether embodied cognition is truly necessary for developing artificial intelligence

Exact timeline and requirements for achieving artificial general intelligence

Suggested Compromises

Using a combination of public data, specially generated data, and synthetic data to train AI models for biology applications

I think AI applied to science is a lot richer than just the language models and things like AlphaFold. These are kind of bespoke models that are built using the same principles as the general models but are then applied specifically to particular domains.

speaker

Demis Hassabis

reason

This comment highlights the unique potential of AI in scientific applications beyond general language models, introducing the idea of specialized AI tools for specific scientific domains.

impact

It set the tone for discussing concrete applications of AI in biology and drug discovery, moving the conversation beyond general AI capabilities.

Despite decades of research we still really don’t understand how the brain works. There’s a lot of efforts on making sure we understand all the connections. We can look at which neurons fire at the same time but despite all of this work if you look at pattern of neurons firing in any animal model and say can we predict what behavior is going to come next? Maybe we can do it in C. elegans the worm that have only 300 neurons but anything complex like mammals we have absolutely no idea and that’s been one of the holy grails of neuroscience not just to predict behavior but as we talk about more complex thoughts, intelligence, consciousness where is that where does this come from and I think AI really has the potential to answer this because that’s what it’s very good at generate lots of data to let it come up with hypotheses

speaker

Ardem Patapoutian

reason

This comment articulates a fundamental challenge in neuroscience and suggests how AI could potentially address it, bridging the gap between AI and neuroscience research.

impact

It sparked a discussion on how AI tools could be applied to complex neuroscience problems, leading to considerations of virtual cell simulations and other advanced applications.

Why can’t we revolutionize the whole of the drug discovery process, bring it down from many, many years, a decade or something, to do the average time it takes to go from target to a candidate and put it into the clinic? Why can’t we get that down to maybe months or weeks?

speaker

Demis Hassabis

reason

This comment presents a bold vision for dramatically accelerating drug discovery using AI, challenging current timelines and processes in the pharmaceutical industry.

impact

It shifted the discussion towards the transformative potential of AI in drug discovery and development, leading to exploration of specific ways AI could improve efficiency and effectiveness in this field.

The biggest surprise to me is the way that language is being cracked, without explicit need for concepts or other things that a lot of neuroscientists, perhaps, thought would be necessary. Even if you go to some of the famous people at MIT, like Chomsky, and the kind of old school of thinking about language systems as sort of like logic systems, innate to the brain, it doesn’t seem, I feel like that needs to all be updated in a sense.

speaker

Demis Hassabis

reason

This comment challenges established theories in linguistics and neuroscience based on recent AI developments, suggesting a paradigm shift in our understanding of language acquisition and processing.

impact

It prompted a reconsideration of long-held beliefs about cognition and language, leading to a discussion about the implications of AI advancements for our understanding of human intelligence and brain function.

I do think one or two other things, only a handful, small handful of probably big breakthroughs are needed. It might be none. It might be just scaling from here. But I suspect that there may be one or two things that are missing, which will take more of a five-year timescale.

speaker

Demis Hassabis

reason

This comment provides a measured perspective on the timeline for achieving artificial general intelligence (AGI), countering some of the more aggressive predictions in the field.

impact

It brought the discussion back to a more grounded assessment of AI’s current capabilities and future potential, encouraging a more nuanced view of AI development timelines.

Overall Assessment

These key comments shaped the discussion by consistently pushing the boundaries of how AI is perceived in relation to biology, neuroscience, and drug discovery. They challenged existing paradigms, introduced novel applications of AI in scientific domains, and encouraged a balanced view of AI’s current capabilities and future potential. The discussion evolved from specific applications like AlphaFold to broader implications for understanding brain function and revolutionizing drug discovery processes. Throughout, there was a tension between excitement about AI’s transformative potential and recognition of the complexities and challenges still to be overcome, leading to a nuanced and forward-looking conversation about the intersection of AI and biology.

How can AI be used to predict behavior from patterns of neuronal firing in complex animals?

speaker

Ardem Patapoutian

explanation

This is a key challenge in neuroscience that AI could potentially help solve, advancing our understanding of brain function and behavior.

How can AI be applied to understand complex phenomena in biology, such as simulating a virtual cell?

speaker

Demis Hassabis

explanation

This represents a significant potential application of AI in biology, which could revolutionize our understanding of cellular processes.

How can AI accelerate the drug discovery process, potentially reducing it from years to months or weeks?

speaker

Demis Hassabis

explanation

This could dramatically speed up the development of new treatments and personalized medicine.

How can sufficient and reliable data be obtained for AI models in drug discovery, given that much of it is held by pharmaceutical companies?

speaker

Ardem Patapoutian

explanation

This is a crucial challenge for applying AI to drug discovery, as the quality and quantity of data directly impacts the effectiveness of AI models.

How can AI be used to discover new areas of scientific research or generate new hypotheses?

speaker

Demis Hassabis

explanation

This represents a potential for AI to not just accelerate existing research but to actively contribute to the direction of scientific inquiry.

How can AI help in understanding the relationship between neuronal activity and complex behaviors or thoughts?

speaker

Ardem Patapoutian

explanation

This is a fundamental question in neuroscience that AI could potentially help answer, leading to breakthroughs in our understanding of brain function.

How might our understanding of language acquisition and processing need to be updated based on the success of large language models?

speaker

Demis Hassabis

explanation

This suggests a potential paradigm shift in our understanding of language and cognition, with implications for linguistics and neuroscience.

How can AI contribute to the development of advanced prosthetics and brain-machine interfaces?

speaker

Ardem Patapoutian

explanation

This represents a promising application of AI in medical technology, with the potential to significantly improve quality of life for people with disabilities.