The Deep View: Conversations

Ian Krietzberg:
Hey, everybody. Welcome back to the Deep View Conversations. I'm your host, Ian Kreitzberg, and today we're talking about AI for science. Now, in many respects, AI for science is at the core of AI for good. This idea of how can current models, current systems, enhance and advance scientific research, scientific processes, ranging from drug creation to material generation to climate modeling and forecasting, advanced weather prediction. At the core of all of this is an advanced prediction. And the kind of examination is how far can this prediction take us and what can this prediction allow us to do? So to break all that down, my guest today is Dr. Chris Bishop. Now, Chris is a Microsoft Technical Fellow and the Director of Microsoft Research's AI for Science Unit. He walked me through a bunch of different examples of what Microsoft Research is doing on the AI for Science front, how Microsoft thinks about the kind of cost-benefit analysis of building and deploying AI models, and what the kind of reality of what these systems can do, how they can do what they do, and the impacts that that might have on society. So this is AI for Science. Chris, thanks so much for joining me. That's a great pleasure. There's so much to talk about. I feel like the kind of intersection of AI and science, the idea of AI for good and how it's accelerating the scientists that help us accomplish the good is kind of at the forefront of a lot of people's minds. Before I get into all of that, and I'm going to get into all of it, I want to start with you. Right now you lead AI for science research at Microsoft. You've done so much work in the AI space, but you got your start in physics. And I just wonder that in many ways they seem related and in other ways they're not. And so I kind of wonder at what point you were on the physics path and you thought, you know what? computer science and AI?

Chris Bishop:
Yeah, that's a great question. I've always been fascinated by the idea of AI, going right back to, I guess, 2001, A Space Odyssey, the beautiful Kubrick, Clarke film, which depicted a very abstract form of intelligence, very different from the usual Hollywood sort of robots of the day. But the field itself was not one that interested me very much as a researcher. I was fascinated by physics. I did a PhD in quantum field theory. And then I worked on the fusion program for about eight or nine years. And it was while I was working on the fusion program that Geoffrey Hinton published his backprop paper. And I read that and started to learn more about neural nets. And that really grabbed my imagination. Because instead of traditional AI where you try to write rules that make a machine intelligent, here was a system that could learn by experience. It was mimicking the brain in some superficial way at least. That seemed very inspirational. So I started playing about with neural nets and actually started applying them to data in physics. So I was working at Cullum Lab near Oxford next to the world's largest tokamak fusion experiment. It was sort of big data of the day. We had lots of data to play with. And I started applying neural nets to fusion data. And it was a lot of fun because it was like having a new tool. I had a very non-linear technique. Most people are using linear techniques. So I was able to pick a lot of low-hanging fruit. But then I kind of got even more interested in the machine learning and the AI and started to pivot my career to focus on that primarily. So it was a transition, but via physics.

Ian Krietzberg:
Mm hmm. And I got to wonder because I'm thinking so much about a lot of the applications, the scientific applications of of the the models that we're thinking about today. A lot of what we're looking at kind of relies on other domain expertise that goes beyond computer science. And so I wonder how much you use your physics kind of grounding in your day-to-day still today.

Chris Bishop:
I used to say physics is a great place to come from if you're doing machine learning, even more so than computer science, because you learn about probability, you learn about linear algebra, you learn about calculus, and those are the fundamental tools that you need But, of course, now my career has sort of come full circle and now I'm actually using the knowledge I gained in Schrodinger's equation, when I was doing quantum field theory, or in differential equations when I was working on the fusion program. And so now I'm able, in a much more direct way, to use that knowledge and experience that I picked up as a physicist. So, for me, it's very exciting. I'm getting to work on all the things that I've worked on over my career, but now bringing them together in one place. It's a great time.

Ian Krietzberg:
It's the culmination of everything.

Chris Bishop:
It is. And it's kind of cool. I mean, Jeff, of course, who's been a sort of a mentor and a hero of mine for the last 35 years, last year won the Nobel Prize. And he got the Nobel Prize in physics. So maybe it turns out I hadn't left physics at all. After all, I was just doing a different kind of physics. So it's a lot of fun.

Ian Krietzberg:
Now, when we talk about the ways in which AI can be applied for scientific endeavors, there's a couple of, I guess, major ones that come to mind. But the thing that's interesting to me is that it's all almost targeted enhancements of stuff that we're already doing. You know, you think about AI in the biology fields and finding new molecules and the implications of that for drug research. We already have methods of doing that. And I guess the promise here is that now we can do it much more quickly. And so I wonder for you, because I guess this is different for everyone, but we think about the idea of the true promise, the guiding light of why this should be enhanced, should be advanced. What is that for you? What's the big promise on display here?

Chris Bishop:
I think that's a great point. So I think, first of all, what we're seeing right now is a very real and very dramatic acceleration. Things that we could have done before, but we can now do dramatically faster. I think there's a lot more to it than that and I think over the coming years we'll see a lot more. But it might be good just to spend a few minutes talking about that acceleration because if that's all there was, if that was it, that was the only thing that was going on, it would already be transformational. And by that, I mean, we're seeing very robustly that we can take techniques and accelerate them using AI by at least three orders of magnitude. Pretty much the first thing you try, you get a thousand times speed up for the same accuracy. And if you work at it pretty hard, you can get to maybe 10,000 times speed up. And, you know, we live in a world of millions and billions and trillions. And somebody says a thousand, that seems like a small number. So I sometimes say, well, you know, Imagine if tomorrow that $100,000 sports car, you could buy it for $100. How would that change things? So sometimes a quantitative change, when it's sufficiently big, becomes a qualitative change. It means we can do all kinds of things now that we would not have dreamt of doing before. And it's something that I see in many, many different areas of science. We've even given a rather sort of grandiose title, we call it the fifth paradigm of scientific discovery. But we've given it a name because it's a pattern that we see occur many times. I can just take a minute to explain what these paradigms are. So originally, of course, the foundation of science is empirical. We make observations about the world. I think of some cave dweller having a big rock and a small rock and letting go of both of them, expecting the big rock to land first and they land at the same time. And this is that they've made a discovery. So that's experiment. That's the first paradigm of scientific discovery. And to this day, that remains the gold standard. It's experiment that is the ultimate arbiter of truth in science. And then we have a tremendous breakthrough in really beginning the 17th century with Newton, the development of calculus and the discovery that gravitation, that dynamics could be described with incredible precision using some very simple equations, differential equations. And that got extended in the 19th century by Maxwell to the equations of electricity, magnetism, the electromagnetic field. In the 20th century, quantum physics, relativity. All of that incredible physics can be described by a handful of very simple equations. You know, you could write them on a t-shirt. And yet they describe the world with extraordinary precision, just remarkable precision, all the way from the subatomic to the scale of clusters of galaxies and everything in between. So that's the second paradigm, the theoretical. The problem with those equations, they're very simple to write down. They look deceptively simple. But when you try to solve them, you can only really solve them with a pencil and paper for kind of toy problems. And so really it wasn't until the 1960s and the development of digital computers that we had a general way of solving these equations. So that's the third paradigm, the computational. The fact that we could solve these fundamental equations of physics with great precision. And I refer to that as simulators. The first use of a digital computer was simulating physics and it remains to this day one of the most important applications of digital computing. Today we'd use it, for example, to simulate the weather. We solve some simple equations called the Navier-Stokes equations that describe fluid flows with some extra terms in. But basically we make weather forecasts by solving these differential equations on a supercomputer. But it can be very expensive to solve those equations. Then we come to the end of the 20th century, beginning of the 21st century. Another paradigm arose, again enabled by digital computing, the fourth paradigm, named by Jim Gray, a famous computer scientist who worked on very large scale compute. And he observed that scientific discovery was increasingly being driven by data, by large scale data. Think of the discovery of the Higgs particle at the Large Hadron Collider, for example, an absolute torrent of data coming off this machine and having to be filtered and sifted. And then we can learn from that data. So learning from experimental data, observational data is the fourth paradigm. And of course, machine learning thrives on large amounts of data. So machine learning plays a key role in the fourth paradigm. What we've observed in the last few years is something else, another sort of template that we see occur in many different scenarios. We call it the fifth paradigm, in which again you use computers, again you use machine learning, but this time you're not learning from experimental data from observational data, you're learning from simulated data that comes, if you like, from the third paradigm. So, you simulate data by solving the fundamental equations of physics, but treat that now not as the solution to your problem, but as training data. So, simulate, let's take an example, let's take the weather. So, weather forecasting today mostly works by solving these differential equations on a supercomputer. But instead of taking those solutions and just viewing them as the weather forecast, we can view them as training data for a machine learning system. And once that machine learning is trained, we can view that as an emulator of the simulator. And the incredible thing is that that emulator now, once it's trained, is extremely fast, at least a thousand times faster than the original simulator for the same accuracy. So now we have a model that's recently published and released from the IFA science team called Aurora, which is trained on large volumes of weather forecasting data. and can produce an accurate weather forecast thousands of times faster than that simulation. And that really is transformational. A factor of 1,000 or several thousand really changes the way you think about weather forecasting. We can produce an eight-day weather forecasting in a matter of seconds on a single GPU, for example. So that's pretty remarkable and pretty transformational. But it's not just in weather forecasting. We see that at the molecular scale. We see that in many, many different cases. And I would argue if that's the only thing that was happening in AF for science, that alone would be worth building a team and really going after this because accelerating things by factors of a thousand can really change the game, can allow you to do things that you couldn't have conceived of previously.

Ian Krietzberg:
Yeah, the Project Aurora, and there's increasing models like it, the idea of the kind of climate model, the geospatial models. And well, so your point is interesting. You know, if all it did was advance or enhance the speed at which we could do these things by a factor of, you know, whatever it turns out to be, in this case, a thousand and other factors, I guess it varies. It's enough of a breakthrough, right? But you mentioned that you expect it to do more than just enhance the speed at which we do things, and that that's going to start kind of materializing. What does that side look like?

Chris Bishop:
Right. So let's take Aurora as an example. As you say, quite a few people have built these emulators for different systems, such as for weather forecasting. Aurora is much more actually than just an emulator of a weather forecasting system. It's actually the world's first foundation model for atmospheric flows. So it's trained not just to emulate a particular simulator, but instead it's trained on a large variety of data. So there's a lot of diversity of data. And the idea is to, as it were, force the system not simply to replicate a particular simulator, but rather to gain some, I'll put it in inverted commas, understanding of the dynamics of fluid flows in the Earth's atmosphere. And why do we want to do that? The reason we want to do that is we want to generalize this to new situations. So the idea of a foundation model is you build a typically very large, quite general purpose model, and then later you can specialize it to particular applications. So a good example would be modeling the flow of pollutants around the atmosphere, particularly things like oxides of nitrogen, which are pretty unpleasant pollutants. There's a lot of data around weather. I mean, there are lots of weather stations, often people have them in their gardens and so on. There's very dense, rich data for weather. For pollutants, much less so. There's much less data, it's much sparser. If you trained a model just using the pollution data, you'd be in a very data-scarce regime. You could only build a rather small and impoverished model. But instead, if you start from a foundation model, if you start from Aurora and then fine-tune Aurora using the small amount of pollution data that you have available, it can leverage all of that knowledge, as it were, of atmospheric flows. And now you get a much better, much more accurate model of pollution flows than you could possibly get from the pollution data alone. So that's an example of where you're taking this concept of an emulator but now going into a new domain and be able to do things that you couldn't previously do.

Ian Krietzberg:
Yeah, the pollution tracking side, it brings us to an interesting point because you hear a lot, AI is such a broad term and there's so many different kind of things that are not necessarily the same that are kind of, you know, included under that umbrella when people mention stuff. And the climate promise on hand here is a really interesting one and a kind of, It highlights the cost-benefit analysis of these things to me. Because we do have, like if we're able to track, a lot of the promise of these systems is if we can track it through enhanced data, then we know what's going on, and if we know what's going on, we can mitigate. So if we know where pollutants are coming from, we can take measures to address that, hopefully. That's the idea. But if the cost of the model, which we know some of these GPUs are energy intensive chips, maybe outweighs that kind of promise. Like, I wonder how you think about that, the cost benefit of deployment, considering the impact that building these models could have.

Chris Bishop:
I think that's a great question. If you come back to that fundamental fifth paradigm, there are really three stages to it. The first one is you run the simulation, so that costs you compute. In the case of weather, of course, we benefit from the fact that that happens anyway. Weather centres every few hours launch another big supercomputer run. So there's a tremendous amount of historical data. In other contexts, let's say we're dealing with molecular simulation for drug discovery, we may have to go and generate that data specifically for a particular scenario. So that's the first phase, that's the generation of the training data. The second phase is the training of the machine learning system, and often that's computationally intensive as well. And then the third stage is the inference when you're running the train model. That's where you get the massive speed up. So it makes sense to think of this fifth paradigm if you're going to do that inference many, many times. If you just want to do one calculation, you should just run a simulator. But if you're going to use this many times, you amortize that fixed cost of generating the training data, training the simulator, training the emulator. Those first two stages represent a fixed cost. And now you can amortize that across thousands or millions of calls to the trained emulator, which is very much more efficient. So I think this is true more broadly. What we're seeking here is really knowledge. The goal of science is to gain knowledge, gain a better understanding of the world. And the cost that's incurred, whether it's computer simulations, lab experiments, whatever, some combination of the two, that we need to do to go through the scientific process and gain that better understanding, that cost, if you like, in the broader sense of cost, then gets amortized In a sense, forever more. You know, we discovered the presence of bacteria and that they cause disease. And that's human knowledge that we'll have for the rest of time and that we can leverage to cure disease and have better hygiene and save lives and all the rest. So I think that's a rather nice feature of the fifth paradigm, that we're taking some commutationally expensive processes, but now amortizing that cost over many, many downstream uses. So in that sense, it's very efficient.

Ian Krietzberg:
let's take Aurora here since we've been talking about it, the foundational model for atmospheric flow. It's trained on specific data. I wonder, and I think part of this is as people are kind of starting to gain more of an understanding about what AI is and what it means, a lot of, I feel like the public perception of AI kind of centers around and stops and starts at GPT. But something like what you're doing with Aurora is, is not the same. It's, it's done different. I wonder if you can kind of explain how this model, how Aurora functions differently. And I wonder also in the push for knowledge that you're talking about, the kind of limitations that. the generative foundation models like CHAT-GPT have, like hallucinations, I wonder how that kind of carries over to this atmospheric model.

Chris Bishop:
Right. I mean, there are differences, but also similarities as well. I mean, we're talking here about a field called machine learning and a particular approach to machine learning called deep learning. And that was a breakthrough that really took off from about 2012 onwards and has been applied to many, many different domains. I used to call the field machine learning rather than AI. So I think of intelligence is sort of one thing that we might apply to this. These terms are not really well defined, nobody can really define intelligence. But the underlying technology can be quite similar. So actually, Aurora is based on transformer technology that's actually not so different from the kind of technology that goes into chat GPT. And the idea in a model like GPT, you have a sequence of words, tokens, and you train the system to predict the next token. And that's so essentially it's sort of self-supervised. The training data doesn't need to be labeled by anybody. You just have a lot of text. and you take a whole bunch of text as input, and then the target, the thing you train the model to produce, is the next token or the next word. And we could do something similar in the case of atmospheric modeling. It's a little bit different. We don't have a one-dimensional sequence. It's what I might call sort of two-and-a-half dimensional. I mean, the Earth's atmosphere is two-dimensional to a first approximation, but then we have height is important too. So you can think of a model as being like a stack of two-dimensional videos. You know, we've all seen animations or satellite video of weather patterns swirling across the Atlantic and hitting the UK and so on. And you have things like speed, pressure, temperature at different levels in the atmosphere. So imagine this sort of stack of videos. But now you can turn those into sequences, into one-dimensional sequences, through little patches of images for each of these different layers, and sort of map it onto a problem that doesn't look too different from the one that GPT is solving. So a lot of the underlying technology actually has a lot in common.

Ian Krietzberg:
I feel like when we think about all the research side, it's kind of hard to take it from the kind of abstract to making it real. And I guess a lot of the build-out that's going on right now is working on making it real. And we were just talking about the kind of enhancements and advancements of scientific research. you know, how can we do something a thousand times faster while maintaining accuracy? But I kind of on the flip side of that, something that I've been thinking about when we see modern engineering disasters, right, accidents that happen from oil spills to I was even thinking about the Titanic, the whole, the fact that, you know, you yank the wheel and then you rip the whole sides of it open. from the iceberg and, and, you know, rising plastic pollution or, you know, even, you know, accidents and plane crashes. Are we approaching a world where these kinds of approaches, these kinds of technologies kind of offer a mitigation against some of these engineering failures that have kind of happened?

Chris Bishop:
Well, I think it's clear that over the last many decades, computation, digital computing has really transformed the engineering practice. I mean, now things are designed in the computer, they're manufactured by computer-controlled machines. Our ability to model, you talked about some extreme weather events, our ability to model those, Aurora tends to be rather good at that as well. You can also fine-tune Aurora on extreme weather events and again leverage that power of that foundation model. So I think this is part of a continuing trend and for sure these AI-based techniques are just accelerating that process. Our ability to model the world and make predictions is getting better all the time and AI is part of that story for sure, yes.

Ian Krietzberg:
And so I guess it's the question of how far can predictions go? And I would assume your impression is that they can take us pretty far.

Chris Bishop:
I'm sure there's a lot of scope for improvement and, you know, making better, more accurate predictions. Of course, some things are intrinsically difficult to predict. I mean, we know that weather systems can also be, you know, what's called chaotic, which is a very precise mathematical term. meaning that there's a tremendous sensitivity to the initial conditions. If you change the conditions of your simulation slightly, you get a very different outcome many days later for certain kinds of weather systems. And so there are some things we wouldn't expect to be able to predict. We wouldn't expect to be able to predict very precisely whether a particular storm is going to land in a particular place way ahead of time when you're in one of those chaotic scenarios. But from a practical sense, the ability to forecast things a few days ahead and so on, that's improving all the time and AI is definitely part of that mix. So while we won't ever be able to predict the future with precision in all circumstances, our ability to do so for sure is going to improve over the next few years, thanks in large part to AI.

Ian Krietzberg:
massive, large-scale prediction. It's interesting. And, you know, we've talked about Aurora, that there's a bunch of other projects that the AI for Science unit is working on. And I want to kind of dive into some of them. Before we do, I'm very curious how you go about deciding what sectors are ripe for an exploration with AI. What does that kind of process look like before you get to the point of we're putting together a model?

Chris Bishop:
That's such a great question. So, you know, in research, there are so many things to go explore. And I sometimes say ideas are cheap, right? There are many, many ideas you can pursue. And so I think part of the secret of good research is being very tasteful in deciding what to do, the very question you're asking. And I sometimes think, as researchers, we see an opportunity, we go after it very quickly, and sometimes we should just pause a little bit and say, well, you know, we only have a certain amount of time, we can only do a small number of things, what should we really focus on? And in the AFS science team, we've been very thoughtful about that. So science is a vast, vast field, and there are many areas where we could do interesting things. we've deliberately chosen to focus most of our effort at the molecular scale. And that's because of two reasons coming together. First of all, because we can see potential for very rapid and important advances at the molecular scale. But also we see a lot of practical applications and a lot of utility coming from that. And so that combination of the opportunity to make things dramatically better coupled with very clear use cases that are of great practical value make it a very natural thing to go pursue, especially in an industrial research lab where we're looking to do research that has practical utility.

Ian Krietzberg:
Is this a process that's kind of undertaken all internally? Do you go out to other domain experts at universities or whatever and identify problems? And once you identify the problem, once you say molecular, how do you narrow down and pin down this is the project that makes sense? How many projects? get started and don't go past a couple of initial phases, it doesn't pan out?

Chris Bishop:
It's a great question. We certainly have a very deliberate internal process of thinking, what are we going to work on? Because as I said, there are so many possibilities. So that's very much an internal process. But it's strongly informed by the fact that as a research organization, we're deeply connected with the international research community. So we go to conferences, we publish papers in conferences, we read papers produced by other teams at conferences, we have collaborations and partnerships with the academic world and other organizations externally. So we're strongly influenced and informed by that external landscape. And then we bring to bear our own lens, our own perspective, and we think about what opportunities do we have. And that helps inform which areas that we should particularly focus on.

Ian Krietzberg:
Now, I do want to get into some of these areas now. I feel like we've been teasing it. But so you mentioned the kind of focus on the molecular level. For starters, of the projects that are kind of ongoing, maybe in different stages of development, is there one that you personally think, this is it, this is a really crazy, really cool thing?

Chris Bishop:
Right, that's a bit like asking who's my favorite child, so it's a tricky question. But I would say there are two broad areas that we're pursuing. One of them is the design of new materials and the other is design of new small molecules for use as drugs. So the two sort of application areas, materials and drugs. They also have a lot in common and so we develop techniques that are shared across both of those, but there are also differences between them. But they both have the They both have the property, as I alluded to earlier, that we know the fundamental equations. I think maybe it was Feynman who once said, well, biology is complex chemistry, and chemistry is complex physics, and we know the equations of physics. But of course, we can't explain biology just by solving the laws of physics. There's too big a gulf between those. But nevertheless, starting from the laws of physics, we can produce models that can describe molecules and things like small molecules in a drug context or materials. in the quest to create new materials for a whole variety of applications. And through the fifth paradigm, we can accelerate our ability to solve those equations very dramatically. So let's take materials design as a good example. We've recently published papers from two projects. One is called MatterGen and one is called MatterSim. So MatterGen is an example of a generative AI model. Its goal is to propose or design new materials that have certain desired properties. So the way to think about it is this, everything is made of atoms, everything around us, our own bodies, the computer screens we're looking at right now, everything is made of atoms. And if you can configure, there aren't many different types of atoms, but you can put them together in a beyond astronomical number of new combinations. So the space of new molecules and new materials is gargantuan. Take small molecules for use as drugs. The space of those molecules has been well characterized. It's about 10 to the power of 60, one with 60 zeros. It's like the number of atoms in the solar system. It's a gargantuan space. So think of it as a vast haystack and we're looking for some needles in that haystack that might treat a particular disease. And a particular lens that we bring to that is that we can start from those fundamental equations of physics and be informed by those. And we can use that fifth paradigm to create training data to train some of those systems. So let's talk about Matagen for a moment. So Matagen is a generative model. Think of it as analogous to the kinds of AI systems that you're familiar with where you want an image for your next presentation and you type in a text prompt and it generates some images. You ask it for a mouse wearing a superhero cape or something and it generates some images. So it's going from the text prompt to the image. We want to do the same thing with materials, crystalline materials. So we want to be able to say effectively produce new material that has certain properties, certain mechanical properties, magnetic properties, electrical conductivity, whatever the properties might be. And then have the system generate lots of different candidates that hopefully all have roughly those properties. That's what MatAgen does. And the reason that's so powerful is that the The typical approach to finding new materials is through screening. You consider lots of candidates and then you sift them down through a funnel to find ones that look good. That's like searching through the haystack trying to find the needle. Think of Matagen. as like a kind of laser that tells you which region of the haystack to look in, to find good needles with the particular properties that you're looking for. So Matagen really homes in on a region of chemical space that looks very promising for, you know, you're trying to produce some magnets that have, that use less of certain kinds of rare elements or something, and it will give you lots of suggested starting points. It's not perfect. It doesn't come up with the one answer. That would be a very tough challenge. But it gives you a good place to look. What MataSim does is the opposite. It's like that emulator that screens, if you like. You can sift very quickly through lots of bits of hay to check whether one of them is a needle or not. And so MadaSim is a great example of the fifth paradigm. It's a great way of screening things thousands of times faster than you could do by running the basic physics calculations directly. MadaGen tells you which corner of the haystack to look in. So that's also like an acceleration of factors of thousand because you can now focus your screening on a much more relevant part of space. So you combine those together. Once they're working in tandem, with Matagen proposing areas to look and then Matasim screening them, then you have effectively millions, factors of millions of acceleration in terms of your ability to explore chemical space and find interesting solutions. Now, millions is still tiny compared to the gargantuan size of space, but it's a tremendous acceleration compared to where we have been even up until a couple of years ago. So just really excited about the possibilities of this kind of technology.

Ian Krietzberg:
Yeah. Uh, MatterGen is a really interesting system. What can you tell me about in the development process of MatterGen, the kind of challenges in validating its performance, making like, how, how do you determine in that early stage before you publish the paper that this is in fact, doing what we want it to do. It's doing it accurately. How do you test that?

Chris Bishop:
Yeah, that's a really key question. Because at the end of the day, you're producing, you're proposing in the computer these materials and you're claiming that they're stable and they have certain properties. Of course, in principle, we can calculate the properties of materials if we could solve those basic equations exactly. So there's a lot we can do computationally to test that we're on the right track. But ultimately, as I said earlier, the experiment is the ultimate arbiter, the ultimate ground truth in science. And so it's very important to us that we actually propose a material, have it synthesized by collaborators, have its properties measured, and confirm that it does agree with the predictions of the model. I mean, we chose a task to do with producing materials with particular mechanical properties. We specified what they should be, had Madagen propose some new materials, picked a good candidate, and then asked collaborators to actually synthesize that in the laboratory and measure it. And we found pretty good agreement, I mean, within 20% of the desired value. So we're super happy about that. So that gave us the confidence to then go public with the with the model.

Ian Krietzberg:
At this stage for something like this, what's the kind of roadmap to broader use of it? Once you've had it validated? How does it kind of get out of the Microsoft lab and into the hands of researchers who will use it to, I don't know, create a lithium ion battery that is less, I don't know, sustainably, you know, impactful than current systems, for example?

Chris Bishop:
Sure, I think there are a couple of different things we do. One thing is we often open source the models, so we just make it available generally to the community. That allows other people to pick them up, experiment with them, try different things. We can learn from that process as well. So that's one approach. The other approach is that we actually have particular partnerships. If we think there's a particularly interesting application area, we'll go and talk to a potential partner, might be an industrial partner, a laboratory, typically an organization or a team that has complementary expertise to our own. We know a lot about AI. We don't have laboratories. We don't synthesize and characterize materials directly. So actually looking for those partnerships. and then working in close collaboration with a partner to take that technology forward, usually in the context of a particular domain, so it might be batteries, whatever it might be, where we think there's a particular opportunity and a particular, where this technology has a particular advantage and could lead to some rapid advances. So both of those mechanisms are used.

Ian Krietzberg:
Has MatterGen been open-sourced?

Chris Bishop:
We have open-sourced MatterGen. We've made that available and we're Pretty excited about that and we're looking forward to, yes, feedback from the community.

Ian Krietzberg:
I mean, this is a really interesting part of the whole AI in this area, in the generation of new molecules, whether it be for construction or for drug development, because there is a process that happens after you find and identify the kind of needles in that haystack that you were talking about. How much does this accelerate? Like I think about drug development, for example. It takes a decent amount of time to identify the drug candidate that then you have to test. And then there's a whole pipeline, there's clinical trials that it could be 10 years after you've identified it. So if the AI models can help on the earlier side, we still have to go through the clinical trials. I wonder how that kind of factors in terms of the timing that you expect we might start seeing new enhancements in construction materials or drugs from a model like MatterGen. Is it a short-term thing? Is it very industry dependent?

Chris Bishop:
I think there are similarities and differences. So in both, you're comparing materials design with drug discovery. And in both cases, you're exploring chemical space to find a new configuration of atoms that has some design properties. Of course, if you're going to take this new configuration of atoms and give it to a human being to swallow, there are some pretty high standards you need to meet that are even higher than those you need to meet for, let's say, a battery or something like that. So that process of clinical trials and so on is critically important. But even then, what you may hope is that if you can better target the regions of chemical space, you'll have fewer failures in clinical trials. because many drugs that get as far as clinical trials don't ever get to market because they fall short somewhere in the clinical trial process. So I think that entire end-to-end pipeline can be accelerated in all sorts of ways through AI. We're particularly focused, when drugs are concerned, with the so-called discovery phase. You've already identified a particular disease you want to treat. You've understood something of the biology of how that works. You've identified a particular protein that's going to be the target for this drug. And then this is where a lot of the techniques that we're developing really kick in to try to accelerate that process of going from, well here's a target to here's a candidate for a drug that's ready to go forward to clinical trials. It's an iterative process. A lot of the emulators that I talked about can be very powerful in accelerating that process. But it is iterative. It involves the lab. It involves doing tests on actually synthesizing the molecules and testing how well they bind to proteins. I see increasingly a shift from wet lab experiments towards computer experiments. But we'll always do wet lab experiments. There'll always be a balance. But clearly AI has the power, the potential to accelerate that design-make-test cycle to allow us to go around that cycle fewer times, because we can make each pass more effective, but also spin the wheels faster. So I think that aspect is ripe for acceleration, and that's something that we're particularly excited about.

Ian Krietzberg:
You said the wet lab experiments will never go away, which is good and makes sense. But they're slimming down. They're becoming less prevalent because you can do so much more. And if you're doing it anyway, it's cheaper. It's more cost efficient to do these simulations, do these tests on the computer. Is there a risk of, I think about this a lot with AI and certain different applications, the kind of risk of a skill atrophy over time, where we kind of are shipping off a part of something that we used to do a lot, because now this thing can automate it, and with AI compared to other machines that we've had in the past, Here we're talking about an automation of thought processes and maybe in science experimentation processes. And so I wonder for you, do you think there's a risk in science of an atrophy of the ways in which people think about and undertake, I guess, classical analog research, right? Is that a concern?

Chris Bishop:
Actually, I think I see a tremendous opportunity here, because let's stick with the drug discovery process and the wet lab process. So first of all, it's slow and painstaking to do these syntheses and these tests. And with automation, we can potentially actually run many more experiments, so do them faster, but actually even run lots of experiments in parallel. That's high throughput methods. And AI can certainly accelerate those enormously. And I think from the point of view of a medicinal chemist, it's really sort of upskilling. You know, chemists used to learn to sort of pipette things by sucking them up with, you know, pipettes in there, literally, you know, by mouth. I mean, that's all been banned for many years. So there's a skill that's completely atrophied and that's a good thing. I think what most medicinal chemists want to do in drug discovery is to be able to think at a higher level, to be able to operate at that design level and perhaps spend less time doing some of that detailed manipulation to be able to do more experiments, do them faster, do them in a more controlled way, but free up the human intellect to do things that are at a higher level of the design process, really get rid of a lot of the drudgery. So, in a sense, yes, I think some skills will atrophy simply because they'll be replaced by perhaps AI-driven robotics that will allow us to do a lot more experiments, do them much faster, get much better results, and really free those medicinal chemists up to think at a higher molecular design level rather than at the very detailed process of actually synthesizing individual molecules and testing them. That would be my hope. So, yeah, I think the skill sets will shift, but I think it will shift in a very positive way.

Ian Krietzberg:
Now, I want to circle back to something that you kind of said at the top. You said you used to refer to the field as machine learning rather than artificial intelligence. Intelligence is not very well defined. It's not well defined. The AI of it all, all these terminologies are not well-defined, they don't have clear definitions that everybody agrees on. Largely when someone mentions something, if you talk to someone else, they think of something completely different or in a nuanced way, it's different. But so everything that we've talked about so far, it's pretty straightforward machine learning models. It takes advantage of incredible advancements that the field has had in the past few years, but it's pretty straightforward. Now, at the same time as that is going on. And I can see in your face, you probably know where I'm going. I have to ask this. There's a kind of explicit race to develop a general intelligence, which is dubious because, again, intelligence isn't well defined. And so what does it mean to have an artificial general intelligence? And for anyone listening, This is a source of heated debate in the community about, is this possible? How do we test whether it has been achieved, right? What are the safety risks? In the context of AI for science, these kind of very specific applications, Do you think something like an AGI is even kind of necessary to achieve what we're talking about? And is it even worth exploring in the way of there's a lot that we can do with other stuff and AGI is hypothetical, why spend time on it?

Chris Bishop:
Yeah, it's a great question, even if it isn't, as you say, sort of well-formulated or it's difficult to give a precise answer. I think of it in these terms. I think of human cognitive capability There are many, many different axes to this. We can be creative and write poetry. We can do math puzzles. We can add up numbers rather slowly and inaccurately and recognize cats and dogs and many, many different things. And machines can emulate many of those different capabilities. So machines have long been much better at adding up columns of numbers than humans. There are other areas where humans' skills exceed those of machines. And so I think artificial general intelligence, whatever it means, I think it probably means really capturing many of these different axes. But I think there are some in particular that are quite interesting, which has to do with ideation and has to do with reasoning. And we're seeing a lot of capabilities emerging in AI systems that are very relevant to the scientific domain, because at its heart, science is an iterative process. We formulate hypotheses, we then design experiments, we conduct those experiments, we learn from the results and we refine those hypotheses. Now the hypotheses might be which parts of chemical space will be a good drug for this protein target, for example. AI is accelerating not only different components and different tools within that iterative cycle, but actually the iterative cycle itself. So, Again, I think the human will interact with that cycle at a higher cognitive level than perhaps we've done before. And I see it as freeing up some of our time from the mechanics of implementing that cycle to being more in control of the overall design process. And I think that's a very powerful thing, not only in accelerating the route to the end product. Because we want to get better drugs, we want to get them faster, we want them to be safer and so on. But also I think even just in terms of the role of the scientist itself, it's very exciting to have new tools. I mean if you're an astronomer and then somebody invents the telescope, wow, you've suddenly, you've got an amazing new tool, or the biologist when the microscope comes along and so on. So these are enormously powerful tools that will undoubtedly greatly accelerate scientific process and I think deliver a lot of power to scientists themselves as well. As for artificial general intelligence, it's very difficult to see any particular limits on the capability of the technology. It's moving incredibly fast. You asked also a little bit about some of the sort of challenges, some of the hallucinations and those aspects, then for sure, yes, these technologies have limitations and we need to be mindful of those. And again, if I come back to the scientific discovery process, you know, our Matagen system, it produces amazing candidates and we've had some of them tested and we found the system works really well. But I don't think we would ever want to completely omit the experimental components of it. I think that's why we always say, at the end of the day, experiment is the ground truth and systems can be extremely good, very powerful, give you great candidates, but you always just want to go back and just check in the lab, really make sure that it does what you say. And we should also remember that although we're seeing these incredibly rapid advances, The world, even just the molecular world, is extraordinarily complex and we're a long, long, long, long way from really having any kind of complete solution to this. So I think there's going to be a lot of exciting research to be done for many decades ahead.

Ian Krietzberg:
complex world of molecules for sure. And the last point I want to leave off on, and it's somewhat related to what we were just talking about, the kind of AGI specter. There's a lot of feelings about AI, about this enhancement as it becomes more integrated, more prevalent. It's been referred to by many as a kind of dual-use technology. There's kind of risks related to current systems in algorithmic discrimination, when you kind of uncritically adopt systems and reduce human oversight when, like you're saying, we don't want to get rid of the experiment, we don't want to get rid of the human here. And then there's the kind of perceived threats of more powerful systems and stuff. And so a lot of people are concerned. in that environment of the kind of dual use potential, where you sit at the kind of pinnacle of, you know, if we're gonna call it the AI for good section is the AI for science section. Do you think that we'll be able to kind of battle through some of these challenges to get to a point where we can actually start, you know, reaping the benefits that these kind of advancements that you're talking about offer?

Chris Bishop:
I mean, you're right. This is a very general technology. It's extremely powerful. It can be used in all kinds of ways. It's open to misuse as well. I'm fundamentally a tech optimist and see an enormous amount of benefit that can flow from these technologies. But it's a bit like, you know, maybe an analogy would be the invention of electricity. It's incredibly powerful. It can be used for all kinds of things. It will become completely pervasive. There are sort of risks. Electricity can kill you if you grab hold of the wires at the wrong time. We learn to insulate the wires and put in fuses and circuit breaks and so on. And we learn to manage the technology. And we can't completely get rid of all of the risks, but the benefits hugely outweigh the risk. This is a very new technology. It will pose different kinds of challenges, ones that we've not encountered before. There are two things that encourage me. The first is I just see such enormous potential on the upside from the technology. So I find that very encouraging. But also a lot of people are thinking about these risks. I think the different categories of risks have been enumerated, they've been discussed, and a lot of people are giving a lot of attention to this. And so that again gives me, encourages me that it will, that we'll find ways of, you know, there'll be a few bumps in the road, I'm sure, but I think overwhelmingly we'll find ways to mitigate the worst of the risks and get the best of the benefits, as we have done with many other technologies in the past.

Ian Krietzberg:
Yeah, there will certainly be a lot to watch in the spaces it develops. Chris, thanks so much for your time. This was a lot of fun.

Chris Bishop:
No, great. Thank you very much. I enjoyed it. Thank you.