Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk

The Thesis Behind Periodic

0:00Okay, we're here at Periodic. We're here today with Liam Mendoj from Periodic. Welcome, and thanks for having us at yours. Yeah, great to be here, and yeah, thanks for coming in. I want to start off with one of these quotes that I love from your website. It says, intelligence is necessary but not sufficient. New knowledge is created when ideas are found to be consistent with reality. It seems so straightforward, but why is it non-obvious? I think that's sort of the thesis and premise behind Periodic that Mendoj and I thought when

0:33we were building this, which is you can't just think your way to a solution. The universe is so complicated that in order to actually push the frontier of knowledge and to make progress, you need to create these conjectures and then actually see whether or not it holds. No amount of rereading that textbook or paper and just thinking and disappearing into a room is going to allow you to think through all possible experimental outcomes, expected and unexpected, and you really need this iterative process. That's our core belief for putting this together.

1:08That's why from the very beginning, we need to have the AI systems, the simulations of the physical world, but then also to build up the physical high-throughput experiment. We think a different type of intelligence emerges from that. I think it's also notable, like with OpenAI and Google, you don't have that resources in those big labs. Like you had to come out and do your own thing because you are kind of inventing your own playbook as you go along. Yeah. I mean, a team like this has never existed, I think. We tried to bring together salt-state chemists, salt-state physicists, experimentalists, theorists, hardware engineers,

1:42LLM experts, computer scientists. And some of these technologies are very recent, like the high throughput experimentation, the robotic arms have only been tried in the last three years or so. The force field expertise, some of these force fields are also very recent. But yeah, we felt like having a lab is very important. Having this focus and bringing all these people together to work together is very important. And I guess like the closest examples from history are places like Bell Labs, where incredible theorists and experimentalists and chemists work together and achieved incredible things. So we're trying to do the same here, yeah.

AI in Physical Labs

2:17I want to talk about force fields in a bit, but before we get to that, yeah, like what are they? Before we get to that, I think one of the, just for context, one of the biggest differences between, you know, some of our audience are AI, you know, engineers, and some of the audience is scientists. So let's go and talk to the engineers real quick and say like, what is the difference? If you've been working at OpenAI as a big lab, what's the difference between what you do there and what is, you know, what is here? I think you've already hinted at this, but maybe explicitly, what's, what's the, how do you have to reframe your thought process?

2:49Well, I mean, I think like one interesting thing is now our reinforcement learning environments literally derive from the environment, from our physical labs. Our data comes from our physical labs. And this is sort of our ultimate truth. It's not enough just to do optimization against some answers that were known in some papers or textbooks, because we're going beyond that. And I think that's one of the biggest differences. But along with that, there's issues of say, decision-making under uncertainty. So when you're doing optimization against math, there's a high

3:22precision to it. You're not really dealing with, you know, kind of like variants or uncertainties or, you know, aberrant measurements, whereas that's very key to, to our process. Like, for example, when we are doing materials discovery loop, things don't come off the, out of the furnace labeled, right? Even the labeling process can be stochastic and noisy. And sometimes you'll have aberrations between different machines. Maybe we'll have insufficient telemetry. So we thought we ran it at this temperature, but really the temperature was, you know, some delta away. And an intelligence that

3:58can take in this noisy data and make intelligent decisions like a scientist would is a sort of a different set of reasoning strategies. So I think there's a huge amount of commonalities. So the standard mid-training, reinforcement learning, the construction of tool-using agents, variance reduction, you know, making sure infra doesn't have mismatches between like training and inference. But we have to go beyond that and really think about how do you do accurate work when there's a high amount of uncertainty? How do you make really incredibly efficient use

4:34of a limited set of data? So we're scaling things up significantly, but still it's, it's very different than a digital environment where you can arbitrarily add more environments or more rollouts. We don't have that same capacity. So sample efficiency is another key thing. This seems a little bit abstract to me, and it is a theme which has been on the science pods for a bit, but I think it might be helpful when you talk about experimental uncertainty to explain like, what would a specific process that you were doing experimentally look like? And like, what would a

5:04human, how would a human approach this problem before we jump into how do you go into the AI side? Do you have like a specific example of a type of material you make and like what sort of uncertainty and the measurements you would experience in the process?

The Reality of Uncertainty

5:17I think there are many dimensions to this. So one of them is, I think every scientific experiment has to do some dimensional deduction. And this is very different than coding or math. So when you're doing math or coding, all the context could be available to the human or the LLM. It's a bunch of axioms, some corollaries people have derived from them. It's a bunch of like functions and API calls. So everything you need to reason through is available to you. And then you just have to be very smart and like figured out. In physics, as you know, we start with more atoms than we could ever

5:48store on a computer. So clearly we'll have to go from the original number of dimensions and bits that represent the system to the number of bits we can fit in the computer. And, you know, this was, I guess, the original premise of thermodynamics. They were originally really confused about how steam engines worked, but then turns out you can find like five or six thermodynamic variables that explain what's going on pretty well, which is incredible. That's the beauty of physics. Like you have a room full of atoms and there are, you know, 10 to the 23 or 10 to the 27.

6:19There's something atoms in this room. And yet you need temperature and pressure and, you know, simple variables. And that tells you everything. It tells you mostly what you need to know for most cases, right? So like, let's say you make a material and you could do something like look at an x-ray structure. And this thing is a very lossy projection, right? It's like not actually a structure or it doesn't tell you the true structure. It tells you some principle components of that. So how would a human take these simple principle components and then kind of reason about

6:53them? And then how do you extend that to an AI? So on the theoretical physics side, right, we've decided that energy, the average value of energy and the fluctuations of energy is very important. So a lot of thermodynamics has been derived from this. And then we realized reaction barriers are also very important. The kinetics is very important. So a human kind of tries to reason through this extremely complex system by coming up with some, you know, reduced dimensionality descriptors. People also really like atomistic picture. Like,

7:24do you think about what the local neighborhoods look like, whether they're from octahedra, tetrahedra, and that especially chemists really benefit from that. Okay. So that's on the theoretical side. And on the experimental side, you would try to collect as much data as possible. So you'd have a lab notebook and you'd write everything you see, etc. But with the understanding, of course, that you'll miss a lot. So like as Liam mentioned, there are all these issues that come up. For example, your furnace will degrade over time because as you're using the furnace, some of the things you're baking will evaporate and cover the heating element. So every time

7:58you use the furnace, it gets worse and worse. Another issue is like you have a furnace, but the temperature isn't perfectly uniform. So where you put in the furnace will affect the result. We don't have this issue yet, but we will maybe someday. The optical instruments suffer a lot from vibrations. So like every time somebody walks, like I remember when I was doing my PhD, one lab had this big issue because sometimes their results will be very different than other times. And they brought it down to at a certain time in the night, upstairs, somebody will be walking and that vibration affected the laser setup. So yes, science is very hard, but that's what makes it special, right?

8:32Like the other thing Liam and I have been talking about is a lot of the current improvements focus on math and coding and theoretical computer science because it's easier for LLMs. But in real life, most things that require intelligence are actually more like science. Like there's all the uncertainty, there's all the noise, there's all the missing context, but you have to be the intelligent being and figure out what to do next. I think like a really interesting point to build on that too is the optimal reasoning strategies to do math, to do theoretical computer science, to do these types of things may not be the optimal reasoning strategies to do science. How do you define a RL reward function

9:08when your output is a noisy crystal structure or some number of, I don't know, like collective variables or something? Maybe I'll give one part of the loop and I'll just kind of dig into that. So as part of the materials discovery loop, we have to first figure out what to make. So that's the stability of the atoms together, as well as do we expect it to have the properties we want. So we don't just want a novel configuration of the atoms, we want them when put together to have the property of interest.

9:40Next is how do you synthesize it? How do you actually make the thing? That's also non-trivial. So even if you know something will hold together, the process of actually, you know, what are the processing conditions to actually bring that forth is highly non-trivial. But again, like the thing, once you've done those two steps and you've made something, it doesn't come off labeled. So you actually have to characterize it and figure out what you made. So in that case, rather than thinking about the RL environment of we're just going to kick off an experiment, wait a couple days and try to do like a update on like, did you find a room

10:12temperature superconductor or not? That's completely infeasible. I was gonna ask that. Yeah. Like you can't leave your GPU sitting for a week or something. Yeah. It's completely infeasible because you don't have enough agents to reduce the variance sufficiently. The rollout time is going between the different instruments you're waiting. There's just some physics limits of the experiments for like the actual synthesis or furnace time. So it's just too slow, too noisy. So the way we think about the AI program is

Characterizing Material Phases

10:42basically constructing agents around this set of data that we've produced. And so in the case of characterization, you're looking for RL environments that are rewarding the identification of the phases actually present. Are you able to, given the raw experimental data, fit that effectively? So the way we do this is we'll shoot x-rays at the material. X-rays have a frequency or a wavelength roughly measurable or commensurate to the spacing between atoms. So you get nice diffraction patterns that kind of give a

11:14fingerprint of the crystal structure. And doing this identification of what actually is present is highly non-trivial. And so some of our early AI work has been on basically producing systems to do this. And this allows us to move much more quickly, too, because if you're running so many experiments, you really quickly become bottlenecked on your ability to understand what was produced. But again, now the reinforcement learning environment is fairly straightforward, which is you're going to reward the identification of the phases present. You're going to penalize spurious phases or things that don't

11:49have any, like, real chemical plausibility. So that's like a, you know, a little succinct thing where you can kind of do this for the different pieces. And when you stitch this together, that's the end-to-end discovery loop. Maybe like another piece would be, as we build up a large basket of experimental data, you can think of another thing of basically like timestamping the state of the world. So you could say, at this date, this was our experimental evidence up until that point. And you can construct reinforcement learning environments where you say,

12:22okay, given that body of experimental evidence, and given, you know, the choices, what was like the next choice that the scientist made, or what was an outcome of that experiment? And this is really interesting, too, because this allows us to do types of programs that would be harder externally, because if the pre-trained model has already memorized some of this data, and then you try to do reinforcement learning on that, if it already knows the answer, and you construct reinforcement learning tasks that depend on that answer, it can fake

12:54work. So, and then, I mean, because it's like, oh, well, it already knows the answer, so it doesn't actually have to do the hard physics reasoning, it doesn't have to actually do the calculation simulations, and it gets to the right answer, and then you reinforce that policy, and you upweight those reasoning strategies. Those reasoning strategies will not generalize to novel systems, so it's not useful to us. But as we build up baskets of experimental data, and like stitching things together, having the lineage to the whole process, that gives us the ability to construct new types

Why Physics Simulations Fall Short

13:24of environments that are just not plausible elsewhere. I want to ask my question, but I'm worried that I'm taking it off the rails. I'm just going to go for it, because I, again, I come from the non-scientists, but engineering, I'm familiar with the ML side, but not the physics side. Okay, there's a few versions of this, but I think the basic question is, don't we have, like, most of the laws of physics worked out, and, like, how come we don't have perfect simulators already? Oh, that's a great question. Like, this is, like, very basic, like, and he was like, ha, ha, that's cute. But, like, I'm like, dude,

13:57like, like, I have all these physics books, so what do you mean, what do you mean we're not done? Yeah, so there are a couple of things going on here, right? So, I think, first of all, we're definitely not done with any of the laws of physics. Like, even the... Like, a quantum scale, okay, but, and, like, maybe, like, you know, very, very large scale, but, like, at human scale, at, like, you know, material scale, we're, like, what's left? We're not, so it's a really good question, and it's very interesting why we're not done yet, right? So, there's a very

14:29common story people talk about. Like, when Dirac was figuring out quantum mechanics, like, late 1920s, he wrote a textbook about quantum mechanics, and he kind of framed it as it's all done. And then there's a phrase people like to quote, I'm not sure actually how accurate it is, but apparently Dirac said, the rest is chemistry. The idea being, like, he could solve the hydrogen atom. Just compose everything. Yeah, he could maybe solve, like, a chain of 1D hydrogen atoms, but then he cannot currently solve, say, nitrogen interacting with oxygen, but that's okay, that's just chemistry. Yeah. Physics, physicists really love single, like, atom, really simple

15:05systems, and that's all you can solve. Spherical cow. Yeah, yeah, spherical cow. But it turns out, I think what we learned in the last hundred years is, first of all, that's not true. It's not, like, a simple extension of what it was to do with hydrogen. And it's not just, the rest is not just chemistry. There's actually a lot of physics there. I mean, I guess one of the things we still haven't figured out is high temperature superconductivity. But there are a lot of other things. And high for you is, like, 175, 200. Oh, when people say high temperature superconductivity, what they mean there is unconventional superconductivity. So, there's the conventional superconductivity that's just mainly driven by electron-phonon coupling.

15:37And in that case, you can kind of see the isotope effect. Like, if you take the same system, but just different weight for one of the elements, you can see the superconducting temperature drops at the rate you expect. So, that one is conventional. But then higher temperatures for conductors, like Coup rates, like the ones that are above 77 Kelvin, like 93 Kelvin, turns out don't obey that physics. But we don't know what physics they obey. They're just incredible superconductors. But it's not just that. So, that definitely is one of these, like, very popular topics. We don't know the theory for yet. But there's so many other things we don't understand. Like, for example,

16:08even strong correlation in simple quantum mechanical systems, we cannot simulate yet. Density functional theory is a very powerful tool. It's incredibly accurate on some things. But as soon as there's, like, some strong electron correlation, it actually fails to capture some of the effects. I mean, there's so much to figure out. That's why I'm really excited for AI to be applied to this field. Because theoretically, computationally, and experimentally, there's so much to figure out. And every improvement we make here should improve human life. Because, you know, the better we can understand materials, solid state physics, the better devices we can make

16:41for them. There's this famous quote by Phil Anderson, which is, more is different. Which is basically, you may understand all the basic laws of something. But when you add many things, they behave qualitatively, distinct, differently from how the simple, like, physics should tell you. So, understanding how lots of things behave together, it seems like it should be simple. Like, beginning laws are simple, but the collective behavior is so complicated that it's really hard to model. It's, like, emergent and universal, which is crazy. And we're also seeing this with deep learning models,

17:12right? Like, we see power laws everywhere in deep learning. We don't understand. But it's, like, it's very reminiscent of physics, where there's emergence and universality. There's been many theory papers in the physics world about emergence and neural networks. Yeah, absolutely, yes. And then my other follow-up question was also just on the process of, let's call it the process and the end result. Like, let's call it, like, you want to have these properties that you're targeting for materials, and then you have to figure out the process to get there. Is it worth separating these things so that you can have, like, you just train a model that perfectly reverse

17:46engineers any process whatsoever, given whatever theoretical end state you want? Is that meaningful? Maybe going back to the conversation we had about how, since there are more than Avogadro's number of atoms, we'll never know the exact context of what happened. I think that might be a reason why we'll never be able to reverse engineer everything perfectly. But we're just trying to reverse engineer sufficiently to be able to improve the performance of the material, basically. Yeah. What you're asking is, is there a forward model where you put in an input and you can predict

18:16the output, and you want to understand, you want to create, like, a surrogate model, which can understand? Like, the input is, like, the final state of that, this configuration of atoms, and then the prediction is... Because then you can split up the work, almost. Like, you can have one team do the process side, the other team do the property side, and then, you know, just raise them. Yeah. I think there's a lot of progress we can make in terms of these, like, you know, splitting up these agentic workflows to these different areas. I think there's a lot of, like, empirical work we can make on synthesis prediction as well.

18:49Yeah. Whereas, like, you know, taking in that configuration of atoms, what tools can you use? Given this experimental evidence, given prior literature, prior papers, how do you actually get to these things? And you ultimately are testing, you know, are you able to actually replicate it or not? The reason I say this maybe also is just thinking about it in terms of a company. You know, the way that TSMC is kind of like a fab for semis, but they don't design the semis. You know, you could be the TSMC of materials, where, like, people come to you with, like, well, here's what I want, and then, like, you can figure out the process.

19:21Yeah. It's like a matter compiler. Yeah. It's like, given these requirements... Very Star Trek. Very Star Trek. Yeah, yeah, yeah. But it's like, given these requirements, is this actually, like, a valid set of configurations? Can this exist? Is it a good objective function? I don't know. It's too big. Yeah. Well, I mean, we kind of talk about, like, one of our missions as, like, synthesis super intelligence. So, I think it's in line with this.

Phase Transitions in Practice

19:43Yeah. So, I want to double-click on something that you said a little while back, which was, you talked about phases. And I think this is, you know, going to Doge's comment about what is, like, emergence. Like, the concept of a phase transition, I think, might be foreign to a lot of the audience. So, can you explain, what is a phase transition, and then why is this, like, a helpful signal for something like an RL environment? Phase transitions are fascinating. I mean, I think maybe the one that people can most easily relate to in their life is

20:15ice melting, probably, or water boiling. Like, you know, you increase the temperature of ice, and it still looks like ice, and it behaves like ice. But there's a certain temperature at which it just stops raising its temperature, and then turns into liquid. So, it suddenly goes from this solid phase to liquid phase. Another one that's very relevant to us, of course, is the Ising model. And I think computer science and mathematicians also study this maybe under different names. But basically, you can have spins up and down, and they have different interaction terms

20:46if they're both up versus one is up, one is down. And at high enough temperature, they're usually just randomly up or down, entropy wins. You lower the temperature, it's still looking the same. But then at some point, suddenly they perfectly set aligning, yeah, the phase transition. Phase transition is really nice because it makes physicists' life easier for studying certain things. They tend to have certain spatial correlations that really help us as well. Okay. So, in our lab, phase transitions come about usually because we mix the precursors that are

21:16different crystals, basically. But then we raise the temperature or do something else to encourage them to react. And then atoms start reacting, and they form into this new crystal. And this shows up in the XRD pattern because usually if the geometry of the atoms change drastically, the XRD pattern changes a lot. But what really happens in like a typical practical material discovery campaign is, when you first try something, it doesn't work. And the outcome isn't just like a clear yes or no, but it's usually like a very mixed phase.

21:46It will usually have some of the precursors, it might have some amorphous phase, and then it will have like a bunch of phases that you maybe didn't try to make. So, that becomes a challenge. And that's why we found it very difficult to bring together simulations and AI to do the characterization. So, the simulations can always say, oh, this phase doesn't look like what we predicted, but it's like slightly variation of it. Let me now do force field calculation to see if it's still stable. Or like the AI can say, based on the previous experiment on this one, this is probably not the phase we want to make, so let's change the conditions.

22:18For listeners, XRD is X-ray diffraction, which you already described. Yeah, in some cases, there's like a phase that just hasn't been recorded before. It doesn't exist in any papers or database. And so, the AI has to actually go and sort of use these tools to say, well, what configuration atoms would actually explain these phases. But this becomes incredibly important for directing the scientific process because let's say we have a phase target in mind, we really need to sort of hill climb that to get better measurements on that. So, if it's 1% and we're like, actually, we want to increase this phase period,

22:51we want to have a very good measurement on this experimental data to say, you know, what is present and what's not. So, one of the advantages is you have just a distinct, like, variable that you can just observe. Like, going back to the statement about how do you deal with uncertainty, it seems like part of your answer is you make your observable signature somewhat unambiguous. There's no interpretation there. Is that right? Is that, like, the logic behind this? Or is this, like, literally, you just need to define a new phase and so this is the thing you care about? I mean, I think there's still ambiguity in a lot of these cases.

23:22I mean, like, you know, a very dumb thing is, like, replicates. Yeah. We run replicates, of course. Yeah, yeah, yeah. But no, I mean, I think they're, you know, it's a challenging task because it's not just fully deterministic. And there are, like, well, there are, like, ambiguities and other things. Two different phases can actually be consistent with the same pattern. Yeah. And so I think that's why it's really important for the system to use some chemical intuition to say, like, okay, well, this would be highly unlikely given the synthesis conditions, given the prior knowledge.

23:55And so that can be used to help disambiguate. So you're somewhat injecting priors there based on what you expect. Yeah. I mean, I think thermodynamics is the biggest prior, right? Right. And then physics is a big prior. Another one that we really benefit from here is multimodality or materials characterization. So, like, we can do XRD and, yes, some phases might look similar on the XRD pattern. But then we can also measure their electrical properties, magnetic properties. We can measure their morphology, like some kind of electron microscopy. And then some of the phases that look similar on XRD will look different in some of these dimensions.

24:27And that kind of multimodality really helps. And again, AI is really helpful here because, like, humans have limited context and computation power as well. So if you give humans, like, 10 different modalities at the same time and, say, analyze these all consistently, it's a bit difficult. But it's actually not super intelligent for an AI to do it. And it's great. Yeah. Yeah. It's like the super intelligence is it can just do so much calculations. It can look at everything. And this kind of gets to that phrase about, like, better decision making under uncertainty, where now it's stitching together the signal across many different instruments longitudinally across different experiments across the replicates

25:04and kind of getting to this, like, underlying better model rather than just over-indexing on one measurement from one instrument. And let me just, like, so, you know, you started, I think, a while ago also saying that there's just more data than you can fit in any reasonable computer or process or anything. So, at some point, you have to throw out data, even though you have all these replicates, even though you have, like, presumably, like, 10 different ways to measure the same thing. Because what if one of your temperature things is wrong? So, like, when do you throw that out? How do you decide what to do? As a data-hungry ML guy, I just want everything, right?

25:35And then I just, you know, waddle, learn everything. I don't know what I don't know, you know, just throw the machine at it. Yeah, so we don't throw in data. I think maybe what I meant was potentially, like, if God was observing the experiment, there is more data in the experiment than humans can fit in a computer. Unfortunately, us mortals cannot observe all the atoms. So, even though there's more data than we could store, we actually can't access that data. For example, I cannot trace all the atoms' positions.

26:05So, unfortunately, we're still in a regime where any data we collect is very precious and we don't delete any of it. Maybe a way of thinking of it to connect to the AI side is, like, think about probing, like, linear probing of a foundation model, right? This model has all of this data. So, God has this entire picture of what the material is doing. And then the experiment is this little tiny linear probe which has, like, eight bits or something of output. And so, you're taking this humongous thing and, of course, grading it down to a little bit of information.

26:35And then it's now either traditionally human's job or now your agent's job to reconstruct from those eight bits of information what is actually happening inside. Like, it's an analog where you can't have full visibility over every parameter weight, over every activation for that input. Yeah, you're getting, like, some sort of subset on it or some, like, you know, coarse-grained features across the whole thing. Yeah. But you probably find that you probably do find Maxwell's daemon argument very interesting because that kind of… Can you recite it because I don't…

27:07He's mentioned it before. I keep forgetting it. So, you know, in the 1800s, people realized that entropy or a system has to increase for a closed system or stay the same, but it can't go down. Yeah. It's the second law of thermodynamics. Yeah. And there was a thought experiment that was asking what would happen if there was a very tiny, all-knowing daemon that could kind of look at atoms. And then whenever an atom had low energy, kind of open a door and let it into this other room and keep doing this to reduce entropy of the system, which would break second law thermodynamics.

27:41And for a long time, I think there wasn't, like, a very satisfying answer for why that was the case. But then there was a more recent paper in the 1900s, more recently to us, Landau came up with the argument that you'd have to spend energy to delete information. And Maxwell's daemon would basically have access to so much information by doing this thing he's doing that he'd have to delete some. And when he's deleting information, he'd have to spend energy, which would then make sure entropy increases. So even though we're not at the level where we can be Maxwell's daemon, I think your question is, like, very visionary towards, like, a future where we store so much data we have to delete some.

28:16Yeah, no, I wasn't obviously going to that extent. It does remind me of, like, air conditioning. That's kind of how air conditioning works. But also, I think it's just, like, a question of, like, why don't you just buy every sensor in the world and just, like, you know, ridiculously over-instrument, over-replicate everything. How we do that. Right? Okay, yeah. Right. All right. Yeah, I think it's a matter of- Confirmed. Yeah, exactly. Are you getting visibility to individual atoms, which is infeasible? But no, we add a huge amount of telemetry. And that basically reduces the number of, like, hidden variables, if you will.

28:49Yeah. And just out of curiosity, do you worry that you're just in, like, California? Like, don't you want to do this in, like, Nepal and, like, Australia? Yeah, we could. I mean, we have plans to, you know, open more labs. Because, obviously, gravity and, like, you know, where you face and, like, where you are matters. Also, like, expertise can change, right? Like, here we have certain kinds of expertise. For example, Stanford physics department, engineering department, has certain kinds of expertise that really helps us because we can hire from there. Our resources can collaborate.

29:19Like, as you said, like, geography can matter. But then different cities can have some different sets of expertise just in physics, chemistry. Yeah. Oh, I mean, we have Zoom. Yeah. For that. Just like your experiments are going to be affected by, like, where you are. Yeah. Yeah, I mean, the future of Periodic is not a single lab. I mean, I think you've got to hit up, like, your buddy Jeff and, like, go to space and, like, you know. Right, right. Yeah, I mean, I think it's necessary to have this sort of network of labs. And, like, yeah, I think as they're just kind of pointing out, like, there's an optimal place for each of these based on expertise or, you know, materials constraint.

29:53There's many different constraints permitting. Okay. So, one of them is, so one thing I've been curious about is how do you deal with computational tools?

Integrating Computational Tools

30:01Or, like, how do you integrate computational tools into a decision-making workflow? Especially when, let's say, computation and experiment don't necessarily, you know, meet. It's also, like, theory in terms of, like, high-level, like, understanding. And, yeah. How does that sort of work together? Yeah. So, I mean, one thing that's really helpful here is there are certain things that are easier to do in simulation and more accurate in simulation. There are certain things that are easier to do in experiment and more accurate in experiment.

30:32I mean, in general, of course, experiments are more accurate, but certain measurements are very hard to make with experiment. So, you make a bad version of it, so you get inaccurate data. Like, let me just give you some examples. So, one of them is we use density functional theory to estimate formation enthalpy or materials. Sorry, quick pause. Can you explain density functional theory and formation entropy? Yeah. Sorry, enthalpy. Yeah. So, density functional theory is probably the most commonly used simulation method for materials.

31:02It kind of comes from Kohn-Honenberg theorem and Kohn got a Nobel Prize for this. When he realized that, sure, quantum mechanics is very expensive because you basically have this exponential Hilbert space for every wave function. So, you have to solve this, like, very difficult problem, especially as the system size grows. Kohn and Honenberg realized that, turns out, you don't need to deal with that exponential space. All you have to do is think about the charge density, at least for the ground state of quantum mechanical systems.

31:32So, it's actually, the theorem is pretty simple. So, even I think an undergraduate in physics can understand. But the idea is all the properties of a quantum mechanical system ground state is a functional of the charge density, which is amazing because charge density is just this three-dimensional object, whereas the Hilbert space and wave functions are exponential. So, that was a really interesting observation. And then, Kohn published another paper, I think, with his postdoc, SHAM, Kohn-SHAM wave functions,

32:02kind of had a practical way to solving some approximation of quantum mechanical properties of materials. And today, we use it very much, both in our lab and externally. And then, formation enthalpy is one of the measurements we can make to understand the energy of the material. Like, basically, we want to ask, what is the energy that this material has? Because if that energy is higher than other materials these atoms can come into, then it's unlikely that this material will be able to be made.

32:32So, stable just means, like, it's on the convex hull of other materials and their energies. So, was it clear enough, or should we... I mean, maybe a simple, like, high-level point. So, like, density functional theory is a way of taking something which is exponentially hard in number of electrons or number, you know, in your system size, and then reducing it approximately with some introduced error to, like, maybe usually an in-cubed approximation. So, now you can compute quantities which are, you know, you can now compute things which are just uncomputable, but at some cost.

33:08All right, if you could do it perfectly, you would, we wouldn't need the lab. But, you know, we can. This may be a good callback to our, actually, our second episode ever. Heather Kulik had said, she famously said, there is no alpha fold for materials, not just computationally, but the ground truth just doesn't exist. Like, we don't really know what material crystal structures look like. And DFT is generally our best route towards. Yeah, I think so. I mean, maybe two things I would add is DFT, by definition, doesn't have to be an approximation.

33:40The Kohn-Honenberg theorem shows that it can be exact. But I think, as you're pointing out, the exchange correlation functional, we don't yet have access to. And then, for the charge density-based models, we don't know what the kinetic energy functional is. And then, even if we had perfect DFT, I still think we'd need a lab. Because we, even if we have perfect DFT, we cannot fit 10 to the 23 atoms in the computer. Yeah. You're very obsessed with that, Amber. No, okay, okay. Yeah. But with exponentially growing compute, you can argue that in encubed scaling size, eventually you could, in principle, calculate things.

34:14You just wait until the law is long enough. Oh, yeah, we need massive computers. Yeah. But as long as in-cubed is a good scaling, you can... As the non-scientist, I will make an observation, which I don't know if you want to throw around or anything like that. So, I came from finance, we had the Gaussian copula, which was a way of pricing credit default swaps by doing correlations and reducing everything into a single kernel. And it feels also spiritually similar to the VAE, in terms of condensing all these things into a single parameter.

34:48I wonder if there's the same trick everywhere. Yours sounds a bit more high-dimensional than mine. Which, you know, VAE is just one-sigma thing. But it sounds the same? I guess the charge-sensity file is still quite a big file. Yeah. Right. You hand-wave a bit, but it's good enough. Like, this is the highest order bit. Like, yeah. I personally don't know of a better method to predict the stability of a new material. It's definitely not perfect, but it's definitely better than other methods I can think of.

35:20And yeah, so going back to your question, the reason we use them is, for certain things, it's easier to simulate than experiment, and they kind of complement each other well. And then we can, like, do this in a loop. Like, we feel like simulations will never be enough by themselves. But in the loop of simulations, AI, and experiments, I think we can make progress much faster than before. One thing that simulations let you do is scale up much more quickly. I mean, there's now, like, compute is much easier to bring online than bringing on new labs.

35:50How do you avoid, or let's say you scale up a bunch of DFT, how do you avoid having your models over-index on that? And still be grounded that the real world is ultimately the truth you care about? If you, you know, how do you sort of balance that? Well, we scale up the lab as well. Yeah, you scale up the lab. But I mean, you could, I could imagine scaling up DFT to millions, hundreds of millions given modern compute, whereas your lab, I would imagine your hundreds, thousands a day, I don't know. Like, something still, like, fairly, even in a really high throughput case, like, you know, you're still fairly limited by comparison, right?

36:26Yeah, I mean, one thing is the humans, but also the LMs, are kind of pretty aware of the limitations of DFT. For example, even if we can do, as you said, 100 million trials with DFT, we could never actually try their microstructure. And microstructure is this idea that, like, DFT usually simulates the perfect crystal, but crystals usually actually aren't perfect. And their microstructure, which is the structure at like a medium order, can really affect the properties. Other issues, of course, like some properties like superconducting temperature cannot be easily simulated by DFT.

37:00Again, like even if you do 100 million trials with DFT, it won't be enough. So we have to rely on heuristics, experiments, etc. And so there's a huge filter from all those calculations to actually what gets executed in the lab. Is that human mediated or LM or mix? Can be a mix. Like, for example, for a long time, materials project, which was the best open source DFT database, use experimental calibration for DFT. So they would actually use experimental data to calibrate DFT. We also do this internally here.

37:31Like, whenever there's a certain chemical system we're interested in, you get experiments, you get simulations, and you can calibrate them. This is another thing. I've talked to a lot of biologist friends, and they are always very confused when you say materials, you can't just do an XRD structure and actually know what the structure is. Because if you're in the biology world, you can look at a crystal structure of a protein, and typically, more or less, with, like, sub, well, they order Engstrom accuracy reconstructed. What is the difference between the materials, XRD, you know, X-ray diffraction?

38:06Diffraction, sorry. Experimental XRD for materials versus, let's say, the biology world. Like, what information do you lose, and why is this sort of like a lossy projection? It's hard not to sound like crazy when answering this question, but do you not feel like organic chemistry and biology almost stem from a lower VC dimension, like lower Kolmogorov complexity model? Like, you can almost represent it as a 1D sequence, and it can almost be compiled. So, like, there's a sense in which it's simpler in the Kolmogorov complexity sense.

38:381D? I mean, like, just DNA is, like, just, or RNA. I think most people would argue it's more 2D, but yeah. I'm not a biologist, so I'm sure they're right. But with inorganic chemistry, virtual is crazy. It's clearly not coming from such a low Kolmogorov complexity model because, like, 3D inorganic crystals can host metals, insulators, superconductors, diamonds. It's, like, really crazy different. Like, people tried, but we could never find a 1D representation, right?

39:10It's very hard to find a smile string for 3D inorganic crystals. And the way the atoms interact with each other is quite strange, like, covalent bonds, ionic bonds, metallic bonds.

Inorganic Versus Biological Systems

39:21But yeah, this is very interesting. I wonder sometimes, like, you know, have you ever looked into issues with solid electrolytes? Like, why people cannot replace liquid electrolytes in batteries with solids? There are all these dendrites that form. Like, basically, the lithium will form these structures that go into solid electrolytes and crack it. And you look at that, and you think this is so primitive compared to biology. Like, in biology, these systems can do such complex nanotechnological progress, whereas we can't even, like, make two interfaces work with each other because lithium will break it.

39:55Yeah, but then I guess inorganic stuff can be very resistant to high temperature. You can make space shuttles, you can make silicon, compute Moore's law. So, yeah, it's a tradeoff. Yeah, we've talked about this with several other guests. I think one of the big differences in my mind is that biology, we have a toolkit that, you know, that you can basically borrow, like, from millions of years of evolution. Yeah. That has already given us all the tools, and we can just reproduce those. But also, evolution also constrains the, yeah, the VC dimension to be fairly low for, like, proteins and so on.

40:29The, like, emergent behavior of large-scale systems like cells could be much more complicated. Yeah, and cautiousness. Yeah, yeah, yeah. Yeah, but maybe even just, like, more specifically for, like, if you just look at an X-ray, you had a perfect crystal, and you shoot an XRD at it, and you get the, you know, you get the spectrum. That still doesn't actually tell you uniquely what the crystal structure is, right? No. So, and that's something that. What? It's just average. Yeah.

40:59It's just average behavior. But, like, for example, disproportionation is a very complicated thing, right? Where it's possible that you had the perfect crystal, and on average, it looks like this perfect XRD. But in reality, there is some other pattern going on where, like, atoms slightly shift to the right or left, but on average, they're in the middle. Yeah, it's a difficult life, yeah. Whereas with proteins, you basically know there's some sort of underlying protein that follows some rules, so you can combine a forward model with it. Right. And then you can kind of move and move towards, like, you know, single crystal, XRD, and that can help. When we started the science pod, there were all these, like, elements of AI and science, AI and math, AI and physics, and all these things.

41:35And there was this, like, kind of pissing contest, like, which is harder? And, like, I think, like, you can make the scientific argument that material science is the hardest. Well, we didn't pick it necessarily for hardness. I mean, there's actually, like, a lot of areas where it is easier. I mean, we have these, like, great simulators for large classes of materials. Whereas in biology, it's incredibly difficult to model a cell or an organ or a full organism. So, I mean, that's a huge advantage. Yeah. Isn't that interesting that at the smallest level, it is one-dimensional or two-dimensional?

42:07But then, like, it just has many orders more scale and structure. And that actually introduces the complexity. That's weird. Whereas, like, maybe the complexity in materials is more the microstructure. Yeah, I mean, you still get, like, different complexity at the different length scales for materials as well. So, yeah, microstructure and beyond. Yeah.

The Value of Superconductors

42:27Just sort of like a brief, like, sort of palette cleanser. In terms of science fiction, let's say you can invent whatever materials you want. What is more valuable? You know, I have a list here. There's obviously room temperature superintroductors. But also, you know, you mentioned batteries. I've always thought, like, batteries. Just, like, do better than lithium-ion, you're good. Like, you know, that's stood around for, like, a hundred years. And then the last one is carbon nanotubes, mostly for the space elevators. I don't know if, like, those are three that come to mind.

42:58Are there any that, like, people talk about in your world that, like, are the dream? Definitely superconductors. Yes. Magnets. Right. We can lower the dependence on rare earth or, like, transition metals that are hard to source, like cobalt, reducing cobalt in batteries. One of the things I think is very fundamentally important is if we can get close to Landauer limit for compute energy efficiency. Can you explain Landauer limit? Yeah. So, if you look at how much energy we spend per, like, flop or compute, it is followed the Moore's law type behavior where it's kind of exponentially more efficient.

43:36So, we've been spending exponentially less energy per flop or compute. I mean, I haven't checked recently. This was the case 10 years ago. Yeah, that's great. I remember when I first learned about it, looked at that, and we were, what, like, 10, 20 orders of magnitude away from it or something. Yeah, on the limit. And last time I looked, we're weirdly close. Like, we're within a factor of, I don't know, like, three or four orders of magnitude away. Yeah. It's actually, like, I guess exponentials rolled out over many years. Yeah, yeah, exponentials over many years are, yeah, pretty impressive. What about, like, in your lifetime? Yeah, just since I discovered this, like, what, 10 or 15 years ago.

44:10Yeah, I know. And one of the things we have to do is dissipate heat. Because, like, as these computations are happening, it's producing heat, which has to happen, but then we have to dissipate it. So, like, if we can get close to the Landauer limit for how efficiently we do computation, that's amazing for humanity, right? Because that means now we're doing computation, which is probably one of the most fundamental things we do as humanity, but as efficiently as possible from an energy perspective, which is a real currency in the universe. That get affected by quantum computing? I'm just going to throw it out there. Theoretically, massively, embarrassingly parallel compute.

44:41So, it definitely gets affected by reversible computing. But honestly, I'm not an expert. I don't understand. So, if you can do reversible computing, now you're not, you don't actually have to spend energy to do computation because you're not actually deleting any information. It's reversible. But I don't know if that works. I don't know. Quantum computing, I assume, has to obey these laws somehow, because they are so, like, they have to obey thermodynamics. It's unitary, so it's reversible. One of the most memorable conversations with Elad Gil was, like, he actually was a very skeptical person about quantum computing.

45:14He's like, even if he had it today, there's no applications. I'm like, whoa. Whoa. Like, it breaks security, I always say. Like, that's it. I agree with that. Like, Prudem, people talk about how if you're a quantum computer today, you could simulate things so much better. And I ask them, okay, let's accept that. What would you simulate? Like, you're still simulating perfect crystal. Like, you still have the issues that DFT has if DFT was perfect in its prediction ability. So, yeah, I do think people are kind of glossing over some things because quantum computing is so exciting.

45:44Like, if we can compute in a quantum logic space instead of classical logic, it's just so exciting that people are glossing over what it would actually do when it's made. Maybe that's fine, because maybe once it's made, it will do amazing things we can't even imagine. Good. Okay. I was going to move to the automating the lab side, where you've closed the loop on your experimentation. You can talk about robotic arms, talk about basically just everything you've done here.

Building Intelligent Laboratory Equipment

46:10My favorite quote from you was that every piece of equipment in your lab is going to have 140 IQ. Okay, what does that mean? In the early days, as we were scaling things up, we realized that we had huge bottlenecks imposed just by operating machinery. So, for one set of machinery, we would have technicians and scientists looking for particular morphology and trying to see, okay, what actually were we making, see if this is consistent with our intentions. And we really quickly, as we scaled up the lab, came into these bottlenecks where it just wasn't keeping up.

46:44So, we started doing some programmatic approaches to capturing the data on SEM. And the data that was being surfaced from this very simple program, where it sort of takes a field of view, captures, zooms in, captures, was just not all that useful. It was a little too dumb for what we really wanted to take out. And so, at that point, it was then really pertinent to build AI systems directly onto the machines to start controlling these things. And now they have the full context as to what we were trying to achieve.

47:16So, what was the intent of the experiment? What were we trying to synthesize? What were some other experimental evidence? And it's actually looking in the machine and capturing that data. And this is really valuable now because we've been talking about these hidden variables and not being able to capture everything. But if you can do more intelligent data capture at the time of that experiment, your data for future AI systems and future computational predictions is that much better. You'll have just a richer set of data. So, that's sort of what we mean by just everything on the lab has to be incredibly intelligent to just make the data as useful as possible.

47:53So, that was sort of the motivation and inspiration behind that. What's the state of the art there in terms of putting intelligence on every device, right? Well, I think one interesting aspect is what's the latency of controlling that instrument? Because there's cloud latency, but then there's also device compute. That's right. And then there's sort of a timescale associated with different physical processes. And if calling out to an API or something is simply too slow, so the amount of reasoning or tokens or tool calls is just not matched to the latency of that actual process, then that's infeasible.

48:32Okay, that is right. I mean, it sounds like self-driving cars, right? So, real quick, I just like, I'm actually surprised to hear that because the latency I'd imagine for reasoning seems small compared to a lot of these things take like hours to run, right? Or maybe, you know, maybe your experiments are much faster or high to put or something. I'm just surprised to hear that actually. So, ultimately the experiments do take, you know, many hours, days, but there could be particular steps where latency would matter. Yeah, like you might want to have finer control.

49:03Another thing is how expensive it is because like the same reason some of these models can be really slow in analyzing data also makes them very expensive. And then finally, the human patience, like if a human wants the analysis of a certain XRD pattern and they have to wait two hours before they get a good result, that's very different, I think, than if they can get it in like two minutes. Oh, okay. Yeah. And then the reasoning, you're not doing like a single call to that model, right? Like you're doing potentially many, many tokens to kind of get to these patterns.

49:34Like run simulations in the loop. Exactly. Deep research in the loop. It's not a single tool call. It's not a single inference thread. So, then the latency can blow up quite a bit. I see what you mean. Yeah. So, how much of your latency is tool calls? And I assume tool calls is like largely like DFT or computational. Yeah. Yeah. So, how much of your latency is derived from those versus like actual reasoning? It's really process dependent. Okay. Yeah. And then the other thing I think about is also, I guess, building up from small things to bigger things, where I assume that the general temptation or the typical development is incremental, where like everything is human operated, and then you find ways in which to automate it, and then you sort of build up from there.

50:19So, I worry that sometimes that that is the way that people evolve things, but that's a local minima, optimal. Like short noise and optimization. Right. Exactly. When like actually you should get a humanoid and just put them in there. I think we're in the opinion that solving humanoids would actually be slower to kind of getting to some of our goals. Just checking. Again, a lot of this is just like you do this every day. We like see this, but I don't know the reality of the situation. A lot of people are hyperhumanoids. Yeah. I mean, I think like maybe one process is by having like a mix of humans and automation, you can identify really quickly like what are some of the bottlenecks in the experimental process and you start alleviating those bottlenecks one at a time.

51:00And so, for example, if there's some really tricky dexterity task that humans are excellent at, but you'd have to spend months automating machine, maybe don't spend a ton of time there. And maybe like a more promising thing is something very routine that takes up a huge amount of scientists and technicians time and it's easy to automate. So I think there's like this pragmatism to it, but I mean, ultimately what we want from the lab is a huge quantity of data, high quality data, diverse data. And that those are our goals and full autonomy is a non goal.

51:34It's sort of in the service that we use automation in the service of achieving the goals on the data. Yeah. But also, I mean, I think another aspect to automation is robots will just be make fewer mistakes potentially in the lab. And it's been really helpful having AI systems with a full view over all of our data, because sometimes if there's like a permutation in data, then it can identify it. So, for example, one of our steps at one point had a cyclic error because one of the machines was loaded incorrectly and the patterns were inconsistent.

52:09So the AI was reading through these things and says, well, given what was run, this is not expected. And it's looking at this basket of data longitudinally. And then it realized if I do this like cyclic permutation and reverse it, everything is consistent. And then we're able to go back to the physical infrastructure and understand that a mistake had been made in the loading. And I think like managing data quality is so foundational to doing AI in the physical world. And these are the types of things that we're building in.

52:39And so you can improve the operating process. But at that point, you're like, okay, this is another great opportunity for automation. How do you make this just so reliable, so durable that we never have those types of mistakes again? Yeah, I mean, reducing the noise floor experimental data, I think is very important. Most experimental data in literature have such high noise floor that it's actually usually worse than DFT accuracy. Because, you know, when different people do experiments, different labs, different parts of the world, different times, it really introduces so much fluctuation to the result. So we are hoping that by standardizing these workflows, one of the biggest benefits will be that the noise floor will be lowered.

53:15You've been pretty public about how you use open source models and fine tune them for stuff like this. Is it better? Is it basically, I can imagine a situation where you have like maybe the dumber models do one task, and then you optimize for that task, and then like sort of like a generalist frontier model that like supervises everything. Is that a good mental model to have? Are there more stages to this that I can't think about? I mean, we basically use a mix of open source models and closed source models. You know, we're huge beneficiaries of this from our simulation side, building up the ML stack, the simulation stack, we don't need to push on that axis.

53:50But there's a lot of areas where the latency is too high, the cost is too high. But in some cases, we can actually, you know, by having access to this data, we can actually push beyond the frontier of what these systems can do. Under even some like the highest reasoning efforts. And so it's not necessarily just about cost or Pareto efficiency, but in some cases, you can push beyond it when you have access to data that no one else has. One way to characterize this is in terms of compute efficiencies.

54:20So by having access to this data, you can be that much more compute efficient compared to, you know, some of the frontier models. And that's been able, been like a really instrumental thing in our program. Yeah, I'm so data quality or data access as a trade off of compute, like, you know, I think there's some amount of exchange rate of dollars for compute for dollars for data that people, I feel like the pendulum might be swinging now towards data. I mean, obviously, you would agree with that. Yeah, especially like, you know, talking about like the noise floor of experimental data, like we need to have high quality data.

54:55If you throw a huge amount of compute against a bunch of noise, you're not going to have a good thing emerge. And it's harder for you because you actively try to be sparse, like you actually try to get like null results and learn from that. That's right. Yeah. Yeah, you've talked about null results and several other venues. Like, what does a null result look like for materials? And how do you use that effectively, especially when I think your overall signal is probably quite sparse usually in terms of success? I mean, a null result could be we intended to produce some structure, and then all of the evidence points to us not producing that structure.

55:27Do you ever intend to not produce a structure? Of course. Yeah. Okay, absolutely. You do the negative controls. Absolutely. Okay, okay. Like the impurity phases that will kill your property. Yeah. Okay. It can be toxic. Exactly. 100%. Okay, okay. And a negative result, you know, in some cases is actually, you know, we intended to make the following thing, and then we're actually able to identify a new structure that hadn't previously been identified. So it's negative in some respect, like we intended to do something else, but something else emerged from the data.

56:02Yeah. Yeah. So when you're training a machine learning algorithm, especially like at a very basic level, if it's a classification algorithm, if you don't have negative samples, you can't really train if everything is positive. Yeah. We have a particularly bad problem in material science because people usually publish crystals that could synthesize, but they usually don't publish if they fail to synthesize a crystal. Sometimes they might. Yeah. And also, like we never know really if a crystal is not synthesizable ever, right? It could just be a skill issue. Exactly. A skill issue.

56:32Especially with the literature. Like synthesis method issue, technology. Yeah. So I think it really helps us when we do our own experiments and get negative results in the context of what we tried. So then we can even train a classification algorithm. I think there's also like another valid thing too, of when you have this sort of string of negative results, and then finally through process iteration, you're able to get to that positive result. It's a really interesting set of, I'll call it like process engineering type data.

57:02So it's like through iteration, how did you actually get to that correct result? Because so much of material science has this ambiguity in actually how something was made. And so, yeah, and so there's actually, even if there's a known material, it can be highly non-trivial to replicate that. Some things are, you know, in like a high school textbook, but other things are really at the frontier. And, you know, we're building up that know-how as well. And then building up a system that, you know, given the string of negative results, how do you actually get to that positive case?

57:36Yeah, going to the classifier result, I would almost assume that if you are doing everything in-house, most of your results will actually be negative rather than positive, which is kind of ironic because the literature only gives you positive results. So if you were sort of cold starting this, it seems like the problem is actually the reverse, that you have an abundance of negative results and not enough positive. And as you said, like if you were doing per experiment labels, it would be mostly negative. Yeah. But if you were doing per campaign and assuming campaigns and when we succeed, then it could be a bit more balanced.

Training on Scientific Processes

58:07I see. So your reasoning traces could almost be over entire campaigns. Absolutely. Okay. So this is the key thing. This type of data basically doesn't exist anywhere else. And we spend so much of our time getting the full lineage of the scientific process into the model. And so like tracking all this data, like the conversations, the intuitions, what was executed in the lab or the computations run, what was the code written, stitching all this together is so valuable. And I think like, you know, the overall goal is rather than training on the final output of science, you're training on the process of doing science.

58:43This feels very much like what you would try to do with RSI, where you are having a model trained to train better models. You're now having a model trained to make better experiments, but the difference is it's not going back into the core model for, so it's. Unless the physics models you're developing are improving chips, which then improve AI. So this is the wider. The bigger, yeah, the higher level RSI. That's the bigger. But does that mean that, is it plausible that, you know, Fable 6 or something or, you know, GPT 8 could just have a, which has been tuned on these sort of higher order reasoning traces could actually just do this without any.

59:29Mm hmm. Because it's sort of the same thinking process, right? Yeah. I mean, I think there's decision making under uncertainty, obviously in machine learning there's, there's noise in running like the AI loops. But we do think there's like a different set of challenges when you're actually like interfacing with the physical world. But also I think there's another piece too, which is getting the compression of everything into weights is still really valuable. If inference time reasoning was sufficient, all of the frontier labs would have stopped training at like GPT-4.

1:00:00Like, okay, from now on out, we're going to get really good at inference time improvements. And so we think that by, you know, because of the differences between, you know, physical sciences and machine learning, getting that compressed into our own weights will lead to different types of systems and different types of capabilities. But also we feel like even if like Fable 7 gets really good at, like even better than what it is today, it will still have to run experiments to get results. And the reason for that is right, like machine learning is really good at what's been trained on, but scientific discovery is almost by definition what you haven't been trained on.

1:00:37And that's why we're building these labs, so that whether open models or closed models can use these labs to tinker with the universe, because we don't feel like you can make a big discovery without trying things. Yeah, there's, there's not going to be like, no one's going to like zero shot the room temp supernatural. I think that would be pretty cool. Yeah, yeah, I mean, as someone who has worked closely with, you know, wet labs before, it's certainly I am more skeptical about zero shotting like scientific results than some, but it is something that I think some people might ask so.

1:01:08Yeah, I mean, I think it's, it's really important to distinguish the results we see in math and theoretical physics from the physical world. Right? These are two different things. No, no, it's two very different things. Not even theoretical physics, like so far it's been theoretical computer science coding and math. Right, right. Yeah, yeah. Maybe theoretical physics is next. It'll come. Yeah. Yeah. Can I get a sort of mental model of the levels of extraction that you can go? So for example, in terms of automation, and I'm just on this theme again, where we talked

1:01:38about the campaign, talk about individual essays and running tests and how their humans are bad at it, so we should stop humans from doing it. Are there others? Are there others? So for example, one level higher than campaign could be a physics theory that you're testing, or one level lower than campaign is what? Like, you know? And like, I just, I like to think about it from that point of view, and then think about it from like, okay, well, this is an API call now. Like you never have to touch this again. And this one is, no, no, this is still like 90% human.

1:02:10You know, and maybe we can sort of draw the map of the territory that way. Are there other levels? I can tell you some levels. I feel like you asked a good question. I don't have a very systematic answer, but like, let's talk about some levels. So one level is the atomistic structure. This is an abstraction, right? Like there's no perfect atomistic structure, anything we do, but it's one approximation. Another one is the continuum model. So this one is like not atomistic anymore, but it's a continuous mesh representing the material. Okay. And then in a different dimension, there's the thermodynamics.

1:02:42Like assuming you do this experiment forever, what would be the final state? And then there's another layer of abstraction, which is kinetics. Acknowledging that we're not doing this experiment forever, so time matters. So how quickly will that reaction happen or not, even if it's lower energy or not? So thermodynamics, kinetics, atomistic continuum, what else? Are there other levels of abstraction? Well, I mean, I think like your point about like, okay, there could be an overarching new theoretical advance that informs multiple campaigns. Kind of like as an investor, I want to go like, here's the bottleneck, guys.

1:03:18And like, we get this, we get everything else. I don't know. I mean, from the beginning, our hypothesis has been that one of the big bottlenecks is automatic characterization, because it's not that hard to mix powders to try stuff. But if you can't characterize and analyze it and then decide what the next step should be intelligently, you don't really benefit much from mixing powders randomly. So we focused a lot on automatic characterization, closing the loop with simulations, because that's how you like try a lot of things, make an informed decision, and then decide what to do the next day.

1:03:51Maybe, as a follow-up question, what is the, how do humans live in this loop? Like, at what points do you put, like drop the human in? I mean, I'm guessing you probably- You could randomly drop into any part at value, but like, where do you decide to spend your time? I mean, it seems like it might be all levels. Like, you have people in the lab who are physically moving materials, and then you also have scientists who are guiding decisions about what campaign, and then you have humans who are, what's that? I think I'm just asking the question for you.

1:04:21Yeah, 100%. I think it's always shifting too. Yeah. So I mean, I think even just using like the guidance of campaigns, in the early days that was fully human-driven. Now, increasingly, it's a mix of AI-driven and human-driven. And then Daesh was saying early on, we identified that characterization was a big bottleneck. In the early days, it was human-driven, and the scientists were very much overwhelmed, and the balance has shifted significantly towards AI-driven. And this has allowed the scientists, who were previously spending all of their time doing these refinements, to now elevate their work to something else.

1:04:55But I mean, I think it's really sort of a function of time. It's always changing at each of these levels. Yeah. In terms of just general search, are there things that perform well in, let's say, the physical world that don't perform well in the other worlds, or vice versa? So evolutionary search is pretty popular in LLMs. I don't imagine it works well here. So especially in simulations, people have been using evolutionary search for a long time.

1:05:26It's pretty good at structure prediction, for example, finding low-energy structures. Okay, so this works. And then the process engineers in the semiconductor industry will do DOE, design of experiments. And they'll often use some kind of zero-to-order, whether it's Bayesian optimization or evolutionary search. I don't think they use RL, but it's a similar idea, yeah. You have to have some kind of algorithm to figure it out. And all zero-to-order optimization behaves the same anyway, right? Yeah. Okay, I want to go back to a question I had earlier, which was one of scaling.

1:05:57And so I'm curious, what is your vision for scaling the lab? So there's different ways. I mean, coming from the biology world, there's a lot of cool tricks you can do in biology. You can tag things. You can use DNA sequencing for all sorts of interesting readouts. And if you can map your complicated assay onto sequencing, you can now scale to millions or hundreds of millions or something. Can you do tricks like this? Like, ultimately, every assay has its own runtime.

1:06:28I have a favorite blog post I would want to quote in the show notes. But, like, what does the runtimes look like for these labs? Is your scaling just for materials just fundamentally linear in terms of, you know, you just need more resources to do more things? Or can you parallelize things in clever ways and combine things? And are there tricks there? Well, I mean, I think, like, one way to think about it is the RL environment is not quite trivially run the experiment, wait for that agent to fully roll out, you know, doing the calculations, going through all the instruments to the final outcome.

1:07:06We basically start producing data across all the different instruments and all these different campaigns. And one way you can begin to expand that for machine learning is now reinforcement learning environments can be constructed based on, again, like, dates for, you know, the experimental campaigns or computational campaigns at that point. Or you can also start taking subsets of instruments and you're like, okay, we're going to create reinforcement learning environments on the basis of this instrument alone, or maybe it's across these different instruments to kind of get to some reward state.

1:07:41So that's a way where this finite basket of data, when viewed in different ways, can then be expanded for training purposes. So it doesn't get at the actual experiment run, but from an ML perspective. Maybe I can give you some analogs for the bio examples you gave.

Scaling Up the Lab Network

1:07:57Yeah. So, you know, one thing people have tried is combinatorial sputtering. I don't know if you ever heard of this. Yeah. Like, you have these sputtering targets and you create this compositional gradient. So when you look at the final product, because there is gradient from each precursor, it creates different crystals at different spatial locations. So in one go, you maybe try hundreds, thousands of crystals. This is similar, right, to your example. Yeah, that makes sense. As far as I know, this hasn't worked very well so far, because it turns out, like, there is diffusivity, so things move around.

1:08:30Another one that we thought about, but we haven't done yet, but a lot of people talk about this, is superconductivity measurements are a bottleneck. Because unlike XRD, there's no high throughput superconductivity measurement, and it can take like one hour per measurement. So people have thought about taking the different candidates, mixing them all into one sample and putting it in. And if it has superconductivity. And if it has superconductivity, you know that one of the sources did, and then you can do. Some people are doing COVID testing too. So there are ideas like this exist. Some of them work well, some of them don't. So you, I think, are about to announce a big fundraise.

1:09:03How are you thinking about scaling? Like, we've been doing our design of our labs for a while, and I think the resources allow us to continue to scale up the labs significantly, but also the compute, both from like the AI side and the computational side. But I think really importantly too, it's what are the new types of labs we're able to build. So I think we've been proving out this loop in our initial labs here in Menlo Park, but we'll be continuing to expand to new types of labs and repeat the same process.

1:09:33Yeah, like the kind of lab we're trying to build, I think, haven't been built before at this scale or at this approach. So we have learned a lot from our own build out so far, and those lessons guide us for the next version and the next version. And then we can, as you said, like increase the scale, the ambition, the quality of instruments. Some instruments are very expensive. As we get a really good understand for what we get a lot of return from, we can invest into those more. And it makes hardware engineering has become so core to it. So for example, in the early days, just for speed, we buy some off the shelf instruments.

1:10:05And as we kind of push the scale of the lab. You make your own. We have to make our own. So we realize, okay, for this instrument, it's pretty fast. These components are actually pretty quick, but this weighing machine actually becomes the bottleneck. Or there can be things pertinent to data quality where, well, the resting position of this robotic arm actually is above the plate of previously mixed things. So there's a contamination risk. So in our hardware design, we're going to redesign it so that the resting position lies away from those things.

1:10:36These are the kind of subtle details that allow us to kind of push the noise floor down for our experimental campaigns. So basically take these learnings and scale it up. So, you know, you ship your org chart. You have many teams. I think that we were surprised when we looked at your jobs page and we were like, okay, we don't think we fully map what periodic does. Maybe some organizing principles. So to kind of make sense of the job page. So we hire for AI research and infrastructure, computational, experimental roles and hardware engineering roles and then product roles.

1:11:09We want to achieve synthesis super intelligence, right? And we feel like for that we need the right chemistry expertise, the right physics expertise, simulation theory and thin films, powder. So hardware engineering, like these seem like very diverse roles, which they are, but they're all actually coherent in what they're trying to do, which is a synthesis for intelligence. And this also applies to the LLM researchers, the infra, even the product engineers, because the product engineers kind of make sure all the research gets made into a product that the experimentals in the lab can use.

1:11:41Yeah. So in order to do the end to end loop with the physical world, it really requires your ability to have agency over the physical world to build these labs, to run the campaigns effectively, to create AI systems against that, to make them good users of the computational tools. That's like the difficulty, but also the opportunity of periodic. It's like this group of people has just never been brought together before. It's irreducibly a multidisciplinary problem. Yeah. And one other guiding principle for us that is really important for us is we want the people who are very good at what they do and experience to do it hands on.

1:12:16You know, like the modern life has gotten us to this place where, especially in academia, when someone is really good at research in some area, we give them so much responsibility for grant writing, teaching, that they stop having the time to do research in that area hands on. But if you look back at like Bell Labs, IBM, institutions that made really good progress, it was really experienced people doing hands on work. Like Bardeen was in the lab every day, even though he's a theorist. Alex Mueller was in the lab doing experiments, even though he was the lab lead. So we try to do that here too.

1:12:47So we have some of the world's leaders in different fields, but they're doing hands on work. Do you want to brag or just call them out? Oh, I would love to. Yeah. So, you know, when I was doing my PhD, my favorite computational material scientist, who's kind of from my age group, is Muratan Aikol. He did his PhD with one of our advisors, Chris Wolverton. And he is kind of like our computational material science expert. He's incredible. He's sometimes, you know, a really good experimentalist, even though he does simulations. Like he understands systems really well.

1:13:17Our lab lead is Joe Chekalski, who is a professor at MIT, but he's on leave to work with us full-time. And Joe is just incredible physicist, understands superconductivity really well. But he also was a professor in Japan for a bit. So I learned synthesis and chemistry really well, especially for a physicist. We have Daniel Chica, who was a graduate student in mercury kinesthetist group, which is maybe the best cell state chemistry group in the world. And Daniel was one of his like synthesis experts. So he's like a magician with his fingers.

1:13:48I mean, there's so many examples like this. No, I know. I mean, I think like Dima Bahandau. So he's been leading a lot of our AI and LLM efforts. He was the inventor of neural network attention. Oh, Bahandau. Bahandau attention. Exactly. Yeah. Yeah. So very hands-on and sort of getting into like, you know, the traces, the training distribution, how to connect it to the lab. The chemistry. Chemistry. I mean, he'll go deep. He was very deep. He actually has like solid state synthesis books like at his lab in Montreal. Ray Nakano.

1:14:18So he used to tech lead some of the operator work at OpenAI. He's in the lab. So when we were doing some lab tours, we were initially like, oh, this new scientist, this new technician really looks like Ray. And no, it was literally, this was Ray like actually in the lab with the scientists automating the machines. And it's like, this is sort of the DNA and the type of people we need to kind of pull this off. So we also saw that you have forward deployment engineering roles. We happen to be spinning up a forward deployment engineering podcast because there's so many engineers that would do that role.

1:14:53They may not necessarily know that there even is a role for them at Periodic. What is it? And how do they work at customers? So we've been building our own products, our own tools, our own AI and computational for ourselves. So we've been customer zero. Now what we're doing is we're taking these same tools to different industries and we have a huge focus right now in the semiconductor industry. However, in order to do this, like these are very private companies. We need to be able to operate within some of the most secure environments.

1:15:25And so our periodic forward deployed engineers and researchers will actually be onsite working and actually integrating our AI systems or computational systems to help our partners get to their goal states more quickly. So basically the tools, the know how that we've built in doing our own materials discovery, materials engineering, we're now bringing to industry. And we think this is the most effective way to get partners to their end states and to their goals rather than just like throwing some technology over the wall and telling them to figure it out.

1:15:57Then these forward deployed engineers will also do the inference locally, but then also can train on the data. So again, once we take our system that understands these different areas and we deploy, we can make it expert on the customers or the partners data so that they can own their own intelligence and have systems that understand this much more effectively than just like hitting some API or some untrained model. And so this is a much expanded forward deployment engineering role than typical because this really requires some very deep machine learning expertise.

1:16:34You have to be incredibly precise from like a data training perspective, infrastructure perspective, but also physics perspective. So we're working with highly technical industries. They have to understand different areas of like chemistry is material science, some of the devices. And so those are the roles we're building out right now. Part of it is willingness to spend extended periods on site in Taiwan. It's like, okay, well, I know that kind of customer. It's interesting that like you're monetizing, that is the primary way that you're monetizing now, but it might change in the future.

1:17:08But like, I mean, you know, that's like a model that I think people don't really get about APL Research Lab. Like when you compare yourself to like a Bell Labs, that's what people are thinking, right? I think like a really great analog would be software engineering. Right. So in the early days of software engineering, like we have like, you know, GitHub Copilot, then we had early versions of ChatGPT. People were using these things as co-pilots to help them get to their solutions. And as the automation improved, now we have things like codecs and very few of our engineers are writing code the way they used to.

1:17:38And we think a very similar thing could play out for periodic as well, where you can have systems to accelerate the researchers, the material scientists, the materials engineers, the process engineers. But as the autonomy, intelligence and capability grows, you can begin to price outcomes. So help me get to this kind of goal state. And I think that's a really interesting area for us. Yeah. I mean, we kind of feel like our best contribution to solid state physics and science could be if we made these tools and topics profitable.

1:18:11Similar to how ChatGPT made CS and LLM majors way more popular in colleges before and after. We love to show that solid state physics, material science research can make a big impact commercially. And then that will attract more attention. Young people will want to study physics, which will be a dream. And, you know, Bell Labs did make a huge commercial impact, right? So they failed to commercialize some of their incredible advances. They, of course, did some research that just couldn't be commercialized, like the cosmic background.

1:18:41But they did benefit a lot from like the vacuum tube connecting east coast to west coast by phone lines. Yeah. And I think it's really also interesting to think that technology and capital are incredibly intertwined. So if you look at what was the progress on chatbots for many years, like imagine recounting, okay, what was the progress from chatbots from, I'll say, 2010 to 2015? We'd be kind of at a loss to say over that five year period, how much did they improve? Whereas if you look at the period from 2021 to 2026, you know, it's night and day.

1:19:16And what happened was ChatGPT and these other systems were able to achieve a product market fit and it changed the capital landscape entirely. This changes the landscape for hiring, compute, data, et cetera. The virtuous cycle. Exactly. Like success be success. Exactly. So technology, it's incredibly coupled to the capital and those resources and we want to achieve the same thing in the physical world. Yeah. So one thing I've been seeing is there's this move from science being funded from, you know, government grants and so on to, you know, VCs and private funding.

1:19:51Do you all plan on making anything you're doing open source, like either releasing data sets or actual models? Is this something like in the future, even, you know, there are many models which may be not your state of the art, but which could still be, you know, useful releases. Absolutely. I mean, so we have a couple of things. So we have been contributing to open source to kind of like PyMetGen, the materials project code base, custodian, their DFT runner. TorchSim was actually created by one of the researchers here, Abhijit Gangan and he maintains it still.

1:20:25JaxMD, Abhijit maintains it. So we do a lot of open source contributions. SGLANG, Megatron. Yeah. Yeah. Yeah. So we've been very active contributors back into open source. And we also have an academic grant program where we give academic gift grants to academic groups in universities that we feel like are really advancing this direction towards synthesis superintelligence. So that's also been great to see. I think the first paper from this funding is about to come out. So it will be exciting to just get more papers come out like this.

1:20:56Yeah. Let's still man this, say that you've successfully have 100% lab automation, which does everything, characterizes everything correctly. You have all of the fun agents, which can do everything. I still am somewhat unclear about how you actually get to superconductivity. There's a lot of hard problems to solve. What is the path there when there is no sort of like theory for model for most of the strongly correlated systems or high temperature systems? Yeah. So I mean, our view is that it's not hard to think of chemical spaces that would host superconductivity.

1:21:30What's really hard is to synthesize them. That's why we're really emphasizing synthesis superintelligence. So maybe one example to give historically is Alex Mueller, when he thought that transition metal oxides might be a good path to superconductivity, he was actually trying nickelates, the nickel oxygen and some cations, and that didn't work. And then he tried copper, just copper oxygen and some cations, and that worked, and then he got a Nobel Prize. But then we know that later on, it turns out nickelase was a good idea. I mean, nickel and copper are next to each other. There's only two layers, right?

1:22:00Yeah, exactly. There are only so many 3D transition metals. And then Harold Huang, one of our Stanford professors here, who's incredible, he realized he can make nickelase in thin film form and show the superconductivity. So, you know, if we have a really good synthesis superintelligence in the lab and we can really scale up these things, I think we won't run out of ideas or the LLM won't run out of ideas about what directions to try, like 3D transition metals, oxygen. Like there are very related ideas here.

1:22:30Like many people have been trying like cobaltates, which is like instead of copper or nickel, you try cobalt. So I think that's super exciting. Like maybe one historical example to also study is the Japanese group that discovered magnesium diboride. As you know, that's the highest temperature, ambient pressure, conventional superconductor. And they discovered it by just trying a bunch of materials. I think they tried 30,000 different things. 30 of them seem to host interesting superconductivity and MgB2 was one of them. But magnesium diboride sat on people's shelves as a precursor for decades before then.

1:23:04And BCS theory had been invented back in 1957. So the way people found these materials wasn't directly from theory. It wasn't because they couldn't synthesize until then. It was just like they tried a bunch of things. So now imagine if we had this perfect automation that you're describing, which sounds like a dream. And we tried the 30,000 the Takedo group tried over his career in a month. We just really increased the surface area for luck. No, but it's very inspiring what you've done. Congrats on all your success. I feel like you're creating the modern playground.

1:23:36I imagine recruiting must be like super easy for you. So I'm just like a little bit jealous, but you've worked hard for it to get here. So yeah. Yeah. Well, thank you so much. Yeah, it was great chatting with you both. Yeah, super fun. Thank you.