AI科学家已到来：自主实验室与合成超智能——Periodic Labs

Okay. We're here at Periodic. We're here today with Liam and Ekin Dogus Cubuk from Periodic. Welcome, and thanks for having us at your—

Yeah, great to be here, and 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?

1. Intelligence Meets Reality

I think that's the thesis and premise behind Periodic that Dogus and I had when we were building this: 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 make progress, you need to create these conjectures and then actually see whether or not they hold.

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. You really need this iterative process. That's our core belief for putting this together, and that's why, from the very beginning, we're like, “We need to have the AI systems, the simulations of the physical world, but then also build up the physical high-throughput experiment.” We think a different type of intelligence emerges from that.

I think it's also notable that with OpenAI and Google, you don't have the resources in those big labs. You had to come out and do your own thing because you're kind of inventing your own playbook as you go along.

Yeah, a team like this has never existed, I think. We try to bring together solid-state chemists, solid-state physicists, experimentalists, theorists, hardware engineers, LLM experts, computer scientists, and others. Some of these technologies are very recent, like high-throughput experimentation. Robotic arms have only been tried in the last 3 years or so, and some of these force fields are also very recent.

We felt like having a lab was very important. Having this focus and bringing all these people together to work together is very important. I guess the closest examples from history are places like Bell Labs, where incredible theorists, experimentalists, and chemists worked together and achieved incredible things. So we're trying to do the same here.

I'd like to talk about force fields in a bit—

Yes.

But before we get to that, I think, for context, one of the biggest differences—for some of our audience, there are AI engineers, and some of the audience is scientists—is what happens if you've been working at OpenAI as a big lab? What's the difference between what you do there and what you do here? I think you've already hinted at this, but explicitly, how do you have to reframe your thought process?

2. AI Under Experimental Uncertainty

Well, I think one interesting thing is that now our reinforcement-learning environments literally derive from the environment of our physical labs. Our data comes from our physical labs, and this is our ultimate truth. It's not enough just to do optimization against some answers that were known in papers or textbooks, because we're going beyond that, and I think that's one of the biggest differences.

Along with that, there are issues of decision-making under uncertainty. When you're doing optimization against math, there's a high precision to it. You're not really dealing with variance or uncertainties or aberrant measurements, whereas that's very key to our process. For example, when we are doing a materials-discovery loop, things don't come 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. We thought we ran it at this temperature, but really the temperature was some delta away. An intelligence that can take in this noisy data and make intelligent decisions like a scientist would is a different set of reasoning strategies.

I think there's a huge amount of commonality: standard mid-training, reinforcement learning, the construction of tool-using agents, variance reduction, and making sure the infrastructure doesn't have mismatches between training and inference. But we have to go beyond that and really think about how you do accurate work when there's a high amount of uncertainty, and how you make incredibly efficient use of a limited set of data.

We're scaling things up significantly, but it's still very different from 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 that has been on the science podcasts for a bit. When you talk about experimental uncertainty, I think it might be helpful to explain what a specific process that you were doing experimentally would look like. How would a human approach this problem? Before we jump into how you approach the AI side, do you have a specific example of a type of material you make and what sort of uncertainty in the measurements you would experience in the process?

I think there are many dimensions to this. One of them is that every scientific experiment has to do some dimensionality reduction, and this is very different from coding or math. 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 and some corollaries people have derived from them. It's a bunch of functions and API calls. Everything you need to reason through is available to you, and then you just have to be very smart and figure it out.

In physics, as you know, we start with more atoms than we could ever store on a computer. 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. This was, I guess, the original premise of thermodynamics. They were originally really confused about how steam engines worked, but then it turned out you can find 5 or 6 thermodynamic variables that explain what's going on pretty well, which is incredible. That's the beauty of physics.

You have a room full of atoms, and there are 10 to the 23rd or 10 to the 27th or something atoms in this room, and yet you need temperature and pressure and simple variables, and that tells you everything. Well, it tells you mostly what you need to know for most cases, right?

A lot.

Let's say you make a material and you could do something like look at an X-ray structure. This thing is a very lossy projection, right? It's not actually a structure, or it doesn't tell you the true structure; it tells you some principal components of that. How would a human take these simple principal components and reason about them, and then how do you extend that to an AI?

On the theoretical physics side, we've decided that energy, the average value of energy, and the fluctuations of that energy are very important. A lot of thermodynamics has been derived from this, and then we realized reaction barriers are also very important; the kinetics are very important. A human can try to reason through this extremely complex system by coming up with some reduced-dimensionality descriptors.

People also really like an atomistic picture. Do you think about what the local neighborhoods look like, whether they're octahedra or tetrahedra? Chemists especially benefit from that. That's on the theoretical side.

On the experimental side, you would try to collect as much data as possible. You'd have a lab notebook, and you'd write everything you see, with the understanding, of course, that you'll miss a lot. 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. Every time you use the furnace, it gets worse and worse. Another issue is that you have a furnace, but the temperature isn't perfectly uniform, so where you put the material in the furnace will affect the result.

We don't have this issue yet, but we will maybe someday. Optical instruments suffer a lot from vibrations, so every time somebody walks by, it can affect them. I remember when I was doing my PhD, one lab had this big issue because sometimes their results would be very different from other times. They traced it to a certain time in the night when somebody upstairs would be walking, and that vibration affected the laser setup.

Science is very hard, but that's what makes it special, right? The other thing Liam and I have been talking about is that 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.

There’s a lot of uncertainty, a lot of noise, and a lot of missing context, but you have to be the intelligent being and figure out what to do next.

I think a really interesting point to build on that, too, is that the optimal reasoning strategies to do math, theoretical computer science, and these types of things may not be the optimal reasoning strategies to do science.

Yeah.

Mm-hmm. How do you define an RL reward function when your output is a noisy crystal structure or some number of, I don’t know, collective variables or something?

3. The Materials Discovery Loop

Maybe I’ll give one part of the loop, and I’ll just dig into that. As part of the materials discovery loop, we first have to figure out what to make: the stability of the atoms together, as well as whether we expect it to have the properties we want. We don’t just want a novel configuration of the atoms; we want them, when put together, to have the property of interest.

Mm-hmm.

Next is: how do you synthesize it? How do you actually make the thing? That’s also nontrivial. So even if you know something will hold together, the process of actually figuring out the processing conditions to bring that forth is highly nontrivial.

But once you’ve done those 2 steps and made something, it doesn’t come off labeled. You actually have to characterize it and figure out what you made. In that case, rather than thinking about the RL environment as, “We’re just going to kick off an experiment, wait a couple of days, and try to update on whether you found a room-temperature superconductor or not,” that’s completely infeasible.

Yeah. I was going to ask that. You can’t leave your GPU sitting for a week or something.

It’s completely infeasible—

Yeah.

—it’s completely infeasible because you don’t have enough agents to reduce the variance sufficiently. The rollout time is spread across the different instruments. You’re waiting. There are just some physics limits of the experiments, like the actual synthesis or furnace time. So it’s too slow and too noisy.

The way we think about the AI program is basically constructing agents around this set of data that we’ve produced. In the case of characterization, you’re looking for RL environments that reward the identification of the phases actually present. Given the raw experimental data, are you able to fit that effectively?

The way we do this is we shoot X-rays at the material. X-rays have a frequency or a wavelength roughly commensurate with the spacing between atoms, so you get nice diffraction patterns, and that kind of gives a fingerprint of the crystal structure. Identifying what’s actually present is highly nontrivial, and some of our early AI work has been on producing systems to do this.

This allows us to move much more quickly, too, because if you’re running so many experiments, you quickly become bottlenecked on your ability to understand what was produced. But again, now the reinforcement learning environment is fairly straightforward: you’re going to reward the identification of the phases present, and you’re going to penalize spurious phases or things that don’t have any real chemical plausibility.

That’s a succinct example of how you can do this for the different pieces, and when you stitch this together, that’s the end-to-end discovery loop.

Okay. You—

And maybe another—

Yeah. Oh, sorry.

Another piece would be that, as we build up a large basket of experimental data, you can think of another thing as basically timestamping the state of the world. You could say, “At this date, this was our experimental evidence up until that point.” You can construct reinforcement learning environments where you say, “Okay, given that body of experimental evidence and given the choices, what was the next choice that the scientist made, or what was an outcome of that experiment?”

This is really interesting, too, because it allows us to do types of programs that would be harder externally. If the pretrained 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 work.

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 do the calculations or simulations. It gets to the right answer, and then you reinforce that policy and upweight those reasoning strategies. Those reasoning strategies will not generalize to novel systems, so they’re not useful to us.

But as we build up baskets of experimental data, stitch things together, and have the lineage through the whole process, that gives us the ability to construct new types of environments that are just not plausible elsewhere.

I want to ask my question—

There’s like—

—but I’m worried that I’m taking it off the rails. I’m just going to go for it because, again, I’m a non-scientist, but I come from an engineering perspective. I’m familiar with the ML side, but not the physics side.

4. Physics Is Not Finished

There are a few versions of this, but I think the basic question is: don’t we have most of the laws of physics worked out? How come we don’t have perfect simulators already?

Oh, that’s a great question.

This is very basic. He was like, “Ha ha, that’s cute.”

No, it’s a great question.

I’m like, “Dude, I have all these physics books. What do you mean we’re not done?”

Yeah, so there are a couple of things going on here, right? First of all, we’re definitely not done with any of the laws of physics. Even the—

At the quantum scale, okay. And maybe at the very, very large scale. But at human scale, at material scale, what’s left?

So it’s a really good question, and it’s very interesting why we’re not done yet. There’s a very common story people talk about. When Dirac was figuring out quantum mechanics in the late 1920s, he wrote a textbook about quantum mechanics, and he kind of framed it as if it were all done. There’s a phrase people like to quote—I’m not sure how accurate it actually is—but apparently Dirac said, “The rest is chemistry.”

The idea was that he could solve the hydrogen atom.

Just compose everything.

Mm-hmm.

Yeah.

Yeah.

He could maybe solve a chain of 1D hydrogen atoms, but he couldn’t currently solve, say, nitrogen interacting with oxygen. But that’s okay—that’s just chemistry.

Yeah.

Um—

Physicists really love single-atom representations.

Yeah.

Simple systems. And that’s all you can solve.

Spherical cow.

Yeah, spherical cow.

But it turns out, I think what we learned in the last 100 years is, first of all, that’s not true. It’s not just a simple extension of what it was to do the hydrogen atom, and the rest is not just chemistry. There’s actually a lot of physics there.

I guess one of the things we still haven’t figured out is high-temperature superconductivity. But there are a lot of other—

And high for you is like 175 to 200?

When people say high-temperature superconductivity, what they mean there is unconventional superconductivity. There’s conventional superconductivity, which is mainly driven by electron-phonon coupling.

In that case, you can kind of see the isotope effect. If you take the same system but just use a different weight for one of the elements, you can see the superconducting temperature drop at the rate you expect. So that one is conventional.

But higher-temperature superconductors like cuprates, the ones that are above 77 Kelvin, like 93 Kelvin, turn out not to obey that physics. But we don’t know what physics they obey. They’re just incredible superconductors.

Yeah.

But it’s not just that. That definitely is one of these popular topics where we don’t know the theory yet. But there are so many other things we don’t understand. For example, even strong correlation in simple quantum mechanical systems—we cannot simulate that yet.

Density functional theory is a very powerful tool. It’s incredibly accurate on some things, but as soon as there’s some strong electron correlation, it actually fails to capture some of the effects. 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. Every improvement we make here should improve human life because the better we can understand materials and solid-state physics, the better devices we can make for them.

5. More Is Different

There's this famous quote by Phil Anderson, which is “More Is Different.” Basically, you may understand all the basic laws of something, but when you add many things, they behave qualitatively differently from how simple physics should tell you. So understanding how lots of things behave together seems like it should be simple. The basic laws are simple, but the collective behavior is so complicated—

Yeah.

—that it's really hard to model.

It's emergent and universal, which is crazy. We're also seeing this with deep learning models, right? We see power laws everywhere in deep learning. We don't understand them, but it's very reminiscent of physics, where there's emergence and universality.

There have been many theory papers in the physics world about emergence in neural networks.

Yeah, absolutely, yes.

Yeah.

My other follow-up question was also just on the process and the end result. Let's call it the process and the end result. 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 train a model that perfectly reverse-engineers 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 there 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 materials, basically.

Yeah.

What you're asking is whether there's a forward model where you put in an input and can predict the output, and you want to under—

Yeah.

You want to create a surrogate model which can understand—

The input is the final state of that configuration of atoms, and then the prediction is—

Right.

Right.

Because then you can split up the work, almost. You can—

Right.

—have one team do the process side, and then the other team do the property side—

Yep.

—and then, you know, just raise them.

Yeah, I think there's a lot of progress we can make in terms of splitting up these agentic workflows into these different areas. I think there's a lot of empirical work we can do on synthesis prediction as well.

Yeah.

Taking in that configuration of atoms, what tools can you use? Given this experimental evidence, prior literature, and prior papers, how do you actually get to these things? And you ultimately are testing: Are you able to actually replicate it or not?

The reason I say this is maybe also just thinking about it in terms of a company. The way that TSMC is kind of like a fab for semiconductors, but they don't design the semiconductors, you could be the TSMC of materials, where people come to you with, “Here's what I want,” and then you can figure out the process—

Yeah, it's like a matter compiler.

Yeah.

Given these requirements—

Very Star Trek.

Yeah, yeah, yeah. Given these requirements, is this actually a valid set of configurations?

Yeah.

Can this exist?

Is this a good objective function? I don't know. It's too, too big.

Yeah. We kind of talk about one of our missions as synthesis superintelligence, so I think it's in line with this.

Yeah.

I want to double-click on something that you said a little while back, which was that you talked about phases. I think this is going to what Dogus said about emergence. The concept of a phase transition might be foreign to a lot of the audience. Can you explain what a phase transition is, and then why this is a helpful signal for something like an RL environment?

6. Reading Phase Transitions

Phase transitions are fascinating. I think maybe the one that people can most easily relate to in their lives is ice melting or water boiling. You increase the temperature of ice, and it still looks like ice and 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. I think computer scientists and mathematicians also study this, maybe under different names, but basically, you can have spins that are up and down, and they have different interaction terms if they're both up versus if one is up and one is down. At high enough temperatures, they're usually just randomly up or down; entropy wins. You lower the temperature, and it still looks the same. But then, at some point, suddenly they start aligning. That's the phase transition.

Phase transitions are really nice because they make physicists' lives easier when studying certain things. They tend to have certain spatial correlations that really help us as well. In our lab, phase transitions usually come about because we mix precursors that are different crystals, basically, but then we raise the temperature or do something else to encourage them to react. Then atoms start reacting, and they form into this new crystal.

This shows up in the XRD pattern because usually, if the geometry of the atoms changes drastically, the XRD pattern changes a lot. But what really happens in a typical practical materials discovery campaign is that when you first try something, it doesn't work, and the outcome isn't just a clear yes or no. It's usually a very mixed phase. It will usually have some of the precursors. It might have some amorphous phase, and then it will have a bunch of phases that you maybe didn't try to make.

That becomes a challenge, and that's why we find it very difficult to bring together simulations and AI to do the characterization. The simulations can always say, “This phase doesn't look like what we predicted, but this is a slight variation of it. Let me now do a force-field calculation to see if it's still stable.” Or the AI can say, “Based on the previous experiment and this one, this is probably not the phase we want to make, so let's change the conditions.”

For listeners, XRD is X-ray diffraction, which you already described.

In some cases, there's a phase that just hasn't been recorded before. It doesn't exist in any papers or databases, and so the AI has to actually use these tools to say, “What configuration of atoms would actually explain these phases?”

This becomes incredibly important for directing the scientific process because, let's say we have a phase target in mind, we really need to hill-climb that to get better measurements on it. If it's 1% and we're like, “Actually, we want to increase this phase purity,” we want to have a very good measurement of this experimental data to say what is present and what's not.

So one of the disadvantages is that you have a distinct variable that you can observe. Going back to the statement about how you deal with uncertainty, it seems like part of your answer is that you make your observable signature somewhat unambiguous. There's no interpretation there. Is that right? Is that the logic behind this, or is it literally that you just need to define a new phase, and so this is the thing you care about?

I think there's still ambiguity in a lot of these cases.

Yeah.

A very dumb thing is replicates. We run replicates, of course.

Yeah.

But no, I think it's a challenging task because it's not fully deterministic, and there are ambiguities and other things. Two different phases can actually be consistent with the same pattern.

Yeah.

So I think that's why it's really important for the system to use some chemical intuition to say, “Okay, well, this would be highly unlikely given the synthesis conditions and the prior knowledge.” That can be used to help disambiguate.

So you're somewhat injecting priors there, based on what you expect.

Yeah. I think thermodynamics is the biggest prior, right?

Right, you just—

Right.

And then physics is a big prior. Another one that we really benefit from here is multimodality in materials characterization. We can do XRD, and yes, some phases might look similar on the XRD pattern, but then we can also measure their electrical properties and magnetic properties. We can measure their morphology, like with some kind of electron microscopy, and then some of the phases that look similar on XRD will look different in some of these dimensions.

Yeah.

And that kind of multimodality really helps. Again, AI is really helpful here because humans have limited context and computational power as well. If you give humans 10 different modalities at the same time and say, “Analyze these all consistently,” it’s a bit difficult. But it’s actually not superintelligent for an AI to do it. It’s great, yeah.

Yeah, it’s like the superintelligence is that it can do so many calculations. It can look at everything, and this kind of gets to that phrase about better decision-making under uncertainty, where now it’s stitching together the signal across many different instruments longitudinally, across different experiments, across the replicates, and then getting to this underlying, better model—

Yeah.

Rather than just over-indexing on one measurement from one instrument.

And let me just— 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. At some point, you have to throw out data, even though you have all these replicates, even though you have presumably 10 different ways to measure the same thing. What if one of your temperature things is wrong? When do you throw that out? How do you decide what to do? As a data-hungry ML guy, I just want everything, right? Then I just model, learn everything. I don’t know what I don’t know, and then just throw the machine at it.

So we don’t throw out data. I think maybe what I meant was potentially, 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. 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. Unfortunately, we’re still in a regime where any data we collect is very precious, and we don’t delete any of it. Yeah.

Maybe a way of thinking of it, to connect to the AI side, is to 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 8 bits or something of output. You’re taking this humongous thing and coarse-graining it down to a little bit of information, and then it’s now either traditionally humans’ job or now your agent’s job to reconstruct from those 8 bits of information—

It’s analogous, yeah. You can’t have full visibility over every parameter, over every activation for that input. You’re getting some sort of subset of it, or some coarse-grained features across the whole thing.

But you probably do find Maxwell’s demon argument very interesting, because that kind of—

Can you recite it? I don’t—he’s mentioned it before. I keep forgetting it.

In the 1800s, people realized that the entropy of a system has to increase or stay the same for a closed system, but it can’t go down. This is the second law of thermodynamics. There was a thought experiment that asked, “What would happen if there was a very tiny, all-knowing demon that could look at atoms, and then whenever an atom had low energy, open a door and let it into this other room, and keep doing this to reduce the entropy of the system, which would break the second law of thermodynamics?”

For a long time, I think there wasn’t a very satisfying answer for why that was the case. But then there was a more recent paper in the 1960s—more recently to us—where Landauer came up with the argument that you’d have to spend energy to delete information. Maxwell’s demon would basically have access to so much information by doing this thing he was 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. Even though we’re not at the level where we can be Maxwell’s demon, I think your question is very visionary toward a future where we store so much data that we have to delete some.

Yeah, no, I wasn’t obviously going to that extent. It does remind me of air conditioning. That’s kind of how air conditioning works. But also, I think it’s just a question of why don’t you just buy every sensor in the world and ridiculously over-instrument and over-replicate everything.

Oh, we do that.

Right? Okay, yeah. Okay.

Yes. Yeah.

Anyways, 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 hidden variables, if you will.

Yeah.

Um—

Yeah, and just out of curiosity, do you worry that you’re just in California? Don’t you want to do this in Nepal and Australia?

Yeah, we could. I mean, we have plans to open more labs.

Because obviously gravity, where you face, and where you are matter.

Also, expertise can change, right? Here we have certain kinds of expertise. For example, the Stanford Physics Department and Engineering Department have certain kinds of expertise that really helps us because we can hire from there, and our researchers can collaborate. As you said, geography can matter, but different cities can have different sets of expertise within physics and chemistry.

Yeah. Oh, I mean, we have Zoom for that.

Yeah.

You know?

Yeah.

But still—

It seems like your experiments are going to be affected by where you are.

Yeah.

Yeah, I mean, the future of Periodic is not a single lab.

I mean, I think—

We need to have—

I think you’ve got to hit up your buddy Jeff—

Yeah, a huge—

—and go to space and, you know?

Right. Right.

Yeah, I mean, I think it’s necessary to have this sort of network of labs. As you’re just kind of pointing out, there’s an optimal place for each of these based on expertise or materials constraints. There are many different constraints, period.

7. The Simulation Experiment Loop

Okay, so one thing I’ve been curious about is how you deal with computational tools, or how you integrate computational tools into a decision-making workflow, especially when, let’s say, computation and experiment don’t necessarily meet. It’s also theory in terms of high-level understanding. How does that sort of work together?

Yeah. One thing that’s really helpful here is that 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 an experiment and more accurate in an experiment. In general, of course, experiments are more accurate, but certain measurements are very hard to make with an experiment, so you make a bad version of it and get inaccurate data.

Let me just give you some examples. One of them is that we use density functional theory to estimate formation enthalpy of materials. And—

Sorry, quick pause. Can you explain density functional theory and formation entropy?

Yeah, of course.

Sorry, enthalpy. Yeah.

Yeah. Density functional theory is probably the most commonly used simulation method for materials. It kind of comes from the Hohenberg–Kohn theorem, and Kohn got a Nobel Prize for this. He realized that quantum mechanics is very expensive because you basically have this exponential Hilbert space for every wave function, so you have to solve this very difficult problem, especially as the system size grows.

Kohn and Hohenberg realized that 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. The theorem is actually pretty simple, so even an undergraduate in physics can understand it. The idea is that all the properties of the ground state of a quantum-mechanical system are functionals of the charge density, which is amazing because charge density is just this 3-dimensional object, whereas the Hilbert space and wave functions are exponential.

That was a really interesting observation. Then Kohn published another paper, I think with his postdoc, Sham, on Kohn–Sham wave functions, which provided a practical way to solve an approximation of the quantum-mechanical properties of materials. Today, we use it very much, both in our lab and externally.

Formation enthalpy is one of the measurements we can make to understand the energy of the material. 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 combine into, then it’s unlikely that this material will be able to be made. So stable just means it’s on the convex hull of other materials and their energies. Was that clear enough?

I mean, maybe a simple high-level point. Density functional theory is a way of taking something that is exponentially hard in the number of electrons—or, you know, in your system size—and reducing it approximately, with some introduced error, to usually an N-cubed approximation.

Right.

So now you can compute quantities which are just uncomputable, but at some cost. Right? If you could do it perfectly, we wouldn’t need the lab. This may be a good callback to our second episode ever. Heather Kulik famously said, “There is no AlphaFold for materials,” not just computationally, but because the ground truth just doesn’t exist. We don’t really know what material crystal structures look like, and DFT is generally our best route toward them.

Yeah, I think so. Maybe there are 2 things I would add. DFT, by definition, doesn’t have to be an approximation. The Hohenberg–Kohn theorem shows that—

It doesn’t—

It can be exact. But I think, as you’re pointing out, we don’t yet have access to the exchange-correlation functional. And then, for charge-density-based models, we don’t know what the kinetic energy functional is. Even if we had perfect DFT, I still think we’d need a lab because we cannot fit 10²³ atoms in the computer.

Yeah.

Um—

You’re very obsessed with that number.

No. Well, okay. Okay. Yeah, but with exponentially growing compute, you can argue that with N-cubed scaling, eventually you could, in principle, calculate things. You just wait until—

Oh, yeah. We need massive computers, yeah.

Long enough. But, yeah. As long as N-cubed is a good scaling, you can—

As the non-scientist, I’ll make an observation, which I don’t know if you want to throw around or anything like that. 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. It feels spiritually similar to the VAE, in terms of condensing all these things into a single sort of parameter. I wonder if there’s just the same trick everywhere. Yours sounds a bit more high-dimensional than mine. The VAE is just one sigma, one sigma thing, but it sounds the same.

The charge density file is still quite a big file.

Yeah.

Right.

You hand-wave a bit, but it’s good enough. This is the highest-order bit.

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. Going back to your question, the reason we use them is that, for certain things, it’s easier to simulate than experiment, and they complement each other well. We can do this in a loop. 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. Compute is much easier to bring online than bringing on new labs. How do you avoid having your models over-index on that and still stay grounded, given that the real world is ultimately the truth you care about?

You scale up the lab as well.

Yeah, you scale up the lab.

Right. Yes.

But I mean, you could imagine scaling up DFT to millions, hundreds of millions, given modern compute, whereas your lab, I would imagine, is doing hundreds or thousands a day. I don’t know. Something still seems fairly limited by comparison, right?

Yeah. One thing is the humans, but also the LLMs are 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. Microstructure is the idea that DFT usually simulates the perfect crystal, but crystals usually aren’t perfect, and their microstructure, which is the structure at a medium order, can really affect the properties.

Other issues, of course, are that some properties, like superconducting temperature, cannot be easily simulated by DFT. Again, even if you do 100 million trials with DFT, it won’t be enough. So we have to rely on heuristics, experiments, and so on.

And so there’s a huge filter—

Mm-hmm.

From all those calculations to what actually gets executed in the lab.

Is that human-mediated, or LLM, or a mix?

It can be a mix.

Yeah.

For example, for a long time, Materials Project, which was the best open-source DFT database, used experimental calibration for DFT. They would actually use experimental data to calibrate DFT. We also do this internally here. 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’re always very confused when you say, “In materials, you can’t just do an XRD structure and actually know what the structure is.” Whereas if you’re in the biology world, you can look at a crystal structure of a protein and typically reconstruct it, more or less, with sub-angstrom accuracy.

What is the difference between materials X-ray diffraction—experimental XRD for materials—versus the biology world? What information do you lose, and why is this a lossy projection?

It’s hard not to sound crazy when answering this question. But do you not feel like organic chemistry and biology almost stem from a lower VC dimension, like a lower Kolmogorov-complexity model, where you can almost represent them as a 1D sequence, and it can almost be compiled?

1D, yeah.

Yeah.

Or DNA or RNA, or—

I think most people would argue it’s more 2D, but—

Okay.

Yeah.

I’m not a biologist, so—

Yeah, yeah.

I’m sure they’re right. But with inorganic chemistry, it’s virtually crazy. It’s clearly not coming from such a low Kolmogorov-complexity model, because 3D inorganic crystals can host metals, insulators, superconductors, and diamond. It’s really crazy different. People tried, but we could never find a 1D representation, right? It’s very hard to find a SMILES string for 3D inorganic crystals. And the way the atoms interact with each other is quite strange: covalent bonds, ionic bonds, metallic bonds.

This is very interesting. 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. Basically, lithium will form these structures that go into a solid electrolyte and crack it. You look at that and think, “This is so primitive compared to biology.” In biology, these systems can achieve such complex nanotechnological progress, whereas we can’t even make 2 interfaces work with each other because lithium will break them.

But I guess inorganic materials 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 trade-off.

Yeah. We’ve talked about this with several other guests. I think one of the big differences, in my mind, is that biology has a toolkit that you can basically borrow from millions of years of evolution, which has already given us all the tools, and we can just reproduce those. But evolution also constrains the VC dimension to be fairly low for proteins and so on. The emergent behavior of large-scale systems like cells can be much more complicated.

Yeah.

And consciousness.

Yeah, yeah, yeah. But maybe, more specifically, if you have a perfect crystal and shoot an XRD at it, you get the spectrum.

That still doesn't actually tell you uniquely what the crystal structure is, right?

No.

What?

It's just average.

Yeah.

It's just average behavior.

Oh, yeah.

But, for example, disproportionation is a very complicated thing, right? It's possible that you have the perfect crystal, and on average it looks like this perfect XRD. But in reality, there is some other pattern going on where atoms slightly shift to the right or left, but on average they're in the middle. It's a difficult life.

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 towards single-crystal XRD, and that can help. When we started The Science Pod, there were all these elements of AI in science, AI in math, AI in physics, and all these things. We were in this kind of pissing contest over which is harder. I think you can make the scientific argument that materials science is the hardest. Well, we didn't pick it necessarily for hardness.

I mean, there's actually a lot of areas where it is easier. We have these great simulators for large classes of materials, whereas in biology, it's incredibly difficult to model a cell, an organ, or a full organism. So I think that's a huge advantage.

That isn't interesting—that at the smallest level, it is one-dimensional or two-dimensional, but then it has many more orders of scale and structure, and that actually introduces the complexity. That's weird. Whereas maybe the complexity in materials is more the microstructure.

Yeah. You still get different complexity at the different length scales for materials as well. So, yeah, microstructure and beyond.

Yeah.

Yeah.

Just as a brief palate cleanser, in terms of science fiction, let's say you can invent whatever materials you want. What is more valuable? I have a list here. There's obviously room-temperature superconductors, but you mentioned batteries. I've always thought batteries: just do better than lithium-ion, and you're good. That has stood around for 100 years.

The last one is carbon nanotubes, mostly for space elevators. I don't know if those are three that come to mind. Are there any that people talk about in your world that are the dream?

Mm-hmm.

Definitely superconductors. Magnets, especially if you can lower the dependence on rare earths or transition metals that are hard to source, like cobalt—reducing cobalt in batteries.

One of the things I think is fundamentally important is if we can get close to the Landauer limit for compute energy efficiency.

Can you explain the Landauer limit?

Yeah. If you look at how much energy we spend per FLOP or computation, it has followed Moore's Law-type behavior, but it has gotten exponentially more efficient. So we've been spending exponentially less energy per FLOP or computation. I haven't checked recently. This was the case 10 years ago when I last—

Yeah, that's great. I remember when I first learned about it, I looked at that, and we were what? 10 or 20 orders of magnitude away from it or something.

Beyond the limit, yes. Right.

And the last time I looked, we're weirdly close. We're within a factor of, I don't know, 3 or 4 orders of magnitude away.

Yeah.

It's just exponentials rolled out over years.

I mean, yeah. Exponentials over many years are—

Yes, right.

Pretty impressive.

I mean, yeah, since I discovered this—what, 10 or 15 years ago?

Yeah.

Yeah. No.

One of the things we have to do is dissipate heat because, as these computations are happening, they're producing heat, which has to happen, but then we have to dissipate it. So if we can get close to the Landauer limit for how efficiently we do computation, that's amazing for humanity, right? That means we're doing computation, which is probably one of the most fundamental things we do as humanity, as efficiently as possible from an energy perspective—which is the real currency in the universe.

Does that get affected by quantum computing? I'm just going to throw it out there. Theoretically, it could enable massively parallel compute—very embarrassingly parallel compute.

It definitely gets affected by reversible computing, but honestly, I'm not an expert. I don't understand. If you can do reversible computing, you're not actually having 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 still have to obey thermodynamics.

It's unitary, so it's reversible.

Yeah, exactly.

One of my most memorable conversations with Elad Gil was that he was actually very skeptical about quantum computing. He said, “Even if you had it today, there are no applications.” I said, “Whoa.”

It breaks security, I always say. That's it.

I agree with that. People talk about how, if you had a quantum computer today, you could simulate things so much better. I ask them, “Okay, let's accept that. What would you simulate?” You're still simulating a perfect crystal. You still have the issues that DFT has if DFT were perfect in its prediction ability.

I do think people are glossing over some things because quantum computing is so exciting. If we can compute in a quantum logic space instead of classical logic, it's 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.

8. Intelligent Instruments

Good. 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—just basically everything you've done here. My favorite quote from you was that every piece of equipment in your lab is going to have 140 IQ. 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. For one set of machinery, we would have technicians and scientists looking for particular morphology and trying to see what we were actually making and whether it was consistent with our intentions.

As we scaled up the lab, we quickly came into these bottlenecks where it just wasn't keeping up. So we started using programmatic approaches to capture the data on SEMs. The data that was being surfaced from this very simple program, where it takes a field of view, captures it, zooms in, and captures it again, was not all that useful. It was a little too dumb for what we really wanted to get.

At that point, it became really pertinent to build AI systems directly onto the machines to start controlling them. Now they have the full context of what we're trying to achieve: What was the intent of the experiment? What were we trying to synthesize? What was some other experimental evidence? The system is actually looking in the machine and capturing that data.

This is really valuable now because we've been talking about these hidden variables and not being able to capture everything. If you can do more intelligent data capture at the time of the experiment, your data for future AI systems and future computational predictions is that much better. You'll have a richer set of data.

That's what we mean by everything in the lab having to be incredibly intelligent: to make the data as useful as possible. That was the motivation and inspiration behind it.

What's the state of the art there in terms of putting intelligence on every device? Well, I think one interesting aspect is what's the latency—

Exactly.

—of controlling that instrument.

Because there's cloud latency, but then there's also device compute, which also—

That's right. And then there's a timescale associated with the different physical processes. If calling out to an API or something is simply too slow, so the amount of reasoning, tokens, or tool calls is not matched to the latency of the actual process, then that's infeasible.

Okay. That is fair.

I mean, it’s almost like self-driving cars, right?

I’m actually surprised to hear that, because the latency I’d imagine for reasoning seems small compared to how long a lot of these things take. A lot of them take hours to run, right? Or maybe your experiments are much faster or high-throughput or something. I’m just surprised to hear that, actually.

Ultimately, the experiments do take many hours or days, but there could be particular steps where latency would matter.

Yeah.

Yeah.

Yeah.

You might want finer control. Another thing is how expensive it is, because the same reason some of these models can be really slow at analyzing data also makes them very expensive. And finally, there’s human patience. If a human wants the analysis of a certain XRD pattern and they have to wait 2 hours before they get a good result, that’s very different, I think, from getting it in 2 minutes.

Oh, okay.

Yeah. And then the reasoning—you’re not doing a single call to that model, right? You’re doing potentially many, many tokens to get to these patterns.

You might run simulations in the loop.

Exactly.

The pre-search—

It’s not a single tool call or a single inference thread.

Sure.

So then the latency can blow up quite a bit.

I see what you mean.

Yeah.

How much of your latency comes from tool calls? And I assume tool calls are largely DFT or computational techniques—

Or DFTs, yeah.

—and therefore, yeah. How much of your latency is derived from those versus actual reasoning?

It’s really process-dependent.

Okay, yeah.

Yeah.

Yeah.

And then the other thing I would think about is building up from small things to bigger things, where I assume the general temptation or typical development is incremental: everything is human-operated, then you find ways to automate it, and then you build up from there.

Mm.

I worry that sometimes that’s the way people evolve things, but that’s a local-minimum optimum.

Short term is an optimization.

Mm.

Right. Exactly. When, actually, you should get a humanoid and just put them in there.

I think we’re of the opinion that solving humanoids would actually be slower for getting to some of our goals.

Just checking. Again, a lot of this is just that you do this every day. We see this, but I don’t know the reality of the situation. A lot of people are hyping humanoids.

Yeah. I mean, maybe 1 process is that, by having a mix of humans and automation, you can identify really quickly what some of the bottlenecks in the experimental process are, and you start alleviating those bottlenecks 1 at a time. For example, if there’s some really tricky dexterity task that humans are excellent at and you’d have to spend months automating the machine, maybe don’t spend a ton of time there. Maybe a more promising thing is something very routine that takes up a huge amount of scientists’ and technicians’ time and is easy to automate.

So I think there’s a pragmatism to it. Ultimately, what we want from the lab is a huge quantity of data, high-quality data, and diverse data. Those are our goals, and full autonomy is a non-goal. We use automation in the service of achieving those goals with the data.

Yeah.

Another aspect to automation is that robots will just make fewer mistakes, potentially, in the lab. It’s been really helpful having AI systems with a full view over all of our data, because sometimes, if there’s a permutation in the data, they can identify it.

For example, 1 of our steps at one point had a cyclic error because 1 of the machines was loaded incorrectly and the patterns were inconsistent. The AI was reading through these things and said, “Well, given what was run, this is not expected.” It was looking at this batch of data longitudinally, and then it realized, “If I do this cyclic permutation and reverse it, everything is consistent.”

We were able to go back to the physical infrastructure and understand that a mistake had been made in the loading. 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, so you can improve the operating process. But at that point, you’re like, “Okay, this is another great opportunity for automation. How do we make this so reliable, so durable, that we never have those types of mistakes again?”

Yeah. Reducing the noise floor of experimental data is very important. Most experimental data in the literature has such a high noise floor that it’s usually worse than DFT accuracy. When different people do experiments in different labs, different parts of the world, and at different times, it introduces so much fluctuation to the result.

We’re hoping that, by standardizing these workflows, 1 of the biggest benefits will be that the noise floor will be lowered.

You’ve been pretty public about how you use open-source models and fine-tune them for stuff like this. Is it better? I can imagine a situation where you have maybe the dumber models do 1 task, and then you optimize for that task, and then you have a generalist frontier model that supervises everything. Is that a good mental model to have? Are there more stages to this that I can think about?

We basically use a mix of open-source models and closed-source models. We’re huge beneficiaries of this from our simulation side, building up the ML stack and the simulation stack. We don’t need to push on that axis, but there are a lot of areas where the latency is too high or the cost is too high.

In some cases, by having access to this data, we can push beyond the frontier of what these systems can do, even under some of the highest reasoning efforts. It’s not necessarily just about cost or Pareto efficiency. 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. By having access to this data, you can be that much more compute-efficient compared to some of the frontier models, and that’s been a really instrumental thing in our program.

Yeah. I’m sold on data quality or data access as a trade-off for compute. I think there’s some exchange rate of dollars for compute to dollars for data. I feel like the pendulum might be swinging now toward data. Obviously, you would agree with that.

I mean, yeah, especially when talking about the noise floor of experimental data. We need to have high-quality data. If 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. You actually try to get null results and learn from that—

That’s right.

Yeah.

9. The Value Of Null Results

Yeah, you’ve talked about null results in several other venues. 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 in terms of success?

Mm. A null result could be that we intended to produce some structure, and then all of the evidence points to us not producing that structure.

Do you ever intend not to produce a structure?

Of course.

Yeah.

Yeah.

Okay. Cool.

Yeah. Absolutely.

You do the negative controls regularly.

Absolutely.

Okay.

There are impurity phases that will kill your property.

Yeah.

Some can be toxic.

Yeah, exactly. 100%.

Okay.

A negative result, in some cases, is actually that we intended to make the following thing, and then we were able to identify a new structure that hadn’t previously been identified. It’s negative in some respect: we intended to do something else, but something else emerged from the data.

Also, in general, when you’re training a machine-learning algorithm, especially at a very basic level, if it’s a classification algorithm, you can’t really train it if you don’t have negative samples and everything is positive.

Yeah.

And this is a particularly bad problem in materials science because people usually publish crystals they could synthesize, but they usually don’t publish when they fail to synthesize a crystal.

Sometimes they might. Also, we never really know if a crystal is not synthesizable, ever, right?

It could just be a skill issue.

Exactly.

Yeah.

It could be a skill issue—

Especially with the literature.

—or a synthesis-method or technology issue, 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 another valid thing, too. 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—process-engineering-type data.

Through iteration, how did you actually get to that correct result? Because so much of materials science has this ambiguity in how something was made. Even if there's a known material, it can be highly nontrivial to replicate that. Some things are in a high school textbook, but other things are really at the frontier, and we're building up that know-how as well, and then building up a system that, given the string of negative results, can figure out how to get to that positive case.

Going back 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 cold-starting this, it seems like the problem is actually the reverse: you have an abundance of negative results and not enough positive results.

And as you said, if you were doing per-experiment labels, it would be mostly negative. But if you were doing per-campaign labels, and assuming campaigns end when we succeed, then it could be a bit more balanced.

I see. So your reasoning traces could almost be over entire campaigns and not…

Absolutely.

Okay.

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. Tracking all this data—the conversations, the intuitions, what was executed in the lab, the computations run, or the code written—and stitching all this together is so valuable.

I think the overall goal is that, rather than training on the final output of science, you're training on the process of doing science.

This feels very much like what you would try to do with RSI, where you are having a model train to train better models. You're now having a model train to make better experiments.

Absolutely.

But the difference is, it's not going back into the core model, 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 answer.

That's the big unified wheel. Yeah.

But does that mean—is it plausible that in, you know, Fable 6 or something, or GPT-8, you could just have a model that has been tuned on these higher-order reasoning traces and could actually do this without any—

Mm.

—logic? Because it's sort of the same thinking process, right?

Yeah. I think there's decision-making under uncertainty, obviously, in machine learning. There's noise in running the AI loops. But we do think there's a different set of challenges when you're actually interfacing with the physical world.

I also think there's another piece, too, which is that getting the compression of everything into weights is still really valuable. If inference-time reasoning were sufficient, all of the frontier labs would've stopped training at GPT-4 and been like, “Okay, from now on, we're going to get really good at inference-time improvements.”

We think that, because of the differences between 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 we also feel like even if Fable 7 gets really good—even better than what it is today—it will still have to run experiments to get results. The reason is that machine learning is really good at what it's been trained on, but scientific discovery is almost by definition what you haven't been trained on.

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 not going to be—no one's going to zero-shot the room-temperature superconductor.

Think their way to—

That would be pretty cool. Yeah. As someone who has worked closely with wet labs before, certainly I am more skeptical about zero-shotting scientific results than some.

Yeah.

But it is something that I think some people might ask.

Yeah. I think it's really important to distinguish the results we see in math and theoretical physics from the physical world. These are 2 very different things.

No, it's 2 very different things.

Yes.

And not even theoretical physics. So far, it's been theoretical computer science, coding, and math.

Yes.

Right.

Maybe theoretical physics is next.

It'll come.

Yeah, that'll be next.

Yeah.

Oh, yeah.

Can I get a sort of mental model of the layers, the levels of abstraction that you can go? For example, in terms of automation—and I'm just on this theme again, where we talked about the campaign, talked about individual assays and running tests, and how humans are bad at it, so we should stop humans from doing it—are there others? One level higher than a campaign could be a physics theory that you're testing.

Mm.

Mm.

Or one level lower than a campaign is what? I just like to think about it from that point of view, and then think about it from, “Okay, well, this is an API call now, and you never have to touch this again.” This is still 90% human.

Mm-hmm.

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 ask a good question. I don't have a very systematic answer, but let's talk about some levels.

One level is the atomistic structure. This is an abstraction, right? There's no perfect atomistic structure in anything we do, but it's one approximation. Another one is the continuum model. This one is not atomistic anymore, but it's a continuous mesh representing the material.

In a different dimension, there's thermodynamics. Assuming you do this experiment forever, what would be the final state? Then there's another layer of abstraction, which is kinetics: acknowledging that we're not doing this experiment forever, so time matters. How quickly will a reaction happen—or not—even if it's lower energy?

Thermodynamics, kinetics, atomistic, continuum. What else? Are there other levels of abstraction?

Well, I think your point that there could be an overarching new theoretical advance that informs multiple campaigns—

Maybe the difference…

Kind of like, as an investor, I want to go—

Yeah.

“Here's the bottleneck, guys, and—

Mm.

—when we get this, we get everything else.” I don't know.

From the beginning, our hypothesis has been that one of the big bottlenecks is automated characterization, because it's not that hard to mix powders and try stuff. But if you can't characterize and analyze it, and then decide intelligently what the next step should be, you don't really benefit much from mixing powders randomly.

We focused a lot on automated characterization, closing the loop with simulations, because that's how you try a lot of things, make an informed decision, and then decide what to do the next day.

Maybe as a follow-up question: how do humans live in this loop? Where—at what points do you drop the human in? I mean, I'm guessing you probably do—

You could randomly—

Yeah.

—drop in at any part—

Of course. Yeah.

—and add value.

Yes.

But where do you decide to spend your time?

I mean, it seems like it might be all levels. You have people in the lab who are physically moving materials—

Yeah.

And then you also have scientists who are guiding decisions about what campaign to run, and then you have humans who are... Well, sorry, I think I'm just asking you the same question. Yeah, yeah.

No, yeah—100%. I think it's always shifting, too.

So I think even just using 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, as Ekin 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.

This has allowed the scientists, who were previously spending all of their time doing these refinements, to now elevate their work to something else. But I think it's really 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? 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. It's pretty good at structure prediction, for example, finding low-energy structures.

Okay, so it works. Yeah, okay.

And then process engineers in the semiconductor industry will do DoE, design of experiments, and they'll often use some kind of zeroth-order method, whether it's Bayesian optimization or evolutionary search. I don't think they use RL, but it's a similar idea.

Mm-hmm. You have to have some kind of algorithm to figure it out.

Yeah. And all zeroth-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. I'm curious, what is your vision for scaling the lab?

Mm.

There are different ways, and coming from the biology world, there are 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? Ultimately, every assay has its own runtime. I have a favorite blog post I would want to quote in the show notes. What do the runtimes look like for these labs?

Is your scaling for materials just fundamentally linear? You do more—you just need more resources to do more things, or can you parallelize things in clever ways and combine things? Are there tricks there?

Well, I think one way to think about it is that the RL environment is not simply: run the experiment, wait for that agent to fully roll out, do the calculations, go through all the instruments to the final outcome. We 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 that reinforcement-learning environments can be constructed based on, again, data for 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 get to some reward state. So that's a way where this finite basket of data, when viewed in different ways, can then be expanded for training purposes. It doesn't get at the actual experiment run, but it does expand things from an ML perspective.

Maybe I can give you some analogs for the biology examples you gave.

Yeah.

One thing people have tried is combinatorial sputtering. I don't know if you ever heard of this. You have these sputtering targets, and you create this compositional gradient. When you look at the final product, because there is a gradient from each precursor, it creates different crystals at different spatial locations. So in one go, you maybe try hundreds or thousands of crystals. That's similar to your example.

Oh, yeah. That makes sense.

And as far as I know, this hasn't worked very well so far, because it turns out there is diffusivity, so things move around. Another one that we thought about but haven't done yet, but a lot of people talk about, is that superconductivity measurements are a bottleneck because, unlike XRD, there's no high-throughput superconductivity measurement, and it can take 1 hour per measurement.

So people have thought about taking the different candidates, mixing them all into 1 sample and putting it in, and if it has superconductivity, you know that 1 of the sources did. Then you can do... It's a bit like how people are doing COVID testing.

Yeah. Mm-hmm.

So there are ideas like this; some of them work well, and some of them don't.

Yeah.

So you are about to announce a big fundraise. How are you thinking about scaling?

We've been doing the 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 the AI side and the computational side. But I think really importantly, too, it's about what new types of labs we're able to build.

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.

Yeah, the kind of lab we're trying to build, I think, hasn't been built before at this scale or with this approach. So we have learned a lot from our own build-outs so far, and those lessons guide us for the next version and the next version.

Then we can, as you said, increase the scale, the ambition, and the quality of instruments. Some instruments are very expensive. As we get a really good understanding of what gives us a lot of return, we can invest more in those.

And it means hardware engineering has become so core to it. For example, in the early days, just for speed, we'd buy some off-the-shelf instruments. As we pushed the scale of the lab—

You make your own.

—we had to make our own. So we realized, 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 the resting position of this robotic arm is actually above the plate of previously mixed things, so there's a contamination risk. In our hardware design, we're going to redesign it so that the resting position lies away from those things. These are the kinds of subtle details that allow us to push the noise floor down for our experimental campaign, so basically—

You shared your org chart. You have many teams. 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 to make sense of the jobs page: we hire for AI research and infrastructure, computational and experimental roles, hardware engineering roles, and then product roles.

We want to achieve synthesis superintelligence, right? And we feel like for that we need the right chemistry expertise, the right physics expertise, simulation theory, and expertise in thin films and powder. So, hardware engineering.

These seem like very diverse roles, which they are, but they're all actually coherent in what they're trying to do, which is achieve synthesis superintelligence. This also applies to the LLM researchers, the infrastructure, and even the product engineers, because the product engineers make sure all the research gets made into a product that the experimenters in the lab can use.

Yeah, 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, and to make them good users of the computational tools. That's the difficulty but also the opportunity of Periodic.

This group of people has just never been brought together before. It's irreducibly a multidisciplinary problem.

Yeah. One other guiding principle for us that is really important is that we want the people who are very good at what they do and experienced to do it hands-on.

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 and teaching that they stop having the time to do research in that area hands-on. But if you look back at Bell Labs, IBM, and institutions that made really good progress, it was really experienced people doing hands-on work.

Bardeen was in the lab every day, even though he was a theorist.

Alex Müller was in the lab doing experiments even though he was the lab lead. So we try to do that here too. We have some of the world's leaders in different fields, but they're doing hands-on work.

Do you want to brag and just call them out?

Oh, I would love to. When I was doing my PhD, my favorite computational materials scientist from my age group was Murat Aykol. He did his PhD with one of our advisors, Chris Wolverton, and he's our computational materials science expert. He's incredible. He's sometimes a really good experimentalist even though he does simulations; he understands systems really well.

Our lab lead is Joe Checkelsky, who is a professor at MIT but is on leave to work with us full-time. Joe is an incredible physicist who understands superconductivity really well, but he was also a professor in Japan for a bit, so he learned synthesis and chemistry really well, especially for a physicist. We have Daniel Chica, who was a graduate student in Mercouri Kanatzidis's group, which is maybe the best solid-state chemistry group in the world. Daniel was one of his synthesis experts, so he's like a magician with his fingers. There are so many examples like this.

No, I know. I think Dima Bahdanau has been leading a lot of our AI and LLM efforts. He was the inventor of neural network attention.

Oh, Bahdanau. The Bahdanau attention.

Yeah.

Exactly, yeah.

Yeah.

So, very hands-on and getting into the traces, the training distribution, and how to connect it to the lab.

The chemistry—

Chemistry.

—parts. I mean, he'll go deep.

He's very deep. He actually has solid-state synthesis books at his lab in Montreal. Reiichiro Nakano used to be the tech lead for some of the Operator work at OpenAI. He's in the lab, so when we were doing some lab tours, we initially thought, "Oh, this new scientist, this new technician really looks like Ray." No, it was literally Ray, actually in the lab with the scientists automating the machines. This is the DNA and the type of people we need to pull this off.

We also saw that you had 4 forward-deployed engineering roles. We happen to be spinning up a forward-deployed engineering podcast because there are so many engineers who would do that role. They may not necessarily know that there even is a role for them at Periodic. What is it, and how do they work with customers?

We've been building our own products, our own tools, and our own AI and computational—

For yourself. Yeah.

—for ourselves. We've been customer zero. Now what we're doing is taking these same tools to different industries, and we have a huge focus right now in the semiconductor industry.

However, these are very private companies, so we need to be able to operate within some of the most secure environments. Our Periodic forward-deployed engineers and researchers will be on-site, working and integrating our AI and computational systems to help our partners get to their goal states more quickly. Basically, the tools and know-how that we've built in doing our own materials discovery and materials engineering, we're now bringing to industry. We think this is the most effective way to get partners to their end states and goals, rather than just throwing some technology over the wall and telling them to figure it out.

These forward-deployed engineers will also do the inference locally, but then they can also train on the data. Once we take our system that understands these different areas and deploy it, we can make it expert on the customers' or partners' data, so they can own their own intelligence and have systems that understand this much more effectively than just hitting some API or using some untrained model.

This is a much more expansive forward-deployed engineering role than typical because it requires some very deep machine learning expertise. You have to be incredibly precise from a data-training perspective, an infrastructure perspective, and also a physics perspective. We're working with highly technical industries, so they have to understand different areas of chemistry, materials science, and some of the devices. Those are the roles we're building out right now.

Part of it is the willingness to spend extended periods on-site in Taiwan. It's like, okay, well, I know that kind of customer. It's interesting that you're monetizing—that that is the primary way that you're monetizing now—but it might change in the future. That's a model that I think people don't really get about an applied research lab. When you compare yourself to Bell Labs, that's what people are thinking, right?

I think a really great analogy would be software engineering.

Right.

In the early days of software engineering, we had GitHub Copilot, then early versions of ChatGPT. People were using these things as copilots to help them get to their solutions. As the automation improved, now we have things like Codex, and very few of our engineers are writing code the way they used to.

We think a very similar thing could play out for Periodic as well, where you can have systems to accelerate researchers, materials scientists, materials engineers, and process engineers. As the autonomy, intelligence, and capability grow, you can begin to price outcomes. Help me get to this kind of goal state. I think that's a really interesting area for us.

We kind of feel like our best contribution to solid-state physics and science could be if we made these tools and topics profitable, similar to how ChatGPT made CS and ML majors much more popular in colleges. We'd love to show that solid-state physics and materials science research can make a big impact commercially, and then that will attract more attention. Young people will want to study physics, which would be a dream.

Bell Labs did make a huge commercial impact, right? They failed to commercialize some of their incredible advances. They of course did some research that just couldn't be commercialized, like the cosmic microwave background, but they did benefit a lot from the vacuum tube connecting the East Coast to the West Coast by phone lines.

Oh.

Yeah.

I think it's also really interesting to think that technology and capital are incredibly intertwined. If you look at the progress on chatbots for many years, imagine recounting, okay, what was the progress on chatbots from 2010 to 2015? We'd be at a loss to say, over that 5-year period, how much did they improve? Whereas if you look at the period from 2021 to 2026, it's night and day.

What happened was that ChatGPT and these other systems were able to achieve product-market fit, and it changed the capital landscape entirely. This changes the landscape for hiring, compute, data, and so on.

The virtuous cycle.

Exactly. Technology is incredibly coupled to capital and those resources, and we want to achieve the same thing in the physical world.

Yeah.

Success breeds success.

Exactly.

One thing I've been seeing is this move from science being funded by government grants and so on to VCs and private funding. Do you plan on making anything you're doing open source, either releasing datasets or actual models? Is this something that might happen in the future? There are many models which may not be your state-of-the-art, but could still be useful releases.

Absolutely. We have a couple of things. We've been contributing to open source, including pymatgen, the Materials Project codebase, and Custodian, their DFT runner. TorchSim was actually created by one of the researchers here, Abhijit Gangan, and he still maintains it. Abhijit also maintains JAX-MD. So we do a lot of open-source contributions—

SGLang—

—including LLMs—

Megatron.

Yeah.

We've been very active contributors back into open source.

We also have an academic grant program where we give academic gift grants to academic groups at universities that we feel are really advancing this direction toward synthesis superintelligence. That's also been great to see. I think the first paper from this funding is about to come out, so it'll be exciting to get more papers like this. Yeah.

Let's steelman this: say that you've successfully achieved 100% lab automation, which does everything and characterizes everything correctly. You have all of the agents that can do everything. I'm still somewhat unclear about how you actually get to superconductivity. There are a lot of hard problems to solve. What is the path there when there is no sort of theory or model for most of the strongly correlated systems or high-temperature systems?

Yeah.

10. The Path To Superconductors

Our view is that it’s not hard to think of chemical spaces that would host superconductivity. What’s really hard is to synthesize them. That’s why we’re really emphasizing synthesis superintelligence.

Maybe one historical example is Alex Müller. 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. Then he tried cuprates—copper, oxygen, and some cations—and that worked, and then he got a Nobel Prize. But later on, it turned out nickelates were a good idea. Nickel and copper are next to each other on the periodic table.

Those are always 2D layers, right?

Yeah.

Yeah.

Exactly. There are only so many 3D transition metals.

Yeah.

Then Harold Hwang, one of our Stanford professors here, who’s incredible, realized he could make nickelates in thin-film form and show the superconductivity.

If we have a really good synthesis superintelligence in the lab and 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. There are related ideas here involving 3D transition metals and oxygen. Many people have been trying cobaltates, where instead of copper or nickel, you try cobalt.

I think that’s super exciting. 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. Thirty of them seemed to host interesting superconductivity, and MgB₂ was one of them.

But magnesium diboride had sat on people’s shelves as a precursor for decades before then, and BCS theory had been invented back in 1957. So the way people found these materials wasn’t predicted from theory. It wasn’t because they couldn’t synthesize them until then. It was just that they tried a bunch of things.

Now imagine if you had this perfect automation that you’re describing, which sounds like a dream, and we tried the 30,000 materials the Taketo group tried over his career in a month. We’d really increase 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. I imagine recruiting must be super easy for you, so I’m just a little bit jealous. But you’ve worked hard to get here.

Well, thank you so much. It was great chatting with you both.

Yeah, super fun.

Thank you.