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

“think that's sort of the the thesis and premise behind periodic that dojo i thought when we 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.”

Claim: Periodic's thesis is that advancing knowledge requires forming conjectures and testing them against reality; thinking alone is insufficient.

“No amount of rereading that textbook or paper and just thinking and like disappearing into a room is going to allow you to think through all possible experimental outcomes expected and unexpected.”

Claim: Rereading texts and thinking alone cannot anticipate every possible experimental outcome, including unexpected outcomes.

“and you really need this like iterative process and 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 to build up the the physical high throughput experiment and we think a different type of intelligence emerges from that”

Claim: Periodic believes that combining AI systems, physical-world simulations, and high-throughput physical experiments in an iterative process produces a different kind of intelligence.

“Yeah, I mean a team like this has never existed, I think. Uh we try to bring together cell state chemists, cell state physicists, experimentalists, theorists, hardware engineers, all the experts, computer scientists. Um 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.”

Claim: High-throughput experimentation and robotic arms have only been tried for about the past three years.

“The force field expertise, one of these force fields are also very recent. Um but yeah, we felt like having a lab is very important, having 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 worked together and achieved incredible things.”

Claim: The speaker considers having a lab and bringing people with diverse expertise together to work 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 worked together and achieved incredible things.”

Claim: Bell Labs is an example of theorists, experimentalists, and chemists collaborating to achieve significant results.

“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.”

Claim: Periodic derives its reinforcement-learning environments and data from its physical labs, which serve as its ultimate source of 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.”

Claim: Optimizing solely against answers already established in papers or textbooks is insufficient for work intended to go beyond existing knowledge.

“issues of say decision making under uncertainty so when you're doing optimization against math there's a high precision to it you're not really dealing with you know kind of variance 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 at this temperature but really the temperature was some delta away and an intelligence that can take in this noisy data and make intelligent decisions like a scientist would is a sort of a different set of reasoning strategies.”

Claim: In materials discovery, experimental data can be noisy or uncertain because furnace outputs may be mislabeled, labeling may be stochastic, machines may disagree, and telemetry may not reveal the actual temperature.

“maybe we'll have insufficient telemetry so we thought we ran at this temperature but really the temperature was some delta away and an intelligence that can take in this noisy data and make intelligent decisions like a scientist would is a sort of a different set of reasoning strategies.”

Claim: An AI system that makes scientist-like decisions from noisy experimental data requires a different set of reasoning strategies.

“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 infrared doesn't have mismatches between like training and inference.”

Claim: Periodic's approach shares methods with standard AI training, including mid-training, reinforcement learning, tool-using agents, variance reduction, and reducing training–inference mismatches.

“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? of a limited set of data.”

Claim: Periodic's scientific AI work must achieve accuracy under high uncertainty while using limited data efficiently.

“So we're scaling things up significantly, but still 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.”

Claim: Periodic is scaling up, but its physical experimentation cannot expand environments or rollouts as freely as a digital environment, making sample efficiency important.

“I think every scientific experiment has to do some dimensional deduction.”

Claim: Scientific experiments require reducing a system's dimensions to a manageable representation.

“And this is very different than 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, 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 figure it out.”

Claim: Unlike scientific experimentation, coding and mathematics can present the relevant reasoning context—such as axioms, corollaries, functions, and API calls—to a human or an LLM.

“In physics, as you know, we start with more atoms than we could ever store on a computer. So clearly we'll have to go from. the original number of dimensions and bits that represent a system to the number of bits we can fit in the computer.”

Claim: Because physical systems contain more atoms than computers can store, physics requires reducing their full dimensional and informational representation to one that fits computational limits.

“And, you know, this was, I guess, the original premise of thermodynamics. They were originally really confused about how steam engines worked.”

Claim: The speaker characterizes thermodynamics as originating from efforts to understand 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.”

Claim: A small set of five or six thermodynamic variables can explain steam-engine behavior quite well.

“we've decided that energy, the average value of energy and the fluctuations of energy is very important.”

Claim: In theoretical physics, both the average energy and energy fluctuations are considered important.

“So a lot of thermodynamics has been derived from this. And then we realized reaction barriers.”

Claim: Many results in thermodynamics have been derived from considering average energy and energy fluctuations.

“So a lot of thermodynamics has been derived from this. And then we realized reaction barriers. There's are also very important the kinetics is very important Um”

Claim: Reaction barriers and kinetics are important considerations in understanding material systems.

“so a human can try to reason through this extremely complex system by coming up with some, you know, reduced dimensionality descriptors um”

Claim: Humans can reason about extremely complex systems by developing descriptors that reduce their dimensionality.

“People also really like atomistic picture, like do you think about what the local neighbourhoods look like, whether they're from octahedra, tetrahedra, and that especially chemists really benefit from that. Um 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 etcetera.”

Claim: Experimental science involves collecting as much data as possible and recording observations in a lab notebook.

“People also really like atomistic picture, like do you think about what the local neighbourhoods look like, whether they're from octahedra, tetrahedra, and that especially chemists really benefit from that. Um 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 etcetera.”

Claim: Chemists especially benefit from atomistic descriptions of local neighborhoods, including octahedral and tetrahedral arrangements.

“But with the understanding of course that you'll miss a lot. Uh so like as Liam mentioned there are all these issues that come up. For example, your”

Claim: Experimental records inevitably omit some relevant observations or information.

“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 you use a furnace it gets worse and worse. Uh 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. Uh we we don't have this issue yet, but we will maybe someday. The optical instruments uh suffer a lot from vibrations.”

Claim: Vibrations can significantly affect optical instruments.

“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 you use a furnace it gets worse and worse. Uh 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. Uh we we don't have this issue yet, but we will maybe someday. The optical instruments uh suffer a lot from vibrations.”

Claim: A furnace degrades as it is used because evaporated material from samples can coat the heating element.

“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 you use a furnace it gets worse and worse. Uh 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. Uh we we don't have this issue yet, but we will maybe someday. The optical instruments uh suffer a lot from vibrations.”

Claim: Because furnace temperature is not perfectly uniform, sample placement within the furnace affects experimental results.

“like I remember when I was doing my P_H_D_ one lap had this big issue, because sometimes their results would be very different than other times, and they brought it down to at a certain time in the night upstairs somebody would be walking. And that vibration affected the laser set-up. Um so yes, science is very hard, but that's what makes it special, right? 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.”

Claim: In a lab where the speaker did a PhD, walking upstairs at a particular time of night caused vibrations that affected a laser setup and coincided with substantially different results across occasions.

“and they brought it down to at a certain time in the night upstairs somebody would be walking. And that vibration affected the laser set-up. Um so yes, science is very hard, but that's what makes it special, right? 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. science because it's easier for your limbs, but in real life most things that require intelligence are actually more like science. Like there's a lot of uncertainty, there's a lot of noise, um there's a lot of missing context, but you have to be the intelligent being and figure out what to do next.”

Claim: Many current AI improvements focus on mathematics, coding, and theoretical computer science because those domains are easier for AI systems.

“and they brought it down to at a certain time in the night upstairs somebody would be walking. And that vibration affected the laser set-up. Um so yes, science is very hard, but that's what makes it special, right? 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.”

Claim: Science is difficult, and its difficulty is part of what makes it special.

“science because it's easier for your limbs, but in real life most things that require intelligence are actually more like science. Like there's a lot of uncertainty, there's a lot of noise, um there's a lot of missing context, but you have to be the intelligent being and figure out what to do next.”

Claim: Most real-world tasks requiring intelligence resemble science because they involve uncertainty, noise, and missing context, requiring an agent to decide 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.”

Claim: Reasoning strategies that are optimal for mathematics or theoretical computer science may not be optimal for scientific work.

“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 next 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”

Claim: Materials discovery begins by selecting a material whose atoms are expected to be stable together and whose properties are expected to meet the target.

“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 next 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”

Claim: Material synthesis is non-trivial because determining the processing conditions needed to produce a target material is difficult.

“come off labeled so you actually have to characterize that and figure out what you made so in that case”

Claim: A synthesized material must be characterized because the product does not identify itself or come labeled with what it is.

“Rather than thinking about the oral environment of we're just going to kick off an experiment, wait a couple days and try to do like an update on like, did you find a room temperature superconduct or not? That's completely infeasible.”

Claim: It is completely infeasible to use a workflow that runs an experiment, waits a couple of days, and then checks whether a room-temperature superconductor was found.

“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 There's just some physics limits of the experiments for the actual synthesis or furnace time.”

Claim: A slow experiment-and-update workflow is infeasible because there are too few agents to reduce variance sufficiently, instrument transitions create waiting time, and synthesis or furnace time has physical limits.

“So it's just too slow, too noisy.”

Claim: The experimental process is too slow and too noisy for the described workflow.

“So the way we think about the AI program is basically constructing agents around this set of data that we've produced.”

Claim: The AI program constructs agents around the experimental data produced by the team.

“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?”

Claim: For material characterization, the team seeks reinforcement-learning environments that reward correctly identifying the phases present from raw experimental data.

“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 fingerprint of the crystal structure.”

Claim: X-ray characterization uses wavelengths roughly comparable to interatomic spacing to produce diffraction patterns that fingerprint a material's crystal structure.

“Doing this identification of what actually is present is highly non-trivial.”

Claim: Determining which phases or structures are actually present in a material is highly non-trivial.

“And so some of our early AI work has been on basically producing systems to do this.”

Claim: The team's early AI work has included systems for identifying the phases or structures present in materials.

“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.”

Claim: AI systems for material identification can speed the workflow, while high experiment volume can make interpretation of experimental outputs a bottleneck.

“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 have any like real chemical plausibility so that that's like a you know a little sustaining 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 and”

Claim: The material-characterization reinforcement-learning environment rewards correct phase identification and penalizes spurious or chemically implausible phases.

“another another piece would be as we build up a large basket of experimental data you can think of another thing of basically like time stamping 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 okay given that body of experimental evidence and given you know the choices what what was like the the next choice that the scientists made or what was an outcome of that experiment and this is really interesting too because this allows us to do you”

Claim: A time-stamped collection of experimental data can support reinforcement-learning environments that use the evidence available at a given point in time to model scientists’ next choices or experimental outcomes.

“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 work. so it doesn't actually have to do the hard physics reasoning. It doesn't have to actually...”

Claim: When a pretrained model already knows an answer used in a reinforcement-learning task, it can appear to do the work without performing the required physics reasoning, calculations, or simulations.

“after actually do the 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 the whole process that gives us the ability to construct new types of environments that are just not plausible elsewhere”

Claim: Reasoning strategies reinforced after a model reaches the right answer without carrying out the actual calculation or simulation will not generalize to novel systems and are therefore not useful.

“after actually do the 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 the whole process that gives us the ability to construct new types of environments that are just not plausible elsewhere”

Claim: Aggregating experimental data while preserving its process lineage enables the construction of new kinds of environments that would not be plausible to create elsewhere.

“yeah so there are a couple of things going on here so i think first of all we're definitely not done with any of the laws of physics like even the”

Claim: Humanity has not yet completed its understanding of any of the laws of physics.

“So there's a very common 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 a soul done.”

Claim: A commonly told account says that in the late 1920s, while working on quantum mechanics, Dirac wrote a textbook that presented the subject as essentially complete.

“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.”

Claim: The speaker was unsure whether the attribution was accurate, but reported that Dirac was said to have remarked, “The rest is chemistry.”

“he could maybe solve like a chain of one D_ hydrogen atoms. But then he cannot currently solve say nitrogen interacting with oxygen. But that's okay, that's just chemistry.”

Claim: Dirac could perhaps solve a chain of one-dimensional hydrogen atoms, but not a system involving nitrogen interacting with oxygen, which he regarded as chemistry.

“Yeah, but it turns out I think what we learned in the last hundred years is first of all that's not true.”

Claim: Scientific progress over the last hundred years has shown that the claim that everything beyond basic physics is simply chemistry is not true.

“um it's not like a simple extension of what it was to do just 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 temperatures for conductivity. But there are other other things.”

Claim: Understanding complex systems is not merely a straightforward extension of solving the hydrogen atom; many phenomena beyond that involve substantial physics, not just chemistry.

“um it's not like a simple extension of what it was to do just 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 temperatures for conductivity. But there are other”

Claim: Scientists have not yet figured out high-temperature conductivity.

“When people say high temperatures for conductivity, what they mean there is unconventional superconductivity. So there is the conventional superconductivity that's just mainly driven by electron phonon coupling. And in that case you can kind of see the isotope effect.”

Claim: Conventional superconductivity is mainly driven by electron-phonon coupling and exhibits an observable isotope effect.

“hope 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 temperature superconductors like cuprates, like the ones that are above seventy seven Kelvin like ninety three Kelvin turns out don't obey that physics. Or we don't know what physics they obey.”

Claim: For conventional superconductors, changing the mass of one element in an otherwise similar system changes the superconducting temperature at the expected rate.

“hope 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 temperature superconductors like cuprates, like the ones that are above seventy seven Kelvin like ninety three Kelvin turns out don't obey that physics. Or we don't know what physics they obey.”

Claim: Higher-temperature cuprate superconductors, including examples above 77 K such as those around 93 K, do not follow the conventional superconductivity account, or their governing physics is not yet known.

“Um they're just incredible superconductors.”

Claim: The speaker characterized high-temperature superconductors as remarkable superconductors.

“Um but it's not just that. So that definitely is one of these like very uh popular topics we don't know the theory yet for yet, but”

Claim: Unconventional superconductivity is a prominent research topic whose underlying theory is not yet known.

“But there's so many other things we don't understand. Like for example, even strong correlation in simple quantum mechanical systems we cannot simulate yet.”

Claim: Even strong correlation in simple quantum-mechanical systems is not yet simulatable by scientists.

“Density functional theory is a very powerful tool. It's incredibly accurate on some things. But as long as as soon as there's like some strong electron correlation,”

Claim: Density functional theory is a powerful method that is highly accurate for some problems.

“Density functional theory is a very powerful tool. It's incredibly accurate on some things. But as long as as soon as there's like some strong electron correlation, it actually fails to capture some of the effects.”

Claim: Density functional theory fails to capture some effects in systems with strong electron correlation.

“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 material solid state physics the better devices we can make for them.”

Claim: The speaker is excited about applying AI to materials and solid-state physics because substantial theoretical, computational, and experimental work remains.

“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 material solid state physics the better devices we can make for them.”

Claim: Improvements in understanding materials and solid-state physics should improve human life by enabling better devices.

“which is basically you may understand all the basic laws of something, but when you... 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,”

Claim: Collective systems can exhibit qualitative behavior that differs from what one would expect from the simple laws governing their individual components.

“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”

Claim: The collective behavior of many interacting components is difficult to model because of its complexity.

“it's like emergent and universal which is crazy and there are also things with deep learning models right like we see power laws everywhere in deep learning we don't understand but it's like it's very reminiscent reminiscent of physics where there is emergence and universality.”

Claim: Deep-learning models exhibit power laws that are not understood and resemble the emergence and universality observed in physics.

“it's like emergent and universal which is crazy and there are also things with deep learning models right like we see power laws everywhere in deep learning we don't understand but it's like it's very reminiscent”

Claim: Collective behavior in complex systems can display emergence and universality.

“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 the performance of the material basically.”

Claim: SPEAKER_02 thinks it may be impossible to reverse-engineer every process perfectly, but says the goal is to reverse-engineer processes well enough to improve material performance.

“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.”

Claim: SPEAKER_01 believes substantial progress can be made by dividing agentic workflows across different areas.

“I think there's a lot of like empirical work we can make on synthesis prediction as well.”

Claim: SPEAKER_01 believes substantial empirical work remains to be done on synthesis prediction.

“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”

Claim: SPEAKER_01 says Periodic Labs' mission includes synthesis superintelligence and that the proposed materials-process concept aligns with it.

“Phase transitions are fascinating. I mean, I think maybe the one that people can most easily relate to in their life is ice melting probably or water boiling.”

Claim: SPEAKER_02 identifies ice melting and water boiling as familiar examples of phase transitions.

“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.”

Claim: SPEAKER_02 describes ice as remaining solid while heated until a transition point at which it turns into liquid rather than continuing to rise in temperature.

“So it suddenly goes from this solid phase to liquid phase.”

Claim: SPEAKER_02 characterizes melting as a sudden transition from solid to liquid.

“Another one that's very relevant to us of course is the icing model.”

Claim: SPEAKER_02 says the Ising model is relevant to the lab's work.

“And I think computer science and mathematicians also study this maybe under different names.”

Claim: SPEAKER_02 says computer scientists and mathematicians study the Ising model, perhaps using different names for it.

“But basically you can have spins up and down and they have different interaction terms if they're both up versus one is up one is down.”

Claim: SPEAKER_02 says that in the Ising model, spin interactions depend on whether paired spins are both up or have opposite orientations.

“And at high enough temperature they're usually just randomly up or down, entropy wins.”

Claim: SPEAKER_02 says that at high enough temperatures, Ising-model spins are usually randomly oriented because entropy dominates.

“spins, you lower the temperature, it's still looking the same, but then at some point suddenly they perfectly set aligning, yeah, the phase transition.”

Claim: SPEAKER_02 says that as temperature decreases, spins remain apparently unchanged until a point at which they suddenly align.

“Phase transition is really nice because it makes physicists life easier for studying certain things.”

Claim: SPEAKER_02 says phase transitions make some phenomena easier for physicists to study.

“They tend to have certain spatial correlations that really help us as well.”

Claim: SPEAKER_02 says phase transitions tend to exhibit spatial correlations that are useful to the lab.

“so in our lab phase transitions come about usually because we mix the precursors that are different crystals basically,”

Claim: SPEAKER_02 says phase transitions in the lab usually arise from mixing precursors that are different crystals.

“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 materials 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”

Claim: SPEAKER_02 says substantial changes in atomic geometry usually produce substantial changes in the XRD pattern.

“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 materials 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 mixed phase.”

Claim: SPEAKER_02 says initial attempts in practical materials-discovery campaigns usually fail to produce a clear binary outcome and instead yield mixed phases.

“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 materials 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”

Claim: SPEAKER_02 says the lab encourages mixed crystal precursors to react, often by raising temperature, leading atoms to form a new crystal.

“It will usually have some of the precursors,”

Claim: SPEAKER_02 says mixed-phase outcomes usually retain some of the precursor material.

“it might have some amorphous phase and then it will have like a bunch of phases that you maybe didn't try to make.”

Claim: SPEAKER_02 says mixed-phase outcomes may include an amorphous phase and phases that researchers did not set out to produce.

“So that becomes a challenge. And that's why we found it very difficult to bring together simulations and AI to do the characterization.”

Claim: SPEAKER_02 says the team found combining simulations and AI for materials characterization very difficult.

“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”

Claim: SPEAKER_02 says simulations can assess a phase that differs slightly from a prediction by using force-field calculations to check its stability.

“Based on the previous experiment this one. This is probably not the phase we want to make so let's change the conditions.”

Claim: SPEAKER_02 says AI can use prior experimental results to infer that a phase is probably not the target and suggest changing experimental conditions.

“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 put, use these tools to say well what configuration atoms would actually explain these phases. But this becomes incredibly important for direct.”

Claim: SPEAKER_01 says some observed phases have not previously been recorded in papers or databases.

“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 put, use these tools to say well what configuration atoms would actually explain these phases. But this becomes incredibly important for direct.”

Claim: For some material phases that are absent from papers and databases, AI must use tools to infer which atomic configurations could explain the phases.

“It doesn't exist in any papers or database and so the AI has to actually go and sort of put, use these tools to say well what configuration atoms would actually explain these phases. But this becomes incredibly important for direct.”

Claim: SPEAKER_01 says that for phases absent from existing papers and databases, AI must use tools to identify atomic configurations that could explain them.

“It doesn't exist in any papers or database and so the AI has to actually go and sort of put, use these tools to say well what configuration atoms would actually explain these phases. But this becomes incredibly important for direct.”

Claim: SPEAKER_01 says phase interpretation is important for directing the scientific process.

“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 a one percent and we're like actually we want to increase this phase parity we want to have a very good measurement on this experimental data to say you know what is present what's not”

Claim: When targeting a material phase, the scientific process should iteratively improve measurements so researchers can determine what is and is not present in the experimental data.

“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 a one percent and we're like actually we want to increase this phase parity we want to have a very good measurement on this experimental data to say you know what is present what's not”

Claim: SPEAKER_01 says that when a target phase is present at one percent, researchers may seek to increase its proportion and measure the experimental data well enough to identify which phases are present.

“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 a one percent and we're like actually we want to increase this phase parity we want to have a very good measurement on this experimental data to say you know what is present what's not”

Claim: SPEAKER_01 says that when researchers have a target phase, they need to iteratively improve measurements of it.

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

Claim: Many cases of identifying material phases remain ambiguous.

“I mean, I think there, you know, it's a challenging task because it's not just a. fully deterministic and there are like well there there are like ambiguities and other things”

Claim: Identifying material phases is challenging because the process is not fully deterministic and contains ambiguities.

“Two different phases can't actually be consistent with the same pattern.”

Claim: A single measurement pattern can be consistent with two different material phases.

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

Claim: A system identifying material phases needs chemical intuition and prior knowledge to disambiguate interpretations, such as recognizing when a proposed phase is highly unlikely under the synthesis conditions.

“Yeah, I mean I think thermodynamics is the biggest prior,”

Claim: Thermodynamics is the most important prior for interpreting material experiments.

“And then physics is a big prior.”

Claim: Physics provides an important prior for interpreting material experiments.

“Another one that we really benefit from here is multi-modality or materials characterization.”

Claim: Using multiple forms of materials characterization benefits the team's analysis.

“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.”

Claim: Material phases that look similar in X-ray diffraction patterns can be further distinguished by measuring electrical and 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. to look different than some of these dimensions.”

Claim: Morphology measurements, such as electron microscopy, can distinguish phases that appear similar in X-ray diffraction patterns.

“And that kind of multimodality really helps.”

Claim: Combining multiple measurement modalities helps analyze materials.

“And again, AI is really helpful there because like humans have limited context and computation power as well.”

Claim: AI is useful for analyzing multimodal data because humans have limited context and computational capacity.

“So if you give humans like 10 different modalities at the same time and say, analyze these all consistently, it's a bit difficult.”

Claim: Humans find it difficult to analyze ten different data modalities consistently at once.

“But it's actually not super intelligent for an AI to do it.”

Claim: Analyzing multiple data modalities is not especially demanding for an AI.

“Yeah, it's like the superintelligence is it can just do so much calculations, they can look at everything.”

Claim: A superintelligent system can perform many calculations and examine all available information.

“about like this is 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 kind of getting to this like underlying better model rather than just over indexing on one measurement from one instrument”

Claim: Better decision-making under uncertainty comes from integrating signals across instruments, experiments, and replicates to build an underlying model instead of relying too heavily on a single instrument measurement.

“Yeah, so we don't throw away 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. So unfortunately, we're still in a regime where any data we collect is very precious and we don't delete any of it.”

Claim: The team retains all collected data because the data is precious and is not deleted.

“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.”

Claim: An experiment contains more data than humans can store or access, because humans cannot observe every atom or trace every atom's position.

“So you know in 1800s people realized that entropy of a system has to increase or like for a closed system or stay the same but it can't go down as the second law of thermodynamics.”

Claim: The second law of thermodynamics says that entropy in a closed system cannot decrease; it must increase or stay constant.

“And there was a thought experiment that was asking what would happen if there was a very tiny all-knowing demon 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,”

Claim: Maxwell’s demon is a thought experiment in which an all-knowing demon observes atoms and selectively lets low-energy atoms into another room.

“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 of thermodynamics. And 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 nineteen hundreds um more recently to us uh Landauer came up with the argument that you'd have to spend energy to delete information.”

Claim: In the Maxwell’s demon thought experiment, sorting atoms into another room would reduce the system’s entropy, apparently violating the second law of thermodynamics.

“which would break second law of thermodynamics. And 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 nineteen hundreds um more recently to us uh Landauer came up with the argument that you'd have to spend energy to delete information.”

Claim: The speaker says that for a long time the Maxwell’s demon problem lacked a very satisfying explanation, until Landauer argued that deleting information requires energy.

“they have access to so much information by doing this thing he's doing that he'd have to delete some and when 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 demon I think your question is like very visionary towards like a future where we store so much data we have to delete some”

Claim: The speaker considers the interviewer’s question visionary because it anticipates a future in which data storage becomes large enough that some data must be deleted.

“they have access to so much information by doing this thing he's doing that he'd have to delete some and when 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 demon I think your question is like very visionary towards like a future where we store so much data we have to delete some”

Claim: In the Maxwell’s demon thought experiment, the demon would need to delete some of the information it gathers; the energy required for deletion would cause entropy to increase.

“we add a huge amount of telemetry. And that basically reduces the number of like hidden variables, if you will.”

Claim: The speaker’s organization adds extensive telemetry to reduce the number of hidden variables.

“Yeah we could I mean we have plans to you know open more labs also”

Claim: The speaker’s organization plans to open additional labs.

“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 it helps us because we can hire from there, our researchers can collaborate, like as you said, that geography can matter. But then different cities can have some differences of expertise, just in physics, chemistry.”

Claim: The speaker says Stanford’s physics and engineering departments provide expertise that supports the organization’s hiring and research collaborations.

“it helps us because we can hire from there, our researchers can collaborate, like as you said, that geography can matter. But then different cities can have some differences of expertise, just in physics, chemistry.”

Claim: Expertise in fields such as physics and chemistry varies between cities.

“Yeah, I mean the future of periodic is is not a single lab.”

Claim: Periodic’s future will not consist of only one lab.

“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. There's many different constraints permitting.”

Claim: The speaker believes a network of labs is necessary, and that each lab’s optimal location depends on expertise, material constraints, and other factors.

“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 experiment and more accurate in experiment.”

Claim: Depending on the task, simulation or physical experimentation can be the easier and more accurate approach.

“I mean, in general, of course, experiments are more accurate, but certain measurements are very hard to make with experiment.”

Claim: The speaker says experiments are generally more accurate, although some measurements are very difficult to obtain experimentally.

“So you make a bad version of it so you get inaccurate data.”

Claim: Difficult experimental measurements can lead researchers to use an inadequate experimental method and obtain inaccurate data.

“use density functional theory to estimate formation enthalpy of materials and”

Claim: The speaker’s lab uses density functional theory to estimate materials’ formation enthalpies.

“density functional theory is probably the most commonly used simulation method for materials it kind of comes from Kohn-Hohenberg theorem and Kohn got a Nobel Prize for this when he realized that sure quantum mechanics is very expensive because”

Claim: The speaker says density functional theory is probably the most commonly used simulation method for materials.

“density functional theory is probably the most commonly used simulation method for materials it kind of comes from Kohn-Hohenberg 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.”

Claim: The speaker says quantum-mechanical calculations are expensive because wave functions involve an exponentially large Hilbert space, and the problem becomes especially difficult as system size increases.

“Cone and Hohenberg realized that turns out you don't need to deal with an exponential space, all you have to do is think about the charge density, at least for the ground state of quantum mechanical systems. So it's actually the theorem is pretty simple, so even I think an undergraduate in physics can understand,”

Claim: The speaker says the Kohn–Hohenberg theorem allows the ground state of a quantum-mechanical system to be treated through its charge density rather than an exponentially large space.

“but the idea is all the properties of a quantum mechanical system ground state is a functional of the charge density.”

Claim: According to the speaker, every property of a quantum-mechanical system’s ground state is a functional of its charge density.

“Yeah, which is amazing because charge density is just this three-dimensional object, whereas the Hilbert space and wave functions are exponential. Um so that was a really interesting observation and then Cohen published another paper, I think with his post-doc uh Sham, Cohen-Sham uh wave functions kind of had a practical way to solving s Some approximation of quantum mechanical properties of materials.”

Claim: The speaker says the Cohen–Sham work offered a practical method for approximating quantum-mechanical properties of materials.

“Yeah, which is amazing because charge density is just this three-dimensional object, whereas the Hilbert space and wave functions are exponential. Um so that was a really interesting observation and then Cohen published another paper, I think with his post-doc uh Sham, Cohen-Sham uh wave functions kind of had a practical way to solving s”

Claim: The speaker contrasts three-dimensional charge density with the exponentially large Hilbert spaces and wave functions used in quantum mechanics.

“And today we use it very much both in our lab and externally.”

Claim: Density functional theory is used both in the speaker’s lab and outside it.

“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?”

Claim: Formation enthalpy is one measurement used to assess the energy of a material.

“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.”

Claim: The speaker says a material with higher energy than competing materials is unlikely to be synthesizable.

“So stable just means like it's on the comeback soil of other materials and their energies.”

Claim: The speaker says a material is stable when it lies on the convex hull of other materials’ energies.

“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.”

Claim: DFT can reduce calculations that are exponentially hard in electron number or system size to approximate calculations that usually scale cubically, at the cost of some error.

“You can now compute things which are just uncomputable, but at some cost.”

Claim: DFT makes some otherwise uncomputable quantities computable, but doing so has a cost.

“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.”

Claim: Heather Kulik has said that materials lack an AlphaFold equivalent because the ground-truth material structures are unavailable.

“Like we don't really know what material crystal structures look like.”

Claim: The speaker says that material crystal structures are not really known.

“Yeah, I think so. I mean maybe two things I would add is, DFT by definition doesn't have to be an approximation. That the Kohn-Hohenberg theorem shows it can be exact,”

Claim: The Kohn-Hohenberg theorem shows that DFT can be exact in principle, so DFT is not necessarily an approximation by definition.

“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 model we don't know what the kinetic energy functional is.”

Claim: Researchers do not yet have access to the exchange-correlation functional, and the kinetic-energy functional for charge-density-based models is unknown.

“And then, even if we had perfect DFT, I still think we'd need a lab. Because even if you're perfect DFT, we cannot fit ten to the twenty three atoms in the computer.”

Claim: The speaker believes laboratories would still be necessary even with perfect DFT, because a computer cannot represent 10^23 atoms.

“Yeah. No, okay, okay. Yeah, but with exponentially growing compute, you can argue that in in cube scaling size, eventually you could in principle calculate things. You just wait till Moore's But as long as you know, yeah, yeah. But as long as in cube is a good scaling, you can”

Claim: The speaker suggests that exponentially growing compute could eventually make some calculations possible in principle under cubic scaling, given enough time for Moore's law to advance.

“I personally don't know of a better method to predict. predict the stability of a new material it's definitely not perfect but it's definitely better than other methods i can think of and 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 these in a loop like we feel like simulations will never be enough by themselves but in the loop of simulations ai and experiments you”

Claim: The speaker knows of no better method than DFT for predicting new-material stability, although DFT is imperfect.

“predict the stability of a new material it's definitely not perfect but it's definitely better than other methods i can think of and 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 these in a loop like we feel like simulations will never be enough by themselves but in the loop of simulations ai and experiments you”

Claim: The speaker says simulations and experiments complement one another and that progress should use an iterative loop of simulations, AI, and experiments, because simulations alone will never suffice.

“And I think we can make progress much faster than before.”

Claim: The speaker expects an iterative process combining simulations, AI, and experiments to accelerate progress substantially compared with the past.

“One thing that simulations let you do is scale up much more quickly. I mean, because now like compute is much easier to bring online than bringing in bringing on new labs.”

Claim: Simulation work can scale up faster than laboratory work because computational capacity is easier to bring online than new laboratories.

“even if we can do, as you said, 100 million trials with EFT, we could never actually try their microstructure.”

Claim: Running 100 million EFT trials would still not allow researchers to experimentally test the material's microstructure.

“And microstructure is this idea that like EFT usually stimulates the perfect crystal, but crystals usually actually aren't perfect.”

Claim: EFT typically simulates ideal perfect crystals, whereas actual crystals are usually imperfect.

“And their microstructure, which is the structure at like a medium order, can really affect the properties.”

Claim: The medium-order microstructure of a material can substantially affect its properties.

“Other issues, of course, like some properties like superconducting temperature cannot be easily simulated by D_F_T_ Again, like even if you do hundred million trials with D_F_T_ it won't be enough. So we have to rely on heuristics, experiments, etcetera.”

Claim: DFT cannot easily simulate some properties, including superconducting temperature, and even 100 million DFT trials would be insufficient, requiring heuristics and experiments.

“Uh like, for example, for a long time materials project which was the best opens for open source DFT database use experimental calibration for DFT. So they would actually use experimental data to calibrate DFT.”

Claim: The Materials Project calibrated DFT using experimental data for its open-source materials database.

“We also do this internally here, like whenever there's a certain chemical system we're interested in, you get experiments, you get simulations and you can calibrate them.”

Claim: The speaker's team uses experimental results to calibrate simulations for chemical systems it is studying.

“But do you not feel like organic chemistry and biology almost stem from a lower VC dimension, like lower chromogero complexity model, like you can almost represent as a 1D sequence, and it can almost be compiled. So like there's a sense in which it's a uh simpler in the common world complexity sense.”

Claim: The speaker suggests that organic chemistry and biology may have lower effective complexity than materials, with aspects that can be represented nearly as one-dimensional sequences and compiled.

“I think most people would argue it's more 2D, but yeah.”

Claim: The speaker says most people would characterize DNA as more two-dimensional than one-dimensional.

“Virtual is crazy. It's clearly not coming from such a low cholemic complexity model because that 3D inorganic crystals can host metals, insulators, superconductors, diamonds.”

Claim: Three-dimensional inorganic crystals can host a wide range of material behaviors and phases, including metallic, insulating, superconducting, and diamond forms.

“Like people tried but we could never find a 1D representation.”

Claim: The speaker says that attempts to find a one-dimensional representation for three-dimensional inorganic crystals have not succeeded.

“Like it's very hard to find a smile string for 3D inorganic crystals.”

Claim: The speaker says that finding a “smile string” representation for three-dimensional inorganic crystals is very difficult.

“And the way the atoms interact with each other is quite strange like covalent bonds, ionic bonds, metallic.”

Claim: Inorganic crystals can involve covalent, ionic, and metallic bonding between atoms.

“there are all these dendrites that form like basically the lithium will form these structures that go into salt electrolyte 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 technological progress whereas we can't even like make two interfaces work with each other because lithium will break it yeah but then I guess inorganic stuff can be very resistant to high temperature”

Claim: The speaker contrasts biology’s complex nanotechnological capabilities with the difficulty of making two material interfaces work together in batteries, where lithium can break the interface.

“there are all these dendrites that form like basically the lithium will form these structures that go into salt electrolyte 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 technological progress whereas we can't even like make two interfaces work with each other because lithium will break it yeah but then I guess inorganic stuff can be very resistant to high temperature”

Claim: Lithium dendrites can grow into and crack a solid electrolyte.

“there are all these dendrites that form like basically the lithium will form these structures that go into salt electrolyte 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 technological progress whereas we can't even like make two interfaces work with each other because lithium will break it yeah but then I guess inorganic stuff can be very resistant to high temperature”

Claim: Some inorganic materials can withstand high temperatures.

“You can make space shuttles, you can make silicon, compute Moore's law. So yeah, it's a trade-off.”

Claim: The speaker describes inorganic materials as enabling applications such as space shuttles and silicon computing, while emphasizing that their advantages involve trade-offs.

“yeah yeah yeah yeah but maybe even just like more specifically for um 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 what the crystal structure is right no so and that's something that what yeah”

Claim: An X-ray diffraction spectrum from a perfect crystal does not necessarily identify a unique crystal structure.

“it's just average it's just average behavior but like for example disproportionation is a very complicated thing right where it's possible that you're the perfect crystal and on average it looks like it's 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”

Claim: X-ray diffraction can reflect average structure: in disproportionation, atoms may be displaced slightly to either side even though the average pattern places them in the middle.

“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 and 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.”

Claim: The speaker says one could make a scientific argument that materials science is the hardest scientific field.

“Well, we didn't pick it necessarily for hardness. I mean, because actually like a lot of areas where it is easier. I mean, we have these like great simulators for large classes of materials,”

Claim: The speaker says their choice of materials science was not necessarily motivated by its difficulty; many areas are easier, and good simulators exist for large classes of materials.

“it's incredibly difficult to model a cell or an organ or a full organism.”

Claim: Modeling a biological cell, organ, or whole organism is incredibly difficult.

“Yeah, I mean you still get like different complexity at the different length scales for materials as well So yeah microstructure and beyond.”

Claim: Materials exhibit different forms of complexity across length scales, including at the microstructure scale and larger scales.

“especially if you can lower the dependence on rare earth or like transition metals that are hard to source like cobalt, reducing cobalt in batteries.”

Claim: The speaker considers it valuable to reduce reliance on rare earths and difficult-to-source transition metals such as cobalt, including in batteries.

“One of the things I think is very fundamentally important is if we can get close to land our limit for compute energy efficiency.”

Claim: The speaker considers approaching the Landauer limit for computational energy efficiency fundamentally important.

“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 gotten exponentially more efficient. So 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 was.”

Claim: The speaker recalls that energy consumption per flop or computation had improved exponentially in a Moore’s-law-like pattern, but had not checked whether this remained true recently; they said it was true ten years earlier.

“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.”

Claim: Computations generate heat that must be dissipated.

“So like if we can get close to the Landauer limit for how efficiently we do computation, that's amazing for humanity, right?”

Claim: The speaker says that approaching the Landauer limit for computational energy efficiency would be highly beneficial for humanity.

“Because that means now we're doing computation is probably one of the most fundamental things we do as humanity, but as efficiently as possible from energy perspective, which is a real currency in the universe.”

Claim: The speaker says computation is probably among humanity’s most fundamental activities and argues that maximizing its energy efficiency matters because energy is a fundamental resource.

“uh so 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 uh but i don't know if that works i i don't know uh quantum computing”

Claim: SPEAKER_02 was unsure whether reversible computing can avoid energy expenditure by preserving information, and said they were not an expert on the question.

“I assume has to obey these laws somehow because they are so like they have to obey thermodynamics”

Claim: SPEAKER_02 assumes quantum computing must obey the laws of thermodynamics.

“One of my most memorable conversations with Elon Gill was like he actually was a very skeptical person about quantum computing.”

Claim: SPEAKER_00 recalled that Elon Gill was very skeptical about quantum computing.

“It's like even if you had it today, there's no applications.”

Claim: SPEAKER_00 said that quantum computing would have no applications even if it were available today.

“like it breaks security,”

Claim: SPEAKER_00 said quantum computing breaks security.

“it. agree with that like like for them people talk about how if you're a quantum computer today you could could simulating so much better and I asked 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 like 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”

Claim: SPEAKER_02 questioned the practical value of claims that quantum computers could greatly improve simulation, asking what would be simulated and observing that perfect DFT predictions would still leave DFT's existing issues.

“could simulating so much better and I asked 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 like 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”

Claim: SPEAKER_02 thinks excitement about quantum logic leads people to gloss over what quantum computing would actually do once built.

“made maybe that's fine because maybe once it's made it will do amazing things we can't even imagine”

Claim: SPEAKER_02 allowed that quantum computing might do amazing, currently unimaginable things once it is built.

“In the early days, as we were scaling things up, we realized that we had huge bottlenecks imposed just by operating machinery.”

Claim: SPEAKER_01 said that operating machinery became a major bottleneck as the lab scaled up.

“we would have technicians and scientists looking for particular morphology and trying to see, what actually were we making, see if this is consistent with our intentions.”

Claim: SPEAKER_01 said technicians and scientists inspected machinery output for particular morphology and checked whether the results matched their intentions.

“And we really quickly, as we scale up the lab, became into these bottlenecks where It just wasn't keeping up.”

Claim: SPEAKER_01 said that manual inspection could not keep up with the lab's growth, creating bottlenecks.

“So we started doing some programmatic approaches to capturing the data on SEM.”

Claim: SPEAKER_01 said the team began using programmatic methods to capture scanning electron microscope data.

“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 account.”

Claim: SPEAKER_01 said their simple field-of-view capture program produced data that was not useful enough for the lab's needs.

“And so at that point, it was then really pertinent to build AI systems directly onto the machines to start controlling these things.”

Claim: SPEAKER_01 said the lab needed to build AI systems directly into its machines to control them.

“And now they have the full context as to what we're trying to achieve.”

Claim: SPEAKER_01 said the lab's machine-integrated AI systems have full context about the lab's experimental goals.

“So 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.”

Claim: SPEAKER_01 said the machine-integrated AI system captures data from inside the machine in the context of the experiment's intent, synthesis target, and other evidence.

“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.”

Claim: SPEAKER_01 said that more intelligent data capture during experiments produces richer data for future AI systems and computational predictions.

“you know, everything on the lab has to be incredibly intelligent to just make the data as useful as possible.”

Claim: SPEAKER_01 said lab equipment should be highly intelligent to maximize the usefulness of collected data.

“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. process then that that's infeasible um”

Claim: SPEAKER_01 said instrument control must account for physical-process timescales, and that API latency or excessive reasoning and tool calls can make control infeasible when they do not fit those timescales.

“ultimately the experiments do take uh you know mini hours days but um There could be particular steps where latency would matter.”

Claim: SPEAKER_01 said experiments can take minutes, hours, or days, while some individual steps may still require low latency.

“how expensive it is because like the 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 yeah”

Claim: SPEAKER_02 said slow data-analysis models are expensive and that users may value receiving an XRD-pattern analysis in two minutes rather than waiting two hours.

“many, many tokens to kind of get to these patterns. not a single tool call. It's not a single inference thread. So then the latency can blow up quite a bit.”

Claim: SPEAKER_01 said experimental analysis may require many tokens and multiple tool calls or inference threads, causing latency to grow substantially.

“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. I worry that sometimes that that is the way that people evolve things, but that's a local minima, optimum.”

Claim: SPEAKER_00 worries that incrementally automating human-operated processes can trap development in a local optimum.

“right exactly when like actually you should get a humanoid and just put them in there”

Claim: SPEAKER_00 suggested that deploying a humanoid robot directly might be preferable to incrementally upgrading existing processes.

“we we i think we're in the opinion that solving humanoids would actually be slower to kind of getting to some of our goals”

Claim: The speaker thinks solving humanoid robotics would take longer than pursuing some of the lab’s goals through other means.

“we we i think we're in the opinion that solving humanoids would actually be slower to kind of getting to some of our goals”

Claim: SPEAKER_01 thinks solving humanoid robotics would be slower than other approaches for achieving some of the lab's goals.

“yeah i mean i think like maybe one one process is um by having like a mix of uh like 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.”

Claim: Combining human work with automation can quickly reveal bottlenecks in an experimental process, which can then be addressed sequentially.

“And 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.”

Claim: The speaker suggests avoiding substantial automation effort on dexterity tasks that humans perform well when automating them would take months.

“And maybe like a more promising thing is something very routine that takes up a huge amount of scientists and technicians.”

Claim: The speaker considers routine, time-consuming scientific or technical tasks that are easy to automate more promising automation targets.

“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 it's sort of in the service that we use automation in the service of achieving the goals on the data”

Claim: The lab prioritizes producing large quantities of high-quality, diverse data; full autonomy is not a goal, and automation is used to advance those data objectives.

“but also i mean i think another aspect to automation is uh robots will just be uh 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 uh 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 so 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 understand”

Claim: AI systems that can inspect the lab’s data comprehensively have helped identify data inconsistencies, such as a cyclic error caused by incorrect machine loading.

“but also i mean i think another aspect to automation is uh robots will just be uh 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 uh then it can identify it so for example”

Claim: The speaker thinks robots may make fewer mistakes in laboratory work.

“and 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 you know these are the types of things that we're building in and 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”

Claim: The speaker considers data-quality management foundational for applying AI in the physical world.

“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.”

Claim: Differences among experimenters, laboratories, locations, and times can introduce substantial variation into experimental results.

“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.”

Claim: The speaker says experimental data reported in the literature often has a noise floor high enough to make it less accurate than DFT.

“So we are hoping that by standardising these workflows one of the biggest benefits will be that the noise floor will be lowered.”

Claim: The speaker hopes that standardizing experimental workflows will reduce the noise floor in experimental data.

“I mean, we basically use a mix of open source models and closed source models.”

Claim: The speaker’s team uses both open-source and closed-source models.

“We're huge beneficiaries of this from our simulation side, building up the ML stack, the simulation stack.”

Claim: The speaker’s team benefits substantially from its simulation work and from building its machine-learning and simulation stacks.

“We don't need to push on that axis.”

Claim: The speaker says the team does not need to prioritize further progress on the simulation axis.

“But there's a lot of areas where the latency is too high, the cost is too high.”

Claim: The speaker says latency and cost are too high in many areas where their systems are used.

“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 so by having access to this data you can be that much more compute efficient compared to you know some of the frontier models”

Claim: The speaker says access to their data can enable systems to exceed the capabilities they reach even with the highest reasoning effort.

“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 so by having access to this data you can be that much more compute efficient compared to you know some of the frontier models”

Claim: The speaker says access to otherwise unavailable data can improve compute efficiency relative to frontier models.

“And that's been able, been like a really instrumental thing in our program.”

Claim: The speaker says access to data has been instrumental to the team’s program.

“um 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”

Claim: In materials research, a null result can mean that the intended structure was not produced, according to the available evidence.

“exactly 100 percent um and a negative result you know in some cases is actually 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.”

Claim: In some cases, an experiment intended to produce one thing yields evidence of a previously unidentified structure instead.

“Yeah. And also in general, like 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”

Claim: The speaker says a classification algorithm cannot be properly trained using only positive samples and no negative samples.

“this is a particularly bad problem in material science because people usually publish crystals they could synthesize but they usually don't publish if they fail to synthesize a crystal sometimes they might and also like we never know really if a crystal is not synthesizable ever right”

Claim: The speaker says it is generally not known whether a crystal can never be synthesized.

“this is a particularly bad problem in material science because people usually publish crystals they could synthesize but they usually don't publish if they fail to synthesize a crystal sometimes they might and also like we never know really if a crystal is not synthesizable ever right”

Claim: The speaker notes that failed crystal-synthesis attempts are sometimes published, although they are usually not.

“this is a particularly bad problem in material science because people usually publish crystals they could synthesize but they usually don't publish if they fail to synthesize a crystal sometimes they might and also like we never know really if a crystal is not synthesizable ever right”

Claim: The speaker says materials-science publications generally report successfully synthesized crystals but omit failed attempts.

“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.”

Claim: The team finds its experimental negative results useful when contextualized by the attempted procedures, and can use them to train a classification algorithm.

“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.”

Claim: A sequence of failed experiments followed by an iterative path to a successful result can provide useful process-engineering data.

“So it's like through iteration how did you actually get to that correct result because so much of material science has this ambiguity and actually”

Claim: Because materials science involves ambiguity about how materials are made, understanding the iterative path to a correct result is important.

“like 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 you know 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 I'm building up a system that you know given the string of negative results how do you actually get to that positive case”

Claim: The speaker is developing know-how and a system for finding a successful outcome from a sequence of negative experimental results.

“like 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 you know 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 I'm building up a system that you know given the string of negative results how do you actually get to that positive case”

Claim: Replicating a known material can be highly non-trivial, with some materials being textbook-level and others at the research frontier.

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

Claim: Individual experiment labels would be mostly negative, whereas campaign-level labels—assuming campaigns end when they succeed—could be more balanced.

“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.”

Claim: The speaker says this scientific-process data is otherwise unavailable and that the team devotes substantial effort to putting the full process lineage into its model.

“And so like tracking all this data, like the conversations, the intuitions, what was executed in the lab or the... before the computations run or with the code written stitching all this together is 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”

Claim: The team tracks scientific-process information including conversations, intuitions, laboratory actions, computations, and code.

“before the computations run or with the code written stitching all this together is 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”

Claim: The speaker's goal is to train models on the process of conducting science, rather than only on scientific results.

“the physics models you're developing are improving chips which then improve ai so this is the the wider”

Claim: Developing physics models can improve chips, and improved chips can in turn improve AI.

“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.”

Claim: The speaker believes that interacting with the physical world poses challenges distinct from the noise encountered in machine-learning AI loops.

“If inference time reasoning was sufficient, all of the frontier labs would have stopped training.”

Claim: The speaker argues that if inference-time reasoning alone were sufficient, all frontier AI labs would have stopped training their models.

“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.”

Claim: The speaker expects that encoding physical-science knowledge in the team's model weights, given differences between physical sciences and machine learning, will produce different types of systems and 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 it 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 and 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.”

Claim: The speaker predicts that even a substantially improved Fable 7 will still need experiments to produce scientific results, since scientific discovery concerns what the model has not been trained on.

“it will still have to run experiments to get results and the reason for it 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 and 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.”

Claim: The team is building laboratories for open and closed models to conduct experiments, reflecting the speaker's view that major discoveries require trying things.

“there's not going to be like no one's going to like zero shot the room temp superconducting.”

Claim: The speaker asserts that room-temperature superconductivity will not be achieved through zero-shot prediction alone.

“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,”

Claim: Based on experience working closely with wet labs, the speaker is more skeptical than some people about obtaining scientific results zero-shot.

“but it is something that I think some people might have.”

Claim: The speaker thinks some people may believe scientific results can be produced zero-shot.

“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 like yeah these are two different things”

Claim: The speaker considers results in mathematics and theoretical physics importantly different from results in the physical world.

“not even theoretical physics like so far it's been theoretical computer science building and math. maybe theoretical physics is next.”

Claim: The speaker says the work so far has focused on theoretical computer science and mathematics rather than theoretical physics, which may be a future area.

“it'll come yeah that's gonna be the next thing”

Claim: The speaker expects theoretical physics to be the next area of work.

“So one level is the atomistic structure. This is an abstraction, right? Like there is no perfect atomistic structure, anything we do, but it's one approximation.”

Claim: The speaker describes atomistic structure as an approximation and says their work has no perfect atomistic structure.

“Another one is the continuum model. So this one is like not atomistic anymore, but it's a the continuous mesh representing the material.”

Claim: A continuum model represents a material with a continuous mesh rather than an atomistic structure.

“I mean 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 to get 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. much from mixing powders randomly so we focused a lot on automated 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”

Claim: The speaker's team has hypothesized that automated characterization is a major bottleneck because powder mixing alone is of limited value without analyzing results and intelligently selecting subsequent experiments.

“much from mixing powders randomly so we focused a lot on automated 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”

Claim: The speaker's team focused on automated characterization and simulation-based closed-loop experimentation to test many options, make informed decisions, and select subsequent actions.

“So I mean, I think even just using like the guidance of campaigns in early days that was fully human driven. Now increasingly it's a mix of AI-driven and human-driven.”

Claim: Campaign guidance was fully human-driven in the early days and has increasingly become a combination of AI-driven and human-driven decision-making.

“And then Ders 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.”

Claim: Characterization was initially human-driven and a major bottleneck that overwhelmed scientists, but the work has since shifted significantly toward AI-driven processes.

“And this has allowed the scientists who are previously spending all of their time doing these refinements to now elevate their work to something else.”

Claim: AI-driven characterization has freed scientists from spending all their time on refinements, allowing them to take on higher-level work.

“But I mean, I think we're, it's really sort of a function of time. There's, it's always changing at these levels.”

Claim: The balance between human and AI involvement in the work changes over time.

“So especially in simulations, people have been using evolutionary search, For a long time, um, like it's pretty good at structure prediction, for example, finding low energy structures.”

Claim: Evolutionary search has long been used in simulations and performs well at structure prediction, including finding low-energy structures.

“And then uh the processor engineers in semiconductor industry will do DOE, design of experiments, and they'll often use some kind of zero order uh like whether it's Bayesian optimization or evolutionary search or I I don't think they use R_L_ but similar idea, yeah.”

Claim: Semiconductor processor engineers use design of experiments and often apply zero-order optimization methods such as Bayesian optimization or evolutionary search; the speaker thinks they probably do not use reinforcement learning, although the approaches are similar in idea.

“all zero order optimization behaves the same anyway, right? So.”

Claim: The speaker suggests that all zero-order optimization methods behave similarly.

“The RLL 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.”

Claim: The speaker describes the reinforcement-learning environment as nontrivial because an experiment must proceed through the agent's full rollout, calculations, instruments, and final outcome.

“We basically start producing data across all the different instruments and all these different campaigns.”

Claim: The team produces data across its different instruments and experimental 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 gates for, you know, the experimental campaigns or computational campaigns.”

Claim: Reinforcement-learning environments based on experimental or computational campaigns can be used to expand data available for machine learning.

“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 so that's a way where this finite basket of data when viewed in different ways can then be expanded for training purposes”

Claim: Reinforcement-learning environments built from one or more instruments can target reward states and generate varied training experiences from a finite dataset.

“So it doesn't get at the actual experiment run, but from a Ml perspective.”

Claim: The described data-expansion approach does not cover the actual experiment run, but it addresses the problem from a machine-learning perspective.

“So, you know, one thing people have tried is combinatorial sputtering, and if you've ever heard of this, 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 may You maybe try hundreds thousands of crystals”

Claim: Combinatorial sputtering creates compositional gradients from sputtering targets, yielding different crystals at different locations and potentially enabling hundreds or thousands of crystals to be tested in one run.

“as far as I know this isn't worked very well so far because turns out like there is diffusivity so things move around”

Claim: The speaker believes combinatorial sputtering has not worked very well so far because diffusion causes materials to move around.

“Another one that we thought about but we haven't done yet, but a lot of people talk about this is super conductivity measurements are a bottleneck because unlike XRD There's no high throughput conductivity measurement and can take like one hour per measurement”

Claim: The team has considered but not implemented an approach to a conductivity-measurement bottleneck: unlike XRD, conductivity lacks high-throughput measurement, and an individual measurement can take about an hour.

“So people thought about taking the different candidates, mixing them all into one sample and putting it in. And if it has superconductivity, you know that one of the sources did.”

Claim: A proposed pooled-testing method mixes candidate materials into one sample; a superconducting result would indicate that at least one source material was superconducting.

“there are ideas like this exist. Some of them work well, some of them don't. Yeah.”

Claim: Some of the ideas being discussed work well, while others do not.

“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.”

Claim: Periodic Labs believes its resources will allow it to significantly expand its labs and increase compute for both AI and computational work.

“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.”

Claim: Periodic Labs says it has been validating its process in its initial Menlo Park labs and plans to expand to new lab types while repeating the process.

“like the kind of lab we're trying to build I think haven't been built before at this scale or at this approach.”

Claim: Periodic Labs says its lab model has not previously been built at this scale or using 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.”

Claim: Periodic Labs uses lessons from its lab build-out to guide later versions and increase scale, ambition, and instrument quality.

“Some instruments are very expensive.”

Claim: Some laboratory 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.”

Claim: Periodic Labs can invest more in instruments after identifying which ones provide the greatest return.

“And it makes hardware engineering has become so core to it.”

Claim: Hardware engineering has become central to Periodic Labs' work.

“So for example, in the early days, just for speed, we buy some off the shelf instruments.”

Claim: Periodic Labs initially bought off-the-shelf instruments to move quickly.

“And as we kind of push the scale of the lab, we have to make our own.”

Claim: Periodic Labs says increasing lab scale requires it to build its own instruments.

“okay, for this instrument, it's pretty fast, these components are actually pretty quick, but this weighing machine actually becomes the bottleneck.”

Claim: In one instrument, the weighing machine became the bottleneck despite the other components being relatively fast.

“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 like these are the kind of subtle details that allow us to kind of push the noise floor down for our experimental campaign so basically take these learnings and scale it up”

Claim: Periodic Labs identified a contamination risk from a robotic arm resting above previously mixed materials and plans to redesign the hardware so the arm rests away from those materials.

“So we hire for AI research and infrastructure, computational, experimental roles, and hardware engineering roles, and then product roles.”

Claim: Periodic Labs hires across AI research and infrastructure, computational science, experimental work, hardware engineering, and product roles.

“We want to achieve synthesis superintelligence, right? And we”

Claim: Periodic Labs' stated goal is to achieve synthesis superintelligence.

“We feel like for that we need the right chemistry expertise, the right physics expertise,”

Claim: Periodic Labs believes synthesis superintelligence requires the right chemistry and 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 sensor 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. made into a product that the experimentals in the lab can use.”

Claim: Periodic Labs views its diverse hardware, AI research, infrastructure, and product roles as coherent because they contribute to building a sensor for intelligence.

“even the product engineers because the product engineers kind of make sure all the research gets made into. made into a product that the experimentals in the lab can use.”

Claim: At Periodic Labs, product engineers turn research into products usable by experimental staff in the lab.

“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.”

Claim: Periodic Labs says an end-to-end physical-world research loop requires physical agency, lab-building, effective experimental campaigns, AI systems designed for the work, and effective use of computational tools.

“That's like the difficulty, but also the opportunity of periodic.”

Claim: Periodic Labs describes its work as both difficult and an opportunity.

“It's like this group of people has just never been brought together before.”

Claim: Periodic Labs says the multidisciplinary group needed for its work has not previously been assembled.

“Irreducibly a multidisciplinary problem.”

Claim: Periodic Labs considers its research problem inherently multidisciplinary.

“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.”

Claim: Periodic Labs' guiding principle is for highly skilled, experienced people to perform hands-on work.

“You 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.”

Claim: The speaker says academic researchers who are strong in their fields can lose time for hands-on research because of grant-writing and teaching responsibilities.

“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 Bardin was in the lab every day, even though he's a theorist.”

Claim: The speaker attributes strong progress at institutions such as Bell Labs and IBM to experienced people doing hands-on work.

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

Claim: Periodic Labs says some of its field-leading experts perform hands-on work.

“Oh, I would love to, yeah. So um you know i i when I was doing my PhD, my favourite computational material scientist who's kind of from my age group is Muratan Icoc. He did his PhD with one of our advisors, Chris Wolverton and he uh he is is kind of like our computational material science expert. He's incredible.”

Claim: Muratan Icok serves as Periodic Labs' computational materials science expert.

“He's sometimes, you know, a really good experimentalist even though he does simulations. Like he understands systems really well.”

Claim: Muratan Icok sometimes demonstrates strong experimentalist ability despite doing simulations, and he understands systems well.

“Our lab lead is Joe Ciekalski who is a professor at MIT but he's on leave to work with us full-time. And Joe is this incredible physicist, understands superconductivity really well.”

Claim: Joe Ciekalski has strong expertise in superconductivity.

“Our lab lead is Joe Ciekalski who is a professor at MIT but he's on leave to work with us full-time. And Joe is this incredible physicist,”

Claim: Joe Ciekalski leads Periodic Labs' lab and is an MIT professor on leave to work full-time with the company.

“But he also was a professor in Japan for a bit. So learned synthesis and chemistry. really well especially for a physicist.”

Claim: Joe Ciekalski previously taught in Japan and learned synthesis and chemistry.

“We have Daniel Chica who was a graduate student in Mercury Kanazidis' group which is maybe the best solid state chemistry group in the world and Daniel was one of his like synthesis experts. So he's like a magician with his fingers. Um I mean there's so many examples like this.”

Claim: Daniel Chica studied in Mercury Kanazidis' group as a graduate student and was one of the group's synthesis experts.

“No, I know. I mean, I think like Dima Bahandar, so he's been leading a lot of our AI and LLM efforts.”

Claim: Dima Bahandar leads many of Periodic Labs' AI and LLM efforts.

“He was the inventor of neural network attention.”

Claim: Dima Bahandar is credited by the speaker as the inventor of neural network attention.

“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 chemistry”

Claim: Dima Bahandar works hands-on with model traces and training distributions and connects them to laboratory chemistry.

“was very deep um he actually has like solid state synthesis books like at his lab in montreal ray nacono so he used to tech lead some of the operator work at open ai”

Claim: Ray Nacono previously tech-led some 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 the DNA and the type of people we need to kind of pull this off eventually”

Claim: Periodic Labs considers hands-on people who work with scientists and automate laboratory machines essential to accomplishing its goals.

“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 the DNA and the type of people we need to kind of pull this off eventually”

Claim: Ray Nacono worked in the lab alongside scientists to automate machines.

“So we've been building our own products, our own tools, our own AI and computational for ourselves. So we've been customer zero.”

Claim: Periodic Labs has developed products, tools, AI, and computational systems for its own use and has acted as its own first customer.

“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. industry.”

Claim: Periodic Labs is bringing its internally developed tools to multiple industries, with a major current focus on semiconductors.

“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.”

Claim: Periodic Labs says that serving its private-company partners requires operating in highly secure environments.

“And so our periodic for deployed engineers and researchers will actually be on site working and actually integrating our AI systems or computational systems to help our partners get to their goal states more quickly.”

Claim: Periodic Labs plans to place forward-deployed engineers and researchers at partner sites to integrate AI and computational systems that help partners reach their goals faster.

“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 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 then these forward deployed engineers will also do the inference locally but then also can train on the data so again”

Claim: Periodic Labs is transferring to industry the tools and expertise it developed in materials discovery and materials engineering.

“in doing our own materials discovery materials engineering we're now bringing to industry and we think this is the 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 then these forward deployed engineers will also do the inference locally but then also can train on the data so again”

Claim: Periodic Labs believes that working closely with partners to apply its tools and expertise is more effective than simply handing them technology without implementation support.

“in doing our own materials discovery materials engineering we're now bringing to industry and we think this is the 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 then these forward deployed engineers will also do the inference locally but then also can train on the data so again”

Claim: Periodic Labs’ forward-deployed engineers will perform local inference and can train AI systems using partner data.

“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 you know than just like hitting some API or some untrained model and so this is a much expanded for deployment engineering role than typical because this really requires”

Claim: Periodic Labs says it can specialize its systems on partner data, enabling partners to own tailored intelligence instead of relying on a generic API or untrained model.

“There's some very deep machine learning expertise.”

Claim: Periodic Labs’ expanded forward-deployment engineering role requires deep machine-learning expertise.

“You have to be incredibly precise from like a data training perspective, infrastructure perspective, but also physics perspective.”

Claim: Periodic Labs’ forward-deployment engineering role requires precision across data training, infrastructure, and physics.

“So we're working with, you know, highly technical industries. They have to understand, you know, different areas of like chemistry, material science, some of the devices.”

Claim: Periodic Labs’ work with highly technical industries requires its forward-deployed staff to understand chemistry, materials science, and relevant devices.

“And so those are the roles we're building out right now.”

Claim: Periodic Labs is currently expanding its forward-deployment engineering roles.

“I think like a really great analog would be software engineering.”

Claim: Periodic Labs’ speaker considers the evolution of software engineering a useful analogy for the company’s approach to materials research and engineering.

“so in the early days of software engineering like we have like you know guess GitHub Copilot, then we had early versions of ChatGPT, people were using these things as co-pilots to help them get to their solutions.”

Claim: Early AI coding tools, including GitHub Copilot and early ChatGPT, were used as copilots to help software engineers reach solutions.

“And as the automation improves, now we have things like codecs and very few of our engineers are writing code the way they used to.”

Claim: The speaker says that improved automation has meant very few of their engineers still write code as they did previously.

“And 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,”

Claim: Periodic Labs believes AI systems could similarly accelerate the work of researchers, material scientists, materials engineers, and process engineers.

“but as the autonomy intelligence. and capability grows, you can begin to price outcomes.”

Claim: The speaker expects that increasing autonomy, intelligence, and capability could enable outcome-based pricing.

“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.”

Claim: Periodic Labs’ speaker believes its most valuable contribution to solid-state physics and science may be making tools and research in those fields profitable.

“Similar to how ChatGPT made CS and LLM majors way more popular in colleges.”

Claim: The speaker says ChatGPT substantially increased the popularity of computer science and LLM majors 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 but but they did benefit a lot from like the vacuum tube”

Claim: Periodic Labs aims to show that solid-state physics and materials-science research can produce substantial commercial impact, which it hopes will draw attention and encourage young people to study physics.

“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 but but they did benefit a lot from like the vacuum tube”

Claim: The speaker says Bell Labs had substantial commercial impact, including from the vacuum tube, while some of its advances were not commercialized and some research, such as cosmic-background research, was not commercializable.

“Connecting East Coast to West Coast by phone lines.”

Claim: The speaker cites connecting the U.S. East Coast and West Coast by telephone lines as an example of Bell Labs’ commercial impact.

“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,”

Claim: The speaker believes technological progress and capital are strongly interdependent.

“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?”

Claim: The speaker says chatbot progress from 2010 to 2015 was difficult to characterize in terms of how much the systems improved.

“Whereas if you look at the period from 2021 to 2026. You know, it's night and day.”

Claim: The speaker characterizes chatbot progress from 2021 to 2026 as dramatically greater than the progress from 2010 to 2015.

“And what happened was ChatGPT and these other systems were able to achieve a product market fit, and it changed the capital landscape entirely.”

Claim: The speaker says ChatGPT and other systems reached product-market fit and thereby transformed the capital landscape.

“This changes the landscape for hiring, compute, data, etc.”

Claim: The speaker says the transformed capital landscape changes access to or conditions for hiring, compute, and data.

“So technology, it's incredibly coupled to the capital and those resources,”

Claim: The speaker believes technology is strongly linked to capital and the resources available to develop it.

“and we want to achieve the same thing in the physical world.”

Claim: Periodic Labs aims to reproduce in the physical world the mutually reinforcing relationship between technology, capital, and resources that the speaker described.

“I mean, so we have a couple of things. So we have been contributing to open source to kind of like PyMedGen, 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, JAX-MD, Abhijit maintains it. So we do a lot of open source contribution.”

Claim: Periodic Labs says it has contributed to open-source projects including PyMedGen, the Materials Project codebase, a DFT runner, TorchSim, and JAX-MD.

“TorchSim was actually created by one of the researchers here, Abhijit Gangan, and he maintains it still,”

Claim: Periodic Labs researcher Abhijit Gangan created TorchSim and continues to maintain it.

“Abhijit maintains it. So we do a lot of open source contribution.”

Claim: Periodic Labs describes itself as an active contributor to open-source projects.

“SGX, including LLVM, Megatron. Yeah, yeah. So we've been very active contributors back into open source.”

Claim: Periodic Labs says it has actively contributed to open-source projects including LLVM and Megatron.

“Yeah, yeah. So we've been very active contributors back into open source.”

Claim: The company has actively contributed to open-source projects.

“And we also have an academic grant programme where we give academic gift grants to academic groups in universities that we feel like are really advancing this direction towards synthetic superintelligence.”

Claim: The company provides academic gift grants to university groups that it considers to be advancing synthetic superintelligence.

“I think the first paper from this funding is about to come out. So it will be exciting to just get more papers.”

Claim: The first paper produced through the academic grant funding is expected to be published soon.

“Yeah, so I mean 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.”

Claim: The speaker considers it relatively easy to identify chemical spaces that could host superconductivity, but difficult to synthesize materials from those spaces.

“That's why we're really like emphasizing synthesis superintelligence. So maybe one example to give historically is Alex Mueller when he thought that transition metal oxides might be a good good path to superconductivity he was actually trying nickelates, the nickel oxygen and some uh cadions and that didn't work and then he tried cuprates, just copper copper oxygen and some cadions and that worked. And then he got a Nobel prize. But then we know that later on turns out nickelates was a good idea. I mean nickel and copper next each other on”

Claim: Alex Mueller explored transition-metal oxides as a possible route to superconductivity; his nickelate attempts did not work, while cuprates did, and he received a Nobel Prize.

“good path to superconductivity he was actually trying nickelates, the nickel oxygen and some uh cadions and that didn't work and then he tried cuprates, just copper copper oxygen and some cadions and that worked. And then he got a Nobel prize. But then we know that later on turns out nickelates was a good idea. I mean nickel and copper next each other on”

Claim: Nickelate materials later proved to be a promising direction for superconductivity research.

“he realized he can make Nicolet's thin film formation with superconductivity.”

Claim: Harald Wang found a way to form nickelate thin films with superconductivity.

“you know, if we have a really good synthesis for intelligence 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,”

Claim: If the company develops effective lab-based synthesis intelligence and scales it up, the speaker expects that it and its LLM will continue to generate ideas for research directions.

“Like many people have been trying like cobaltates, which is like instead of copper or nickel, you try cobalt.”

Claim: Researchers have explored cobaltates by using cobalt in place of copper or nickel.

“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.”

Claim: Magnesium diboride is, according to the speaker, the highest-temperature conventional superconductor at ambient pressure.

“And they discovered it by just trying a bunch of materials.”

Claim: A Japanese research group discovered magnesium diboride through broad experimental trials of many materials.

“I think they tried 30,000 different things. 30 of them seemed to host interesting superconductivity and MgB2 was one of them. But magnesium diboride set on people's shelves as a precursor for superconductivity.”

Claim: The Japanese group tried approximately 30,000 materials, of which about 30 appeared to exhibit interesting superconductivity; magnesium diboride was among them.

“So four decades before then and BCS theory had been invented back in 1957.”

Claim: BCS theory was developed in 1957, roughly four decades before the magnesium diboride discovery described by the speaker.

“So the way people found these materials wasn't for directly from theory. It wasn't because they couldn't synthesize until then. It was just like they tried a bunch of things.”

Claim: The speaker attributes the discovery of these superconducting materials to broad experimental trial rather than direct theoretical prediction or an inability to synthesize them earlier.

“We just really increased the surface area for luck.”

Claim: The speaker expects automation to greatly expand the number of experimental opportunities for serendipitous discoveries.