A big problem with using artificial intelligence to discover new materials? It struggles to predict beyond its training data. That means AI might be better at optimizing known materials than discovering entirely new ones — like a room temperature superconductor or carbon-capture sorbents.
But since we last covered the topic in September 2024, a few things have changed. OpenAI released its powerful O1 reasoning model. Large language models have also gotten better at math, physics, and coding. And lab automation — robots mixing liquids and powders, running characterization tests — has improved, allowing for a higher volume of experiments.
So, can these improvements overcome AI’s training data problem?
In this episode, Shayle talks to Ekin Dogus Cubuk, cofounder of Periodic Labs, which raised $300 million seed round in September. Last year, Dogus took a more cautious view on using AI for materials discovery. Now though, he’s convinced there’s a clearer path forward for physical science research and development, especially materials discovery. Shayle and Dogus cover topics like:
- Creating experimental and synthetic data to overcome AI’s limitations of predicting beyond its training set
- Why we should focus on breakthrough discoveries over easier, incremental wins
- The different roles humans and AI play in the discovery process
- Period’s focus on automated experimental labs using AI-generated hypotheses
Resources
- Catalyst: Can AI revolutionize materials discovery?
- Latitude Media: This ‘superintelligence platform’ just raised $200m in seed funding
- Latitude Media: Can AI get us closer to fusion?
- The New York Times: Top A.I. Researchers Leave OpenAI, Google and Meta for New Start-Up
Credits: Hosted by Shayle Kann. Produced and edited by Daniel Woldorff. Original music and engineering by Sean Marquand. Stephen Lacey is our executive editor.
Catalyst is brought to you by EnergyHub. EnergyHub helps utilities build next-generation virtual power plants that unlock reliable flexibility at every level of the grid. See how EnergyHub helps unlock the power of flexibility at scale, and deliver more value through cross-DER dispatch with their leading Edge DERMS platform, by visiting energyhub.com.
Catalyst is brought to you by Bloom Energy. AI data centers can’t wait years for grid power—and with Bloom Energy’s fuel cells, they don’t have to. Bloom Energy delivers affordable, always-on, ultra-reliable onsite power, built for chipmakers, hyperscalers, and data center leaders looking to power their operations at AI speed. Learn more by visiting BloomEnergy.com.
Catalyst is supported by Third Way. Third Way’s new PACE study surveyed over 200 clean energy professionals to pinpoint the non-cost barriers delaying clean energy deployment today and offers practical solutions to help get projects over the finish line. Read Third Way’s full report, and learn more about their PACE initiative, at www.thirdway.org/pace.
Transcript
Tag: Latitude Media: Covering the new frontiers of the energy transition.
Shayle Kann: I’m Shayle Kann, and this is Catalyst.
Ekin Dogus Cubuk: I have to say there’s a difference between winning gold medals in math olympiads and scientific discovery. You can practice for the math olympiad by studying previous year’s problems. You can’t really practice how to discover the next big theory, but they were getting better at reasoning on complex problems.
Shayle Kann: Coming up: can AI discover a room temperature superconductor? Volume two. I’m Shayle Kann. I lead the early stage venture strategy at Energy Impact Partners. Welcome. So a year ago, a little over a year ago, I had Doge Cubuk on this podcast to talk about using AI for materials discovery, which has all sorts of interesting applications in the spaces that we talk about here. At that time, doge had been leading efforts in that area for Google DeepMind for some time, and I thought of him as being both very knowledgeable in the space, obviously also pretty sober about it. Fast forward a year, doge left Google DeepMind earlier this year and along with Liam Fedus, who was one of the co-creators of ChatGPT started a company called Periodic Labs, which raised wait for it, a $300 million seed round led by Andreessen Horowitz. Periodic is doing AI for materials discovery and not just that, also physics and chemistry, and they’re also very much hardware in the loop.
The way I like to frame it is that they’re building two kinds of frontier labs at once. There’s a frontier AI lab and a Frontier scientific lab, the type of lab that we used to talk about. And then they’re trying to make those two things work together to make breakthrough discoveries. Notably, one thing we talked about last time was how the AI materials discovery companies at the time tended to start by going after often discovery of something like metal organic frameworks or MOFs for carbon capture, which I think of as less of a breakthrough opportunity really from a global scale. Whereas the big, perhaps biggest breakthrough to prove would be the discovery of a room temperature superconductor. Well, Periodic makes no promises, but they’re very publicly working on high temperature and maybe room temperature superconductors based on that last conversation. To be honest, I wouldn’t have predicted this, so it was time to have Doge back on and hear what changed and how.
Here’s Doge.
Ekin Dogus Cubuk: Doge, welcome back.
Shayle Kann: It’s great to be back.
It’s great to see you. A lot has changed since the last time we talked. So I looked back. So the last conversation that we had was just over a year ago. It was September, 2024, and I was having you explain to me the wild world of AI for materials discovery in particular and the work that you’d been doing at Google DeepMind, but also just the broader landscape. And I’ll tell you my takeaway from that conversation, which you could tell me if I had the wrong takeaway at the time, but my takeaway was promising field, pretty unclear if and when this new wave of LLMs and all the reinforcement learning all the things that have shown up in the past few years, pretty unclear if and when that would generate a real meaningful breakthrough discovery in materials. And we talked through a bunch of the reasons why it’s challenging training data, maybe chief amongst them, but I came away with a pretty, I think a sober view of the path there. Okay, so fast forward a year and you left Google DeepMind, started a company to do that amongst other things. So I guess the first question that I have for you is what changed in the last 12 months to give you conviction that now is the time?
Ekin Dogus Cubuk: Great question. When we talked, I was doing research in the field of computational material science and machine learning specifically. We were using graph neural networks, we were using density functional theory, and we were trying to discover materials. One thing that changed since our discussion was the LLMs have improved even further. So at the time I wasn’t using LLMs much at all, but I think right around when we were talking the o1 came out, the reasoning models started showing up and that was a huge update for me because you might remember that one of my big concerns is machine learning works best on the training set distribution, but in science and technology, we almost only care about out-of-domain generalization. What O one showed is if you spend test time compute, you can get better results. That was very exciting to me because there was one way of investing resources there was beyond the training set.
Shayle Kann: So okay, if I can try to translate that into layperson terms, the reasoning model, like OpenAI’s o1 model, introduced- unlocked a door kind of, that maybe allows you to break this challenge of the limited training data set that you have in materials discovery. That was what we spent a lot of time talking about a year ago, was like you can compare the corpus of data that an LLM trains on to do language, which is enormous. It trains on the internet basically versus the corpus of data that you were dealing with in trying to discover novel materials and it was thousands of data points, not tens of billions or whatever. And so that presumably hasn’t changed at least yet. But you’re saying that the reasoning models have gotten good enough that they are able to sort of get around that challenge via reasoning or possibly generating their own synthetic data. What is it that allows them to break that?
Ekin Dogus Cubuk: So I’m not saying that they’re good enough already, but that was one step in the positive direction. And another thing we’ve seen is they’ve gotten really good at math. So since last time we talked they started winning gold medals in math olympiads and they’re doing similarly well on the coding really well on physics olympiads. And I have to say there’s a difference between those things and scientific discovery. You can practice for math olympiads by studying previous year’s problems. You can’t really practice how to discover the next big theory, but it did show you that they were getting better at reasoning on complex problems. So then what else do we need? So I think the biggest thing we need is to have our own lab because once you have a very intelligent reasoning, LLM, you still can’t discover things unless you make trials, right? Just like humans, the LLMs will be wrong often when they try to predict things outside of their training set. But you try many things and then at some point you get a really cool discovery, and this is, as we talked about history last time, this is quite common in solid-state chemistry, solid-state physics, but a lot of discoveries happen somewhat by accident, but of course with a lot of background understanding of the physical system and a lot of trial and error.
Shayle Kann: So this is what you’re doing at Periodic, right? You’re combining the digital domain with the physical domain. You have a lab in both senses, it’s a frontier lab and an LLM sense and a frontier lab in a laboratory sense in the traditional sense of the word. And you’re sort of merging the two. I’m curious in practice how you imagine that feedback loop working? So is it a traditional, you develop a theory, you run an experiment, you generate data from that experiment, but in this case you feed the experiment back into your customized LLM as an additional set of training data and then that’s the way that the loop works, or is it more complicated than that?
Ekin Dogus Cubuk: Yeah, exactly. I mean it’s pretty simple. I think as you said. So the LLM can propose, for example, synthesis recipes or it can propose simulations to run and because the LLMs are pretty good at tool use, it can actually do it itself and then you get some results back. So the results from experiment could be some characterization data results from the simulation can be some trace or some simulation you did, and now the LLM can go through it with the context of its previous training, maybe the context of relevant papers, textbooks, but also now the results that he just got that no one else have ever seen. And then now it can kind of tweak the experiment, tweak the simulation for the next step.
Shayle Kann: Right. You said one thing in there that I guess is worth pointing out. You’re trying to automate this as much as possible. The LLM might run the experiment.
Ekin Dogus Cubuk: Yeah, absolutely. I mean, one of the other advances that’s been happening recently that I think made Periodic possible is the high throughput experiments have been getting better. There are many examples of this now across academia industry where these robots, they became quite commoditized actually just mixing powders or mixing liquids and then sending it to characterization. I think one thing that isn’t as advanced right now, but we feel like we can do pretty soon is automated characterization itself. So you mix powders, you put it in some characterization tool, you get the result out what is the actual output? I think this pretty difficult right now for AI tools, but we feel like we can improve that pretty quickly.
Shayle Kann: I want to talk about one specific application that I know you’re going after that we actually did talk about a year ago, but also I want to talk about it as a way to see whether one of the other things you described as the fundamental challenges has changed, which was, as I understood it, AI being pretty good at the next incremental discovery, but not necessarily good at the breakthrough discoveries you said through history. Usually that’s done accidentally or often it’s done accidentally because it’s not, you can’t reason your way to this massive breakthrough discovery. So let’s talk about superconductive materials. We talked about this last time where all these companies that existed at the time that were doing AI for materials discovery were starting on things like discovering a novel MOF for carbon capture or whatever. But we said the thing that would be the real breakthrough, the big thing would be a room temperature superconductor. And you guys have since launching, been very public about superconductive materials, maybe not room temperature. I’m curious for you to tell me how likely you think that is, but high temperature superconductors is on the roadmap. So why, first of all, and then second of all this question of do you think you have a path to the truly breakthrough? What would the path be to a truly breakthrough discovery as opposed to finding something that is a material that is superconductive at a ever so slightly higher temperature than the best that we’ve got today?
Ekin Dogus Cubuk: To answer your first question, I think it’s still true that it would be difficult to just reason your way into a much better superconductor. I actually would guess that there’s a law out there that we haven’t discovered yet that says that you can’t just look at your training set that’s different than what we’re trying to discover and just predict it. There’s been rules that we discovered from 1800s on where you connect energy to work. So thermodynamics is the first example. There’s more recently Landauer’s limit, which shows that you have to spend a certain amount of energy to delete information, which can be used to describe Maxwell’s demon contradiction. I bet there’s something similar for how hard it is to discover things. It’s outside of your training domain. So I don’t think that’s been fixed since last time we talked. But because we have a lab internally, we can just try things and try them at large scale and often and hopefully as intelligently as possible.
So even though we won’t reason our way into a much better superconductor, we’ll be able to push our trials in the direction that’s most promising or most promising for us given our training set at the time. So yeah, I think that hasn’t changed, and I think there’s reason to be hopeful because in the big scheme of things is a pretty new field. I mean, if you look at cuprates they were from 1985, there’s been a lot of advances more recently. So yeah, we’re very excited. One reason we chose superconductivity is if you find a good superconductor that’s impactful immediately. Last time we talked about how long it can take to translate materials improvements into products. One nice thing about superconductors is if somebody discovers a room temperature superconductor today, even before it makes it into a product, it has huge impact. First of all, it changes how we think about the universe.
Second, it helps us do physics experiments that wasn’t possible before. And whenever you think about a sci-fi ish technology like quantum computing, fusion superconductors come up because that’s kind of what we need. It’s kind of like one of the most exciting macroscale quantum phenomena. So that’s one reason we picked it because it kind is exciting as soon as we succeed. The other reason is it requires all sorts of improvements to get there. When we think about OpenAI and DeepMind, I remember back in 2016, people used to make fun of these institutions for prioritizing a GI so much because they were saying we’re going to do a GI. But what happened is they developed so many other tools on the way to a GI, they were useful in themselves, but today they have these LLMs that you might consider a GI or something really impressive.
Superconductivity is a bit like that. To discover a exciting superconductor, we probably have to develop so many capabilities on the way there. That’s by themselves very useful. For example, automated synthesis, automated characterization, being able to model or predict high temperatures per conductivity because we don’t have a theory for it yet. So it’s kind of like a nice goal that unites people and requires a lot of other useful things to happen on the way. And it’s one of those things that physicists find really exciting. So the physicists in our company are really excited by this mission, but also computer scientists find it very exciting. It’s one of those things that I think both sides can really appreciate. So those are some of the reasons that we picked it.
Shayle Kann: I guess back to this question of how do you distinguish between the incremental innovation, which to be clear if you develop or discover a superconductor at a higher temperature than anything that we’ve got today, that’s meaningful, but it’s probably orders of magnitude less meaningful than if you discover a room temperature superconductor. And I presume that the scientific challenge is commensurately distinct between those two. And the way that LLMs work, as I understand it, at least in part, is on these reward functions. And so are you setting your AI system a goal of find a room temperature superconductor, and then everything flows back from there? Here are the steps and all the things we have to fix to get to room temperature or do you say improve this characteristic such that we can incrementally build our way there? In other words, are you going to find 10 super? I think of it as sort of a different thing from, but the alternative version of this is what happens in fusion nuclear fusion where everybody is sort of chasing this same goal of Q is greater than one right energy break even, and everybody is getting incrementally closer and closer and closer and eventually NIF breaks it or somebody breaks it. Is it going to look like that or is it going to look like we’ve discovered nothing until we discover the room temperature superconductor.
Ekin Dogus Cubuk: So as you said, I think there are many different ways of improving superconductors without getting a room temperature superconductor. So one of them could be having a significant increase in Tc, but another one could be a really high critical magnetic field, which turns out might even be more important for fusion applications than Tc itself. Another one can be more mundane like some mechanical properties like a superconductor that also is ductile and we can make it into devices. So we wouldn’t rule out all of these very exciting developments just for a room temperature Tc.
Shayle Kann: But how do you set the reward function for your model? What are you optimizing it for then?
Ekin Dogus Cubuk: I think that’s an empirical question. I think one thing I should say is it’s quite nice because it’s hard to reward hack. One of these issues with RL and training LLMs is you might worry about reward hacking and in simulations, again, reward hacking can be a problem even in DFT, but for real life experimental measurement of tc, it’s much harder to reward hack, which we love. So if our reward was increasing tc, that just seems like a nicer unhackable reward. But in terms of what specifically will get us there, we’re not sure yet. I mean, it’s an empirical question. We can probably try all of them. I’ll list the things you propose and we’ll try all of them.
Shayle Kann: I guess that gets to the other question, which is what does the human in the loop look like here? And again, as you said, if we haven’t solved the sort of, AI-is-good-at-incremental-innovation-and-not-orthogonal-breakthrough-innovation thing, but humans are historically at least better at it. Is it like folks on your team developing a theory of something and that gets fed through the model and you get the results and you feed it back in? You see whether it’s a promising category, is the germ of the original idea of what to look for coming from a human or is it coming predominantly from the model and then the humans have to interpret and send it off in various directions?
Ekin Dogus Cubuk: Yeah, I mean, that’s a great question. We are not really prioritizing full automation anyway, so if we get better results with humans doing part of it, that’s great. This is also actually a question for our lab. Do we want to automate every single aspect of the lab? At some point you end up needing humanoids for that. And I think that’s not Liam, my co-founder and I, we are trying to be very pragmatic about it. Our goal is to get the best result possible on the things we care about. And how much of automation comes from the ml, how much of it comes from more traditional tools and how much what gets done by humans? I think that’s kind of again an empirical question. So yeah, I think as you said, it does seem like today there are things that ml AI is better than humans, but one of those things is not hypothesis generation. I mean, there are two options that we either have to improve these LLMs and hypothesis generation, which is possible. Or the other option is we have humans providing some of the hypotheses and then AI doing the execution.
Shayle Kann: I guess the other question here is cost. I mean you guys raised a $300 million seed round, so that implies on the outside that your cost structure will look similar to other companies that basically have to, are going to use just an enormous amount of compute. And so a lot of that cost comes from compute in your context. I could imagine maybe that being true, but also maybe that not being true because again, you just don’t have the same corpus of data. You can’t build a 10-billion-parameter model right now because the data isn’t there to do it. And so instead that cost is going to go more toward the robotic lab and all that kind of stuff. How should I be thinking about how much compute you’ll use and where that costs come from?
Ekin Dogus Cubuk: So honestly, compute is very expensive and we are going to train LLMs. We are going to use GPUs to run simulations. So that end up being a large part of the cost. Yeah, it is funny. Before if you asked me this question 10 years ago, I would’ve thought that the biggest part of the cost must come from the labs because physical is real. You’re building this lab, you’re buying instruments, but turns out the GPUs are so expensive and training LLMs is so expensive. So when we were thinking about how much to raise, we kind of laid it out in terms of the GPU cost, the lab cost, and this was kind of minimum number we felt like was viable. And yeah, we’ll see. The GPUs have been getting more expensive recently. I guess we’ll see how the market dynamics continue.
Shayle Kann: To what extent do you end up building generalized model or models versus models designed to a specific domain, even a specific scientific domain? You guys are doing your material discovery obviously, but physics and chemistry and these things all intertwined. But is the same model going to be equally capable across all domains? Is that the intent or is that just not how they’re supposed to be architected?
Ekin Dogus Cubuk: That’s right, and that’s actually something we’re very excited about. One thing I’ve been kind of noticing is in the past, say three, four years, I had a chance to collaborate with very world class best in their field scientists. And even when you work with them, you realize that while their expertise on a few domains is incredible, maybe best in history, there’s just so much more to know in chemistry and physics that they may not know all the other aspects of it. So this is why I brought up superconductivity because you might actually need to be really good at solid-state chemistry and synthesis or difficult novel materials just because you don’t know which chemistry the superconductor is going to come from. Some of these ideas you might have may not be as stable thermodynamically. So you need to be intelligent about how to kinetically force it into that phase you want.
But at the same time, in addition to solid-state chemistry, you need to be incredible at condensed matter physics because there’s so many different kinds of superconductivity we don’t understand most of them very well, and there’s nobody in the world who knows both of those equally well or sufficiently well. And turns out this is true for many different aspects. If you need to use robots for high-throughput synthesis, again, there are only so many people who understand robots and how to use them for synthesis. So I think this was different in 18 hundreds probably. There was probably a time when a physicist could contribute and be one of the best in the world on many fields of physics, but it’s definitely not true today. And this is one of the reasons I think we are very excited about LLMs because when you talk to them, they seem like they have a pretty good understanding of solid-state chemistry and solid-state physics at the same time already.
And we’re trying to improve them further in the physical sciences specifically because that’s where we are really interested in. And then we are hoping that they’ll be good at multiples of these. And then a really exciting prospect with that is a lot of the exciting discoveries happen to lie in between fields, right? That’s why it’s sometimes easier to be interdisciplinary. And there’s so many of these surface areas between these different fields. It’s like I guess science is kind of like a fractal in the way it’s hierarchically organized and there’s so much surface area that humans have exploited of course. But then there’s probably so much left to exploit and we are excited about an LLM. They can basically do that at a scale that humans couldn’t yet.
Shayle Kann: How good are the LLMs today or the best in class of what you guys have at generating synthetic data in this domain? Or another way to ask this question is if you’ve fast forward three years, you’re fully up and running and you’re operating, how much of the valuable insight you will generate do you think will come from the physical data coming out of your lab versus the synthetic data that the LLMs create on top of that?
Ekin Dogus Cubuk: Yeah, that’s a great question. I obviously don’t know the answer, but it’s great to brainstorm about that because on the one hand, the lab data will be our additional data that other LLMs may not have, and you might think that the lab data will only be as valuable as the results in it. But on the other hand, what’s interesting about scientific data is it’s not just a few bits or numbers. For example, there are certain experiments you can run where the result you get from it is just say three floating point numbers. But the implications of those could be tremendous. It is not just going to be a few bites. It will actually be potentially an incredible amount of understanding just from a few experiments. And this has been how it is in human history. There’s certain experiments that told us so much about how we understand about the universe and the way to do this with synthetic data can of course be you run simulations that relate to that experiment and when you get the experimental result that actually validates or refute so much of the simulations you run.
And then that is a lot of information in itself. So it’s a very interesting question, and I think there’s some actually differences about how you think about synthetic data when it comes to an LLM that’s good at science and exactly. This is one of the reasons I really want to work on this because this opens up questions for LLMs and LM training. This may be different than what the Frontier Labs are thinking about right now. If they’re only thinking about math and logic and what’s on the internet, like accounting tasks, that’s a bit different than if you’re trying to do experimental physics, experimental chemistry. It just seems like a very exciting question to explore.
Shayle Kann: I want to talk a little bit about how you build a business out of this. I mean, you mentioned the Superconductor example and you said there’s a lot of value in this long before this novel superconductor goes into a product, but ultimately kind of to go into a product of some sort for you guys, and I think we talked about this a little bit last time too. There’s this question of, okay, so if your job, your core job at Periodic is to discover new things that are going to be valuable in the world, say you do it. To my mind, there’s sort of a binary decision you have to make at that point. Do you try to sell the discovery license, the technology license the IP to somebody else who’s going to go produce it and turn it into a product, or do you produce it? Do you sell the product? Do you sell the discovery or do you sell the product? Do you ever prior on which direction you want to go here?
Ekin Dogus Cubuk: Great question. And I think the two options can be correct depending on the context, depending on the timelines, but honestly, it also depends on where we are in the company. So at the very beginning, you can imagine our LLMs will be very impactful for other companies doing physical r and d already, like contract RD basically. Yeah, exactly. I mean, there’s a lot of interest in being able to use these LLMs. Sometimes the data restrictions don’t allow it because you don’t necessarily want to put your data on an LLM on the web sometimes. The other issue is you haven’t trained the LM on your data, so it’s actually not as good as it could have been. That kind of improvement could be really impactful because we’ve seen how impactful LLMs can be in other fields where they have access to the data. So there’s a lot of, I think, headroom for impact there.
But in the longer run, you can also imagine a case where we as a field get really good at designing materials intentionally. That hasn’t been the case, but if you look at drug design, there was a time when designing drugs wasn’t very profitable, and I think people will look at it and say, this is not a good business. But what happened with Genentech is the field got so good at designing drugs at some point it became very valuable itself. The machine learning field has been making huge improvements in material science that was kind of hard to predict. So it’ll be interesting to see how far that goes and whether material discovery by itself becomes a very exciting business similar to drug discovery, but for us, we already see this big need and the big potential for impact by providing these LLMs to do physical r and d.
Shayle Kann: Yeah, almost like this is going to be the wrong analogy, but it’s partially right, almost like an AWS, you’re going to have the infrastructure, in this case, the infrastructure is your custom designed LLM that is smart about physics and chemistry and all these domains, and also your physical lab and them being interconnected with each other. And so you have all this infrastructure and scale in that infrastructure that you use to go convince whatever large company that’s doing r and d, that they should just be outsourcing it to you rather than rebuilding the whole same thing, which is not exactly what the cloud providers are, but there’s enough of a relationship there. So that feels right, but it is, I suspect, yeah, I guess is what you’re saying. I suspect a smaller ultimate opportunity than the, you proactively discover a bunch of novel materials that change the world, and then however you monetize them, you prove you’re able to do so repeatedly and then your Genentech and it’s a whole different category.
Ekin Dogus Cubuk: Yeah. One other thing you see is people are so excited about this. They want to see LLMs not just conquer the digital world, but also really impact the physical world and impact the atoms basically. So I feel like this has to be done, and the team has been very excited. It is really amazing. We’re hosting weekly seminars where the physicists will teach the computer scientists about the physics and the computer scientists will teach the physicists about LLMs. And of course there are a lot of people in between. It is actually again, like a fractal. So yeah, I think there’s been a lot of excitement about seeing if these technologies can be used not just for the digital world, but also for constructing the atoms around us.
Shayle Kann: I guess maybe the last question, you said at the beginning, the thing that changed between a year ago and now in part was advancements in the big LLMs, right? The o1 model and so on. Is there a next, what could OpenAI release in a year or two years from now that would be a big leapfrog for you? Are you branching off now from what the big LLMs are going to do and everything that all advancements are going to come from Periodic, or is there something else that they could offer that is a step function change in your capability to discover new materials or whatever?
Ekin Dogus Cubuk: Yeah, great question. I think we actually basically rise with the tide, right? As LLMs get better, there’s so many advantages of that to other applications. For example, the LLMs are getting very good at coding, and that’s not surprising, right? Because programming is a kind of close the environment. You can just simulate in your computer and get valuable feedback and then quickly improve. But as computers get better at coding, that’s huge for science because then you can run simulations more efficiently. The simulations themselves can improve similarly with tool use experiments. So I think as LLMs improve in general, there’s going to help a lot with science applications. There are maybe longer term things that can happen. One of them could be things like hypothesis generation or more out-of-domain generalization. But then a question there is, will that come from status quo, like how LLMs are being trained now? Or will it come from actually labs that try to improve scientific reasoning for these LLMs? Because then may be hypothesis generation emerges naturally, or out-of-domain generalization emerges naturally because that’s what you’re kind trying to get at with your reward. So I think that’ll be a very exciting question to see maybe next time we chat.
Shayle Kann: Love it. Alright. Thank you so much for taking some time. Again, congrats on Periodic. Super excited to see what you guys discover and for when your room temperature Superconductor is shooting electricity all around the world around me.
Ekin Dogus Cubuk: Okay, well thanks a lot. It was a great chat.
Shayle Kann: Dogus Cubuk is the co-founder of Periodic Labs and a former researcher at Google DeepMind. This show is a production of Latitude Media. You can head over to latitud media.com for links to today’s topics. Latitude is supported by Prelude Ventures. This episode was produced by Daniel Woldorff, mixing a theme song by Sean Marquand. Stephen Lacey is our executive editor. I’m Shayle Kann, and this is Catalyst.
