Training AI on the Scientific Process | Liam Fedus (Periodic Labs) | Ray Summit 2026

0:00 Hello. It's great to be with you all today. Our goal at Periodic is to build an AI that can intentionally design the physical world. Something that can engineer matter itself. this is a material science and engineering problem. And much of our world has actually been designed in a Edisonion trial and error fashion, accidental discoveries. And what we're trying to do is bring more intentionality, more efficiency to this. I've spent many years building language models, systems that can understand and generate language.

0:34 But I think a system that can do physics, that understands physics, is going to be more foundational and ultimately more impactful. But in order to do this, we don't think that thinking is sufficient. What that means is science has been conducted by conjecturing how the world is going to be, running experiments, and seeing whether or not your explanations are consistent with reality. And that's really core to our thesis at Periodic. We're building up not just the machine learning infrastructure, the computational infrastructure, but also physical labs.

1:07 So today I'm going to walk through some of the technical pieces of closing the loop. What makes this difficult, and how we're approaching this at Periodic. This is a photo we shared earlier this spring. This is our incomplete lab, and this is one of our experimentalists, Chris, see inspecting the the current state. And again, so in order to close this loop, we have to build these high-throughput physical environments. And in order to do this, you see there's a huge amount of power, ventilation.

1:39 This is basically our source of data. These are our reinforcement learning environments. And it's interesting to think what were the >> >> our approaches to building up our current systems. So, when we built language models, math models, code models, what we did is we took a huge amount of data and we did end-to-end optimization against those tasks. So, when we were building ChatGPT, we were doing end-to-end optimization against neural net surrogates of user preferences. When we built coding models, we were doing end-to-end optimization against coding. The system would write some code and you'd check whether or not the unit test go from failing to passing.

2:24 And we think the same is going to be true for science. You have to do this end-to-end loop. And today's models are basically trained on the finished record of science. What that means is the final paper, the final textbook, it gives us like linear review of how science is done and this results in AI systems that understand the state of science. But at PeriodX, we're training on the scientific process itself. there's a lot of messiness in interpreting noisy data, decision-making under uncertainty. We want to teach a system to do science, not just to know the state, but how do you actually make progress? How do you discover new knowledge?

3:07 And to make this a bit more concrete, I want to start off with just a a brief story. There was a Nature paper from 2023 where a group at Berkeley was using an autonomous lab to do materials discovery. over the course of 17 days, they went after 57 targets. So, these are new materials and they actually succeeded making 36 of them. But there was a few examples where they actually couldn't quite figure out what was going on. And so, one of our scientists was curious and took our AI's to see if it could figure out this unexplained data.

3:41 So, this is actually a full view of the materials that that group was going after. So, there's 57 targets here and the blue bars are those that they were able to successfully synthesize. The red bars are the ones where they failed. And technical details aside, basically the lower values here mean that the that material is predicted to be more stable against decomposing into some other material or some other compounds. But I want to draw attention to this specific one.

4:09 And if you actually zoom in to the top, you'll see a string of X's. So, these are actually 16 failed attempts. So, the autonomous lab is going over and over and done, and it isn't able to actually produce this thing that's predicted to be stable. And this is sort of an interesting example to dig into because when they actually looked at the raw experimental data coming from the machine, this black curve is basically you can think of as a fingerprint of what's coming off. They couldn't explain it.

4:39 And you actually see an orange fit here that's really quite bad. It's missing peaks. sometimes the fit is trying to collapse two different peaks into one. It's just a a really lousy explanation of the data, and that's corroborated by the quantitative metrics. So, we took this as a challenge. We're like, "Okay, well, let's just send our AI look and see if we can explain this." So, we gave it the unexplained experimental evidence in the context of the experiment.

5:06 And then armed with the evidence, it starts going through literature, going through all of our experimental evidence, and it runs a series of calculations on our cluster. So, we train not just the language models, but also neural network surrogates that this uses as a tool. And after a lot of computation, it comes back with this discovery. And here this is just denoting the crystal structure. So, the balls here represent different elements, and then you have the connectivity through the bonds represented by these lines.

5:41 And we looked at all the different papers, all the different databases, and we weren't able to find this. So, this is genuinely a a new system that no one had seen before. And what's really nice about it, too, is it was highly consistent with the experimental evidence. So, the system took this unexplained data, dispatched did a bunch of computation, and it's finding something that seems very consistent with the world. And in the quantitative metrics, seemed to match this really well.

6:08 So, just coming back to our overall intention here. So, we want to create AI to intentionally engineer matter. To do this very efficiently. And we think there's huge gains to actually be had. so, I think it's useful to think Let's take a step back and look at how we've been training all the language models today. So, in the digital world, you start with a sandbox and you throw a bunch of agents in there. if you're doing reinforcement learning and there's too much variance, you can elastically increase the number of agents to reduce the variance, and you can make consistent steps.

6:41 You give these agents some tasks. The tasks typically for us, if you're doing math, code, language, are quite fast. So, the full rollout can take minutes to maybe hours. Maybe we're approaching like a day for some like the longest rollouts. The model's thinking, reading, doing literature search, using tools, but it's it's still pretty efficient. And we run this many, many times, thousands of steps of reinforcement learning, and that gives us the models we know and love, gives us better policies through these very fast iterative loops.

7:13 So, now this leads to the naive application to RL in the physical world. What if we try to do the same? So, even the scale of the experiments that we're doing, you're still going to have fewer agents. We can't elastically scale these things up. We can't just say, "Well, actually the lab needs to put forth 100 x more experiments." You can't do that really quickly. Further, some of our experiments will take days to conclude. So, the model will do some calculations, but then it's dispatching to the physical lab to actually carry out the experiment.

7:45 And in science, there's often a lot of noise in the data. So, there's like some uncertainty. It's not like a definitive you got the answer right or wrong. So, there's a there's a lot of uncertainty in interpreting the results. And so, if you just do the very naive thing, where it's like we're going to do exactly what we did in the digital world, we're going to send our experiments, we're going to wait a few days, you have a scarcity of data and a slow noisy iterative loop.

8:13 The way we think about it then is to break down the discovery process into a few very simple steps. The first is, of course, what are you trying to make? So, we saw a a special example of that in the earlier story. But then, even once you know what you're trying to make, there's a question of how do you make it? And then, finally, you have to actually look. Like, materials don't come labeled from nature. You have to actually figure out what you made. And so, the characterization and the analysis of figuring out what is present is actually non-trivial.

8:43 And so, looking at what to make, you basically want to establish what is stable. So, these atoms are going to hold together in this configuration at the conditions you care about. But we also don't just care about a novel configuration of atoms, we care about the ultimate downstream properties. What are the magnetic properties? What are the superconducting temperatures? That's a key thing in determining what to make. But just knowing what you want to make is not sufficient. And you also have to be able to build systems that can figure out what is the recipe? What is the processing in order to do this? much like cooking, you have to take the ingredients, figure out the processing recipe, and that is going to give you your final material. Again, this is a complicated calculation, doing computational material science literature, etc.

9:32 And then, the final piece is, you have the stuff that comes from the lab, how do you actually make it? So, this is an example of one of the instruments that produces these patterns, this experimental fingerprint of the atom. And so, just for this, we're going to talk about this what was made part. This was one of the first bottlenecks that we identified at Periodic in doing this end-to-end loop. In closing the loop, you have to know what you're doing.

9:57 Anyone can go to a lab and mix a bunch of things together and say they ran experiments, but if you don't actually have a a vision of what you've produced, you're flying blind. And so, our approach at Periodic has been to start with a strong open-source model. open-source progress has been really rapid. We've been building on K2-6, excited to build further on K3. And then we start by adding domain knowledge by doing a continued pre-training. So, industry calls this mid-training, some people call it tail patching, but it's really important to add our specific domain expertise from our lab, our computational data, literature that wasn't well represented in the pre-training into the system.

10:38 And then from there, we teach it specific reasoning strategies and tool use with high compute reinforcement learning. On this axis, we have one thing that is a characterization a measure of how well we're characterizing new materials. Higher is better, and this is basically saying how effective is the system at identifying the phases that are present. And we have a few frontier models where we designed this over the course of many months and built very, very strong baselines.

11:09 But what we started to do is we took Kim i K2.6, which started actually at a very big deficit to some of the frontier models. In this point In this case, something like a 25-point deficit. And on this IID distribution, as we start training more and more, we have a a system that is able to exceed the state of the art. And really importantly, this actually has a 20 to 50 x gain on output tokens from a our inference stack.

11:37 This is very important because at the throughput of the scale I just or of the lab I just showed you, if we just naively use these large language models from like the closest providers, it would cost something like $30 million annually. But by doing our own systems, we're able to get this to under a million dollars and we can put that resource towards new equipment, towards bigger training, and it's been very efficient. So, how are we achieving this? We want to talk about our our infra stack. And we've basically been building on a couple core principles. One is to evolve with the the shared frontier. So, we've been building on open-source, Megatron, SG Lang, and Ray.

12:19 But there's some some pieces where we actually want to own a differentiated core for specific workstreams where it makes sense. So, we've been building on Ray fundamentally for across all of our stack. this serves as our orchestration layer for our training, our inference, and all our workloads. And then we've been building on Megatron for our distributed training stack, SG Lang for our inference stack. But we've been building up our own reinforcement learning libraries, sandboxing, and serving infrastructure, which we'll talk about now.

12:51 So, on inference, we need the ability to scale arbitrarily within our clusters rather than having rate limits on external provider. And we also want to be able to hyper-optimize our workloads depending on what we're trying to do. So, for example, if we have prod, we really care about low latency. For reinforcement learning, we care about really high throughput. And then we also built custom load balancers so we achieve a high cache hit rate. we've been building on SG Lang and we've been able to increase throughput by 3 x and speedups of like 10 x on the weight syncs and on the weight conversions, which we again upstream into SG laying so others can benefit.

13:32 On training, our primary objective is basically to make maximize our research velocity for everyone. This lets researchers fit larger models, iterate with them faster, and scale the same workload with the same amount of compute. So, to fit the bigger models, we've actually been able to get K26 training on 64 GPUs with sorry, 64 H200s, which is smaller than any known configuration. And we're able to reduce the peak memory usage from 320 GB to 130 GB.

14:05 We've also been running much faster than out-of-the-box Megatron. I think roughly around 4 and 1/2 times on the training side. And this has been great because we've been able to scale very linearly from small jobs all the way up to our hero mid-training runs. And the final piece is our sandboxing. So, a lot of our workloads are extremely CPU intensive. Some material science jobs take a ton of CPUs, but thankfully our GPU nodes have a lot of CPUs.

14:38 So, what we've been doing is co-locating our jobs with this, and this allows us to significantly reduce latency, increase the throughput rather than relying on external providers where they're, you know, communicating over the internet could be a little bit slow. And we also needed on-prem deployments, and so we've been building a sandbox environment that runs inside our Slurm allocation. stays within our network, it's close to the workloads, and this has been giving us huge improvements. we also wanted to have a unified compute and scheduling layer. we didn't want to have dedicated sandbox clusters. We wanted to then also, as As as an RL job concluded, send that back to be used for other material science workloads.

15:27 But of course, machine learning infrastructure is only part of the infrastructure. In an addition, we've been building up the doing the hardware engineering and design of the physical lab. And since the beginning of the year, we've been running our own experiments internally at Menlo Park. What we're seeing showing here is from January to April, a comparison of what a traditional lab might do say like MIT. the green line represents something more akin to that automated experiment we saw earlier in the talk.

15:56 And then this is our current throughput. Yet as in all good things in machine learning, we're trying to push the scale and building up the infrastructure necessary to do this. This is allowing us to close the loop. But just like machine learning, we had really nice legibility conferred to to us by the scaling laws. So, as we put more compute and data behind things, we're able to see predictable gains. And similarly, we're scaling up our lab. So, at our designed full throughput, we'll be able to accomplish all the experiments we did in 4 months in a matter of less than 2 weeks. And this is really like akin to the scaling hypothesis and we're going to take the same thing to the sciences.

16:37 And we're now starting to take these tips systems to industry as well. Materials are behind every single advanced industry. So, aerospace, energy, but really importantly, semiconductors. And that's been one of our our first focuses. This also gives us a very nice flywheel. So, we've seen how AI has been contributing to better AI through coding and then increasingly systems that can do AI research. And in the last few years, a lot of big companies have put serious effort into better chip design. So, the Google TPU program was great and then Open AI's jalapeno program was announced or released on some technical details yesterday. So, you have these nice flywheels where AI is improving the chip design.

17:18 But, we want to go a layer deeper with Periodic, and we think that better materials design is also going to accelerate our use of artificial intelligence. Every new generation of logic, memory, packaging, ultimately becomes constrained by the materials and your ability to engineer these systems. I think ultimately every single add, multiply, every communication, everything is physical at the end of the day. So, behind the transformer really is the careful engineering of atoms.

17:49 computation, communication, this is dictated by electrons and propagation of photons. And it's all upstream of how do you configure and engineer these atoms. And a system that can intentionally do this much more quickly allows us to alleviate bottlenecks in these industries much more rapidly. So, AI has been changing our digital world. we're seeing new conjectures fall in math every day. few of us write code anymore. But, we think Nexus is going to help us build our physical world.

18:24 And our goal at Periodic is if the physics allows it, we want to be able to build it. That's that's our constraint, and we want to see the physical world move at some pace of the digital world. So, thank you.