Audio · 2026-04-03 · 29m · 12 moments

AI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus

What happens when you apply the scaling laws of large language models to the physical work of atoms? Elad Gil sits down with Liam Fedus, co-founder at Periodic Labs, which is pioneering an AI foundation lab for atoms. Liam discusses how he pivoted from dark matter physics research to the front lines of artificial intelligence, including stints at Google Brain and working on ChatGPT at OpenAI. He talks about how Periodic is connecting massive language models to the physical world to overcome data ✦ AI generated

Apple Podcasts ↗

GUEST Liam Fedus HOST Sarah Guo HOST Elad Gil

timeline · colored by role

Claim◆

Scientific progress in the physical world cannot accelerate at the same rate as digital progress unless you connect AI systems to real experiments — you have to interface with reality, not just sit in a room thinking.

Liam argues that science requires real experiments and grounding in physical reality, not just abstract reasoning. The central thesis of Periodic Labs is that connecting AI to the physical world is necessary for accelerating materials science.

transcript

Liam Fedus: I think just the inevitability of connecting these systems to the physical world. The opinion that I and others held as part of periodic was you're not going to see the same kind of acceleration in science and technology unless you start connecting these things to the physical world... Science ultimately isn't sitting in a room thinking really hard. You have to conduct experiments. You have to learn from them. You have to interface with reality.

explains mechanism · 1

Context

The AI technology of 2022 was too weak to connect to the physical world — reasoning, test-time inference, reliable error correction, and tool use were necessary foundations that emerged only in subsequent years.

Liam argues that ChatGPT-era technology (late 2022) was insufficient for materials science applications. Only the subsequent emergence of reasoning, test-time compute, reliable error correction, and coding agents made it possible to build systems that interface with the physical world.

transcript

Liam Fedus: The creation of ChatGPT in late 2022 was a, you know, important technology, but it was still far too weak. Like we couldn't have done periodic on technology of that era. I think over the next few years past that, we saw ever improving models. We saw reasoning. I think like test time inference became really important. That led to more reliable error correction, more reliable tool use. And we see like the rise of coding agents and other agents. And I think those were foundational technologies necessary to then connect these systems to the physical world. Like it was just not possible with the AI technology of 2022.

Context

Connecting AI to the physical world simply was not possible with the AI technology of 2022 — it took subsequent advances in reasoning, test-time inference, error correction, and reliable tool use to make Periodic Labs viable.

Liam explains that ChatGPT-era models were too weak for materials science applications. Only with later advances in reasoning, error correction, and tool use could systems be reliably connected to physical experiments.

transcript

Liam Fedus: The creation of ChatGPT in late 2022 was a important technology, but it was still far too weak. Like we couldn't have done periodic on technology of that era. I think over the next few years past that, we saw ever improving models. We saw reasoning. I think like test time inference became really important. That led to more reliable error correction, more reliable tool use. And we see like the rise of coding agents and other agents. And I think those were foundational technologies necessary to then connect these systems to the physical world. Like it was just not not possible with the AI technology of 2022.

Data◆

Published scientific literature is often unreliable for training ML models because reported material properties can span many orders of magnitude — you need ground truth from your own experimental data, and crucially it must be an interactive closed loop, not just a static pool.

Liam describes how literature-extracted values for the same material property can vary wildly, forcing Periodic to use their own experimental data in an active, iterative loop rather than static datasets.

transcript

Liam Fedus: One of the engineers on our team was looking at a reported material property. And it was just sort of extracted values from literature. And it was really interesting to see the reported value spanned many orders of magnitude. And so you train an ML system on that and it's like, well, the best you can do is model this distribution, but you're no closer to like a ground truth. And that's where experimental data comes in, where you now have a grounding in this. But really important, it's not just like a pool of data. It's this interactive closed loop system that is so powerful. Once you have the experimental data, you can look through it, you can look for aberrations, you can look for patterns, you can look for consistency with simulation data, with literature, and then that helps drive the next set of experiments. So it's not just a pool of data, it's a very active loop.

explains mechanism · 1

Mechanism◆

AI systems become limited without access to raw experimental data — literature alone is insufficient because reported values for material properties can span many orders of magnitude.

Liam explains that scientific literature data is unreliable for training ML systems because reported material property values vary wildly across papers. Only experimental data provides grounding to ground truth, and it works best in an interactive closed-loop system where data drives the next experiments.

transcript

explains mechanism · 1provides context · 1

Mechanism

Periodic Labs treats language models as an orchestration layer — LLMs direct specialized neural nets designed for atomic systems, which have symmetry awareness, lower latency, and serve as tools and reward functions.

Liam describes Periodic's architecture: language models act as a copilot/orchestration layer that ingests literature and experimental data, then directs specialized neural nets built for atomic systems. These specialized models have symmetry awareness, lower latency, and function as tools and reward functions within the overall system.

transcript

Liam Fedus: Language models are incredibly powerful. It's a very natural interface. And so we continue to use these. But we think about them almost as like an orchestration layer. So that's sort of a copilot assistant, but also like a system that can direct experiments. And it's almost, it's orchestrating other specialized models as well. So we do construct neural nets that are specially designed for atomic systems where there's like some symmetry awareness and those have much lower latency and they've been like fine-tuned for that. And so basically you kind of think of this like orchestrating layer that can ingest literature, it can go through our experimental data, it can go through different modalities, but they can also use specialized neural nets as tools as reward functions. So it's like an overall system.

Mechanism

Periodic Labs uses language models as an orchestration layer that directs experiments and coordinates specialized neural nets designed for atomic systems, which serve as tools and reward functions.

Liam explains their technical architecture: language models act as a copilot and experiment director, orchestrating specialized neural nets built with symmetry-awareness for atomic systems, treating them as tools and reward functions.

transcript

Liam Fedus: Language models are incredibly powerful. It's a very natural interface. And so we continue to use these. But we think about them almost as like an orchestration layer. So that's sort of a copilot assistant, but also like a system that can direct experiments. And it's almost, it's orchestrating other specialized models as well. So we do construct neural nets that are specially designed for atomic systems where there's like some symmetry awareness and those have much lower latency and they've been like fine-tuned for that. And so basically you kind of think of this like orchestrating layer that can ingest literature, it can go through our experimental data, it can go through different modalities, but they can also use specialized neural nets as tools, as reward functions. So it's like an overall system.

gives example · 2provides context · 1

Prediction

Despite atoms being hard, there is at least an order of magnitude or two of speedup possible in the physical world simply by making sense of huge amounts of data and getting to solutions more quickly.

Liam contrasts the rapidly changing digital realm with the physical world, acknowledging that atoms impose real limits. But he argues those limits don't preclude one to two orders of magnitude improvement through better data utilization and faster solution-finding, with profound implications for semiconductors, aerospace, and energy.

transcript

Liam Fedus: I mean, we see how quickly the digital realm is changing. Software engineering now looks wildly different than even six months ago. But I think we see like similar opportunities in the physical world. Of course, like atoms are hard and so you will have some limits of physics. But just because atoms are hard doesn't mean there's not an order of magnitude or two to speed up. just making sense of huge amounts of data and getting to solutions more quickly. Yeah, so I think what we're trying to do is give humanity this agency for atomic rearrangement synthesis, and we think it's going to just be a huge accelerator.

Prediction

Physical sciences will follow the same industrialization trajectory as machine learning — moving from small-scale, few-GPU research to scaling-law-driven, capital-intensive operations with hundreds of researchers and massive compute.

Liam draws an analogy between the early Google Brain days (few GPUs, few people pushing the frontier) and the current industrialized ML era driven by scaling laws. He predicts physical sciences and engineering will follow the same pattern, with Periodic aiming to bring large-scale experiments and the scaling mindset to the field, enabled by both intelligent systems and automation.

transcript

Liam Fedus: Going back to the early Google brain days where the frontier was pushed forward by a few GPUs and a few people, now you look at this era where it's really like industrialized and there's dozens, hundreds of researchers working together with hundreds of thousands, millions of GPUs, dictated and driven by scaling laws. Everything is about scaling. It's given that predictability. It's allowed us to put huge amounts of capital into this field. And I think the physical sciences, physical engineering will have a very similar property where we establish these scaling properties and bring that mindset. And so periodic in this field is really thinking about how do we bring much larger scale sets of experiments to bear on this. And intelligent systems have enabled this, automation has enabled this, and you really need both.

Prediction

The physical sciences will develop scaling laws analogous to those in AI — once you establish predictability through scaling, you can deploy huge amounts of capital, just as frontier AI labs did with LLMs.

Liam draws a parallel between the early Google Brain era (a few people with a few GPUs pushing the frontier) and today's industrialized AI, predicting materials science will follow the same trajectory as scaling properties are established.

transcript

Liam Fedus: Going back to the early Google brain days where the frontier was pushed forward by a few GPUs and a few people... Now you look at this era where it's really like industrialized and there's dozens, hundreds of researchers working together with hundreds of thousands, millions of GPUs, dictated and driven by scaling laws. Everything is about scaling. It's given that predictability. It's allowed us to put huge amounts of capital into this field. And I think the physical sciences, physical engineering will have a very similar property where we establish these scaling properties and bring that mindset. And so periodic in this field is really thinking about how do we bring much larger scale sets of experiments to bear on this.

explains mechanism · 1extends · 1

Prediction

There is a very clear path for recursive self-improvement in software engineering because it has cheap verifiable environments (unit tests) and no domain expertise gap, but the same does not yet hold for other domains.

Liam argues that self-improving AI systems are on a clear trajectory in software engineering, where testing is cheap and domain knowledge is abundant, but warns that intelligence is not a scalar — systems can be world-class on one domain and poor on adjacent ones.

transcript

Liam Fedus: I think one fallacy is thinking about intelligence as a scalar. We've consistently seen these systems have a very odd spikiness... So basically you can make a system that's like a genius at one thing and not very good at a bunch of other stuff. And I guess the point I was making is those fields can actually be quite adjacent... But one way I think about recursive self-improvement is it's really kind of akin to neural architecture search from roughly 10 years ago. And I think there's a very clear path for software engineering. So these systems have become so incredibly impressive on this domain as a result of huge amounts of data, really cheap verifiable environments. Like you can check unit tests go from failing to passing with just a few CPUs. It's basically instantaneous. There's no domain expertise gap between an AI researcher, software engineer. And obviously this will become and is becoming a larger contributor to the next generation of the system.

explains mechanism · 1provides context · 2

Claim◆

AI systems exhibit non-intuitive spikiness in intelligence — a system can be world-class at one math domain yet degrade substantially with minor question perturbations, resembling a bad high school student on adjacent topics.

Liam challenges the scalar view of intelligence, arguing that AI systems show odd spikiness: they can be geniuses at one task and poor at another even in adjacent fields. He connects this to neural architecture search and notes software engineering is the clearest path for recursive self-improvement due to abundant data and cheap verifiable environments.

transcript

Liam Fedus: I think one fallacy is thinking about intelligence as a scalar. We've consistently seen these systems have a very odd spikiness, and it's actually possible to architect a system that is world-class on some math domain. But then you could do some perturbations to the questions and actually degrade it substantially. So it's like a bad high school student. And so there's this like odd spikiness to these systems. So basically you can make a system that's like a genius at one thing and not very good at a bunch of other stuff. And I guess the point I was making is those fields can actually be quite adjacent. So like sometimes a generalization can be non-intuitive. But one way I think about recursive self-improvement is it's really kind of akin to neural architecture search from roughly 10 years ago. And I think there's a very clear path for software engineering. So these systems have become so incredibly impressive on this domain as a result of huge amounts of data, really cheap verifiable environments. Like you can check unit tests go from failing to passing with just a few CPUs. It's basically instantaneous. There's no domain expertise gap between an AI researcher, software engineer.

Highlight slides

The Core Claim: Physical Science Needs Physical Grounding✦ from: Scientific progress in the physical world cannot accelerate at the same rate as digital progress unless you connect AI systems to real experiments — you have to interface with reality, not just sit in a room thinking.Periodic Labs Thesis: Connect AI to Experiments✦ from: Scientific progress in the physical world cannot accelerate at the same rate as digital progress unless you connect AI systems to real experiments — you have to interface with reality, not just sit in a room thinking.Literature values for a single material property span orders of magnitude✦ from: Published scientific literature is often unreliable for training ML models because reported material properties can span many orders of magnitude — you need ground truth from your own experimental data, and crucially it must be an interactive closed loop, not just a static pool.Solution: your own experimental data in an active closed loop✦ from: Published scientific literature is often unreliable for training ML models because reported material properties can span many orders of magnitude — you need ground truth from your own experimental data, and crucially it must be an interactive closed loop, not just a static pool.Literature Data Is Insufficient for AI Training✦ from: AI systems become limited without access to raw experimental data — literature alone is insufficient because reported values for material properties can span many orders of magnitude.Experimental Data Enables Ground Truth✦ from: AI systems become limited without access to raw experimental data — literature alone is insufficient because reported values for material properties can span many orders of magnitude.The Active Loop✦ from: AI systems become limited without access to raw experimental data — literature alone is insufficient because reported values for material properties can span many orders of magnitude.Intelligence is Not a Scalar✦ from: AI systems exhibit non-intuitive spikiness in intelligence — a system can be world-class at one math domain yet degrade substantially with minor question perturbations, resembling a bad high school student on adjacent topics.Software Engineering: The Recursive Improvement Path✦ from: AI systems exhibit non-intuitive spikiness in intelligence — a system can be world-class at one math domain yet degrade substantially with minor question perturbations, resembling a bad high school student on adjacent topics.

Related episodes

Video · The Cognitive Revolution

Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

→

Article · Import AI

Import AI 465: Open vs closed gaps; Kimi K3; Demis' big policy plan

→

Article · Simon Willison's Weblog

llm 0.32rc2

→