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

2026-04-03 - 30 min - source - Read full transcript

Elad Gil (host)Liam Fedus

Key insights

Materials science AI was not viable until reasoning models and reliable tool use matured; ChatGPT-era models in 2022 were too weak for Periodic's approach to work.

Fedus says the opinion he and co-founders held was that science can't accelerate the way language did without connecting AI systems to the physical world, since science requires conducting experiments, not just reasoning in a room. But that required test-time inference, more reliable error correction, and the rise of coding/tool-using agents, none of which existed when ChatGPT launched in late 2022.

ai-plus-physical-world

A pool of experimental data is not enough; the value comes from an active, closed feedback loop between generating data and using it to decide the next experiment.

Fedus describes finding that literature-reported material properties can span many orders of magnitude, so training on that data alone only models a noisy distribution, not ground truth. Experimental data grounds the system, but the real leverage is looking for aberrations and patterns in that data and using them to drive what gets tested next.

closed-loop-data-generation

Generalization in physical-science models happens at the level of underlying physical laws (e.g. quantum mechanics), not uniformly across all domains.

Fedus says a model that accurately captures quantum-mechanical behavior does not automatically transfer to fluid dynamics, a different level of abstraction. The generalization Periodic sees comes from shared first principles (chemical synthesis steps, quantum mechanics, intermolecular forces like van der Waals) rather than from raw scale alone.

scaling-laws-beyond-language

Periodic uses language models purely as an orchestration layer, not as the source of domain-specific physical modeling.

The architecture pairs a general LLM, which ingests literature, experimental data, and multiple modalities and directs experiments, with specialized neural nets built for atomic systems (with symmetry awareness and lower latency) that the LLM calls as tools or uses as reward functions. Fedus sees this same LLM-orchestrator-plus-specialized-tools pattern emerging across other domains like customer support.

ai-plus-physical-world

Physical science and engineering are expected to develop the same kind of predictable scaling properties that justified massive capital investment in language models.

Fedus draws a direct parallel to Google Brain's shift from a handful of GPUs and small teams to an industrialized field with hundreds of researchers and millions of GPUs, driven by scaling laws that made investment outcomes predictable. He argues Periodic is trying to establish similar scaling properties for physical experimentation, and that doing so requires both more automation and more intelligence, since bottlenecks in either one constrain the other.

scaling-laws-beyond-language

Fedus expects near-term recursive self-improvement in software engineering specifically, not a general flip to fully machine-directed R&D.

He frames near-term recursive self-improvement as analogous to neural architecture search from about a decade ago, enabled by cheap, verifiable environments (unit tests passing or failing, checkable with a few CPUs almost instantaneously) with no domain-expertise gap between AI researchers and software engineering. He explicitly says this is 'happening nowish' for coding, but does not extend automatically to biology or other domains with a knowledge gap.

recursive-self-improvement

AI research self-improvement is a slower loop than software self-improvement because its feedback signal (scaling behavior, convergence, generalization) requires GPU-hours of experiments, not instant unit-test checks.

Fedus contrasts the near-instantaneous verification loop of software engineering with AI research, where checking whether a model converged or generalized well requires actually running the experiment for hours. He argues the same closed-loop principle underlies Periodic's premise for physical science: doing science and engineering both need closed loops connected to the physical world, and other domains will follow this pattern with some delay.

recursive-self-improvement

Intelligence in current AI systems is spiky, not scalar - a system can be world-class in one narrow domain and fail badly on small perturbations of the same problem type.

Fedus calls it a fallacy to think of intelligence as a single scalar quantity. A system can be world-class on some math domain, then degrade substantially when questions are perturbed, behaving 'like a bad high school student' - meaning claims of general AGI/ASI capability need to be domain-qualified rather than treated as a single capability level.

recursive-self-improvement

Robotics is not a prerequisite for Periodic's closed loop, but improving robotic reliability will be a major accelerant for scaling physical experimentation.

Fedus says Periodic currently combines human labor with reliable but narrow automation, using largely off-the-shelf, commoditized robotics rather than heavy custom innovation. A dexterous humanoid able to work reliably in an unstructured lab would remove a major bottleneck, since current physical-system automation requires careful, slow custom design (echoing Elad's account of building custom liquid-handling robots at Color).

robotics-and-automation

Compute, not physical infrastructure, is the dominant capital cost for AI-driven physical science, despite long lead times on lab hardware.

Fedus says GPUs are extraordinarily expensive relative to physical lab infrastructure, to the point that infrastructure cost is sometimes lower than compute cost even though physical systems have much longer lead times and intrinsic difficulty being well-calibrated. He frames Periodic's capital intensity as primarily a compute-cost problem.

scaling-laws-beyond-language

Scientists in academia are systematically undercompensated relative to the value they create, and commercial AI labs can correct that by scaling their impact.

Fedus notes the pay gap between a Stanford postdoc and a machine learning engineer is stark, and says he likes that companies like Periodic can bring highly capable scientists into a setting where their work has more leverage and better compensation, in addition to giving them tools to work at much higher throughput than academic labs allow.

ai-plus-physical-world

Periodic is starting as a software/intelligence-layer business but leaves open a biotech-style discovery-royalty model if certain breakthroughs prove high-value enough.

Fedus compares the choice to biotech's fork between partnering with big pharma for royalties versus building and owning drugs outright. Periodic currently positions itself as an intelligence layer / system of record and control plane for partner companies' experiments, but Fedus acknowledges some discoveries could justify a more proprietary, higher-value discovery-model approach.

ai-plus-physical-world

Books referenced

The Diamond Age - Neal Stephenson - Elad brings up the 1990s novel's AI tutor and matter-compiler ('matter pipes' / 3D printers in every home) as a reference point for what an AI-plus-materials future could look like; Fedus frames Periodic's mission as giving humanity agency over atomic rearrangement and synthesis, echoing the book's premise.

Companies

Periodic Labs - Fedus's company, building what he calls an AI foundation lab for atoms: language models orchestrating specialized neural nets and physical lab experiments to accelerate materials science and chemistry discovery.
OpenAI - Fedus was VP of post-training; worked on productionizing GPT-4, including the original ChatGPT effort pushed by John Schulman to keep the product general-purpose rather than a narrow writing or coding bot.
Google Brain / DeepMind - Fedus was among the first-year AI residents at Google Brain (2016-2017) during the era of distributed training, mixture-of-experts, and the Transformer; Google Brain was later combined with DeepMind.
Anthropic - Cited alongside OpenAI as an example of a frontier lab founded by a physicist (Dario Amodei) and growing fast because language plugs into a massive existing domain of human activity.
Color - Elad's prior company, a genomics/health business where they built custom liquid-handling robotic systems, requiring heavy customization (ML-monitored cameras, 3D-printed parts to cut vibration) because off-the-shelf lab robotics and firmware were inadequate for small liquid volumes.
AlphaFold - Cited as an example of a breakthrough model built on a very specific, decades-deep existing dataset (protein crystal structures), contrasted with the sparser experimental data available in general materials science.

Techniques and frameworks

AI as orchestration layer over specialized models - Periodic's core architecture: general language models act as a copilot/orchestrator that ingests literature and experimental data and directs symmetry-aware, fine-tuned specialized neural nets (used as tools or reward functions) rather than doing all modeling itself.
Closed-loop experimentation - Fedus argues data alone is insufficient; the power comes from an active loop where experimental results are checked for aberrations and consistency with simulation/literature, which then drives the next round of experiments.
Scaling laws applied to physical science - Fedus draws a direct analogy to machine learning's shift from a handful of GPUs and small collaborations to industrialized research with hundreds of researchers and millions of GPUs, arguing physical science and engineering will develop similar predictable scaling properties that justify large capital investment.
Recursive self-improvement via cheap, verifiable environments - Fedus compares near-term recursive self-improvement in software engineering to neural architecture search from about a decade ago, driven by cheap and verifiable feedback (unit tests passing/failing) with no domain-expertise gap between AI researchers and the software they improve.

Summary

Elad Gil talks with Liam Fedus, co-founder of Periodic Labs and a former VP of post-training at OpenAI (where he worked on ChatGPT) and early Google Brain researcher, about building what Fedus calls "an AI foundation lab for atoms." Fedus traces his path from undergraduate dark matter research through Google Brain's 2016-2017 "Cambrian era" (distributed training, mixture-of-experts, the Transformer) to OpenAI, where a small team turned GPT-4 into ChatGPT after John Schulman argued against narrower bot ideas like a coding assistant or a meeting-notes bot. He explains that materials-science AI only became viable once reasoning, test-time inference, and reliable tool use matured; the ChatGPT of late 2022 was, in his view, too weak a foundation for Periodic's approach, since science requires connecting AI systems to real experiments rather than reasoning in isolation.

The core of the conversation is Periodic's architecture and data strategy. Rather than training a single monolithic model, Periodic uses general-purpose language models as an orchestration layer that ingests literature, experimental results, and multiple data modalities, and directs specialized, symmetry-aware neural nets built for atomic systems, using them as tools or reward functions. Fedus is emphatic that a static pool of data isn't enough: literature-reported material properties can vary by orders of magnitude, so the real value comes from an active, closed loop where new experimental data is checked for aberrations and consistency, and that in turn drives the next set of experiments. He distinguishes this from AlphaFold's approach, which leaned on decades of existing protein-structure data; materials science lacks that depth, so generalization instead comes from shared physical first principles like quantum mechanics and chemical synthesis rules, which do not automatically transfer across levels of abstraction (a strong quantum-mechanical model doesn't help much with fluid dynamics, for instance).

Fedus and Gil discuss commercialization, drawing an analogy to biotech's choice between a royalty-generating partnership model and an in-house discovery/IP model; Periodic is currently positioning itself as an intelligence layer and control plane for other companies' physical experimentation, while leaving open a more proprietary discovery model for high-value breakthroughs. Gil invokes Neal Stephenson's novel The Diamond Age, with its AI tutor and home matter-compilers, as a reference point for the kind of world Periodic is chasing: agency over atomic rearrangement and synthesis at a pace that could match how fast the digital world has changed, even though "atoms are hard" and physics imposes real limits.

On AGI and recursive self-improvement, Fedus pushes back on treating intelligence as a single scalar - today's systems can be world-class on a narrow task and fail on small perturbations of it, "like a bad high school student." He expects near-term recursive self-improvement specifically in software engineering, comparing it to neural-architecture-search-era automation enabled by cheap, instantly verifiable feedback (unit tests), but says this doesn't automatically extend to domains like biology where there's a real knowledge gap. AI research itself is a slower self-improvement loop because verifying results (convergence, generalization, scaling behavior) requires actual GPU-hours of experimentation, which is the same closed-loop principle underlying Periodic's premise for physical science.

The conversation closes on capital intensity, talent, and robotics. Fedus says compute, not physical lab infrastructure, is Periodic's dominant capital cost despite infrastructure's longer lead times, and notes the stark pay gap between academic scientists (like a Stanford postdoc) and ML engineers, framing commercial AI labs as a way to better compensate and scale the impact of undercompensated scientific talent. On robotics, he says a closed loop doesn't strictly require advanced robots - Periodic currently blends human labor with reliable but narrow, largely off-the-shelf automation - but that improved robotic reliability, especially dexterous humanoids that can work in unstructured labs, would be a major accelerant. Gil, drawing on his own experience building custom liquid-handling robots at Color, agrees that current lab robotics require heavy customization that a more general, reliable robotic platform would eliminate.

Notable Quotes

"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." - Liam Fedus

"It's not just like a pool of data. It's this interactive closed loop system that is so powerful." - Liam Fedus

"One fallacy is thinking about intelligence as a scalar. We've consistently seen these systems have a very odd spikiness." - Liam Fedus

"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." - Liam Fedus

"Many people working in science, particularly in academic setting, are very undercompensated relative to sort of their societal value." - Liam Fedus