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

Summary

• Periodic Labs’ core wager is that AI’s next major value pool depends on closing the loop between models and physical experiments. Fedus argues that science is not “sitting in a room thinking really hard”; systems must “interface with reality.” Reasoning, test-time inference, error correction, and tool use became foundational as models improved after 2022; 2022-era AI was still too weak for Periodic.
• The key data resource is not a static materials corpus but an active stream of grounded experimental feedback. Periodic can leverage “on the order of tens of trillions of tokens” that went into open-source models for a foundational prior, yet published measurements may span many orders of magnitude. An active loop—spotting aberrations, comparing simulations and literature, then driving the next experiments—grounds the system.
• Periodic combines language models as an orchestration layer with fast, symmetry-aware neural networks built for atomic systems. The general model reads literature, analyzes experimental modalities, and directs work; specialized models serve as tools and reward functions. Generalization can be strong within quantum-governed domains, but Fedus cautions that it does not automatically cross into abstractions such as fluid dynamics.
• The initial business is a software intelligence layer for materials and process engineering, with higher-value discovery economics left open. Periodic is “customer zero,” testing systems that inspect data, debug machinery, improve formulations, and control experiments before pursuing broader advanced-manufacturing opportunities. Asked whether it resembles biotech, Fedus said breakthrough materials “might be more akin to a discovery model,” but the company is starting as software.
• Scaling physical science could unlock an “order of magnitude or two” in productivity across semiconductors, aerospace, and energy. Fedus expects improvements in automation to create bottlenecks in intelligence; physical infrastructure has long lead times and calibration risk, although compute remains the primary capital cost. The ambition is to give humanity “agency for atomic rearrangement and synthesis.”
• AGI will not lift every domain simultaneously because intelligence is spiky and closed-loop verification is domain-specific. Software self-improvement is happening “now-ish” because unit tests provide cheap, immediate rewards; AI research has a slower GPU-intensive outer loop, while biology and physical science require their own data-generating loops. General robotics is not required for Periodic, but a reliable dexterous humanoid would be “a huge accelerator.”

Deep dive

1. The path from dark matter to atoms ran through ChatGPT

• Fedus traced his route from physics and direction-sensitive dark-matter research into machine-learning problems in particle reconstruction. His explanation for physics’ talent migration: the discipline produces “very principled, very hard-nosed scientists,” while high-energy physics became bottlenecked on new apparatus for pushing the next energy frontier and AI offered extraordinary leverage.

• At Google Brain in 2016–2017, Fedus experienced a “Cambrian era” when small teams and a handful of GPUs advanced distributed-training strategies, mixture-of-experts systems, transformers, and sparsity. The field had more “diversity and entropy” before frontier research became industrialized.

• At OpenAI, the immediate problem was productizing a pretrained GPT-4 that had only rough post-training. The team considered writing, coding, and meeting bots, but John Schulman argued for remaining general: “Let’s do chatbot.” Fedus later returned to physics because science must conduct experiments and “learn from them”; 2022-era AI was still too weak for Periodic.

2. Experimental loops turn internet priors into physical truth

• Guo raised the core data question: language models inherit the internet, while materials domains lack comparably accessible corpora. Fedus said Periodic spends “zero effort” improving coding models and benefits from Codex and Claude Code, while using a combination of open- and closed-source models. It leverages “on the order of tens of trillions of tokens” that went into open-source models as a foundational prior.

• That prior brings papers and broad world knowledge, but not reliable physical truth. One Periodic engineer found a material property whose reported literature values spanned many orders of magnitude; as Gil observed, a model trained on them could reproduce the distribution while getting “no closer to a ground truth.”

• The high-value mechanism is therefore “not just a pool of data.” Experimental results are inspected for aberrations, patterns, and agreement with simulations and literature, and those findings drive the next experiments. Fedus said Periodic sees its greatest acceleration where data is abundant and sample efficiency is high.

• Gil invoked AlphaFold’s decades of accumulated structures to ask whether every materials domain must be bootstrapped separately. Fedus said quantum-mechanical systems can generalize across related objects, but the boundary matters: accurate quantum modeling offers little automatic help with “fluid dynamics” at another level of abstraction.

3. Language models orchestrate specialists rather than model every atom alone

• Periodic retains language models because they provide a natural interface, but treats them as an “orchestration layer.” The model can ingest literature and experimental data across modalities, direct experiments, and coordinate specialized tools rather than performing every calculation itself.

• Fedus described dedicated neural networks for atomic systems with symmetry awareness, lower latency, and domain-specific fine-tuning. Those networks can function as tools or reward functions inside the larger system—an architecture Gil noted is also emerging in areas such as customer support.

• The practical implication is a system assembled around complementary strengths: broad models supply general reasoning and interface capabilities, while atomic models supply speed and physical structure. Periodic concentrates its machine-learning work only where the existing frontier is “not sufficiently good for us.”

4. Periodic starts as software but leaves discovery economics open

• Periodic has treated itself as “customer zero,” working closely with scientists to redesign how research is done. Fedus sees similar workflows across advanced manufacturing: engineers interrogate data, identify aberrations, debug machinery, and search for better formulations while materials or process engineering bottlenecks constrain the business.

• Gil’s biotech analogy sharpened the commercialization choice: partner with incumbents and share discovery value, or own the resulting materials. Fedus positioned Periodic first as the “intelligence layer”—a system of record and control plane for experiments—but conceded that exceptional breakthroughs “might be more akin to a discovery model.”

• The longer-range vision is AI that moves beyond essays and software to “literally generating matter.” Fedus named semiconductors, aerospace, and energy, arguing that atoms impose real limits but do not preclude “an order of magnitude or two” of acceleration. If the physical world kept pace with even a fraction of digital change, “life will just feel dramatically different.”

5. Closed-loop verification determines where self-improvement arrives

• Fedus rejected treating intelligence as a scalar: models remain oddly “spiky,” sometimes world-class in one mathematical domain yet fragile to nearby perturbations—“like a bad high school student.” Competence in repository-scale software therefore does not imply understanding biology.

• Software self-improvement is happening “now-ish” because code supplies abundant data and cheap, verifiable feedback: a few CPUs can immediately show unit tests moving from failure to success. AI research follows on a slower outer loop because experiments require GPUs, hours of runtime, convergence checks, scaling analysis, and generalization evaluation.

• Fedus’s most exciting operating model is multidisciplinary: physicists and chemists work closely with AI researchers and engineers. Periodic wants to bring frontier AI’s scaling mindset—larger experiment sets enabled by intelligence plus automation—to physical science, though scientists may themselves become the bottleneck in interpreting the resulting data.

• Physical science needs its own loop of experiments and engineering. Asked whether advanced robotics is required, Fedus answered, “No, but it’s a huge accelerator”: Periodic already combines people with reliable autonomous components and simple, off-the-shelf robotics. Gil’s example from Color—ML-monitored cameras, liquid-handling adjustments, and 3D-printed vibration-reducing parts—underscored why a dependable dexterous humanoid could be a major accelerator for spinning up new labs.