No Priors: Artificial Intelligence | Technology | Startups · Conviction

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

April 3, 2026·29 min·2 clips

John Schulman insisted on a chatbot over a meeting bot, sparking the creation of ChatGPT.

1. No Priors hosts an interview with Liam Fedus, one of the co-creators of ChatGPT and former VP of Post-Training at OpenAI, about his new company Periodic Labs, which applies AI to materials science and physical-world engineering. 2. Fedus is a physics undergraduate who worked on dark matter research, moved to machine learning at Google Brain in 2016-2017 during the creation of the transformer and mixture-of-experts architectures, then joined OpenAI in late 2022 to productionize GPT-4 into what became ChatGPT. 3. The episode's core thesis is that the next frontier for AI acceleration is the physical world — specifically, connecting AI systems to experimental closed loops in chemistry and materials science rather than operating purely on text and code. 4. Fedus notes that many AI researchers have physics backgrounds — citing Dario Amodei at Anthropic, Adam Brown at Google, and his own manager Joshua — arguing that physicists were drawn to AI after the Higgs boson discovery made high-energy physics bottlenecked by the cost of new accelerators. 5. He explains that ChatGPT was chosen over a writing bot, coding bot, or meeting bot because John Schulman at OpenAI was opinionated that the product should remain general — a decision Fedus credits as the 'starting gun' of the current AI revolution. 6. After building ChatGPT, Fedus became convinced that language AI is missing a critical component: the feedback loop of physical reality, because 'science ultimately isn't sitting in a room thinking really hard — you have to conduct experiments.' 7. Fedus states that ChatGPT-era technology in late 2022 was 'far too weak' to have enabled Periodic Labs, and that the advances in reasoning, test-time inference, reliable tool use, and coding agents over the subsequent years were foundational prerequisites. 8. On data, Fedus notes that a Periodic engineer found a single reported material property spanning 'many orders of magnitude' in literature, illustrating that an ML system trained on text alone can only model a distribution — experimental data provides a ground truth that text cannot. 9. Periodic's approach leverages the roughly 'tens of trillions of tokens' in open-source models as a prior on the world, then targets specific chemical spaces where that prior is insufficient, achieving high sample efficiency because the model is not starting from a randomly initialized neural net. 10. The company's architecture uses language models as an orchestration layer — ingesting literature, experimental data, and multiple modalities — while directing specialized symmetry-aware neural nets built for atomic systems as tools and reward functions. 11. Periodic uses Claude Code and Codex for software engineering and says it spends 'zero effort' on improving coding models, directing all ML resources toward domains where frontier models are not yet sufficient. 12. Fedus describes the closed experimental loop as the key innovation: results are scanned for aberrations, cross-referenced against simulation and literature, and used to design the next set of experiments — making it an active learning system rather than a static data pool. 13. On commercialization, Fedus frames Periodic as an 'intelligence layer' for companies bottlenecked by materials and process engineering, with Periodic itself as 'customer zero' — testing the system's ability to transform how science is done before deploying it externally. 14. He draws an analogy to the biotech model — partnering to discover and taking royalties versus developing proprietary materials — while noting the company currently thinks of itself primarily as a software business. 15. Fedus predicts that physical sciences will follow the same scaling-law trajectory as AI research: once scaling properties are established, predictability will attract large capital, industrialize the research process, and shift the bottleneck from intelligence to data throughput. 16. He argues that scientists working at Periodic are experiencing the same shift that early ML researchers experienced when going from 'a few GPUs and a few people' to industrialized training — seeing their fields fundamentally change in real time. 17. On AGI, Fedus challenges the framing of intelligence as a scalar, noting that current AI systems have 'odd spikiness' — world-class in some math domains but degradable by small perturbations, and not generalizing across adjacent fields like biology even when those fields seem close. 18. He argues that software-engineering self-improvement is happening 'now-ish' because the feedback loop (unit tests) is instantaneous and cheap, while materials-science self-improvement requires a slower outer loop — hours of experiments, GPU-intensive convergence checks — but will follow the same path. 19. Fedus is most excited outside Periodic about robotics, predicting that dexterous humanoid robots that can operate in unstructured labs will massively accelerate Periodic's data generation, and that AI-plus-robotics will transform labor-shortage-constrained physical industries broadly. 20. Listeners working in AI research, materials science, deep tech investment, or advanced manufacturing who want a technical but accessible account of where AI for physical sciences is heading will find this episode directly relevant.

As heard by us

A clear look at AI moving from language into labs, experiments, and physical systems.

Periodic Labs is framed as a push to move AI beyond the screen and into atoms, experiments, and the physical world. The clearest section explains the orchestration layer: literature and experimental data come in, specialized neural nets can serve as tools or reward functions,…

Read the full review in PlayNext →

Why you'd press play

You want a conversation with Liam Fettus, a ChatGPT co-creator, about building an AI foundation lab for atoms.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from No Priors: Artificial Intelligence | Technology | Startups