AI for Atoms: How Periodic Labs is Revolutionizing Materials Engineering with Co-Founder Liam Fedus
In a Nutshell
Liam Fedus, ChatGPT co-creator and OpenAI alum, co-founded Periodic Labs to apply AI foundation models to materials science, using language models as orchestration layers for analyzing literature, directing experiments, and leveraging specialized neural nets for quantum mechanics and atom interactions. The approach addresses data scarcity in physical sciences through interactive closed-loop experimentation, sample-efficient generalization from vast code corpora, and hybrid human-robot automation for high-throughput data generation. Vision: accelerate materials discovery by orders of magnitude for semiconductors, aerospace, and energy, enabling atomic-scale agency akin to digital revolutions, with robotics as a key accelerator toward domain-specific AI genius and self-improving systems.
These notes were generated by AI and may contain inaccuracies.
Liam Fedus is a co-creator of ChatGPT, VP of post-training at OpenAI, and previously at Google Brain working on early AI innovations. He discusses Periodic Labs, an AI foundation lab for atoms impacting materials sciences, chemistry, and the physical world.
Liam was a physics major in undergrad, did dark matter research with a directionally sensitive apparatus. Many AI leaders have physics backgrounds: Dario Amodei (Anthropic), Adam Brown (Google), Joshua (Anthropic manager). Physics provides principled, hard-nosed thinking with high leverage in AI. After Higgs discovery, high-energy physicists sought new frontiers as progress bottlenecked on apparatus. String theorists and black hole researchers moved into AI, recreating a Manhattan Project-like effort for intelligence.
In grad school, gravitated to machine learning for particle reconstruction. Switched to computer science, joined Google Brain first-year residency (2016-2017): era of distributed training, mixture of experts, transformer, small collaborations with few GPUs. Google Brain was small, later combined with DeepMind. Focused on architecture, sparsity for efficient model serving at scale, pushing scale toward late 2022.
Joined OpenAI excited about productizing compelling technology. Productionized GPT-4 post-training: ideas included writing bot, coding bot, meeting bot. John Schulman insisted on general chatbot, leading to ChatGPT effort. ChatGPT was the starting gun for AI revolution awareness.
Inevitability of connecting AI to physical world for science acceleration. Science requires experiments interfacing with reality. ChatGPT (late 2022) too weak for Periodic; needed improving models, reasoning, test-time inference, error correction, tool use, coding agents. Data challenge in physical world lacks internet-scale corpus.
Leverage existing models' generalization (tens of trillions tokens from open source like CodeX, Claude Code). No effort on coding models. Strong priors from papers/internet enable sample efficiency in chemical spaces. Literature data unreliable (properties span orders of magnitude). Experimental data provides grounding.
Interactive closed-loop: analyze data for aberrations, patterns, consistency with simulations/literature to drive next experiments.
Abundance of data accelerates advances. AlphaFold succeeded on decades of protein structures. Generalization across quantum mechanical effects, but not to fluid dynamics. First-principles: chemical synthesis steps, quantum mechanics, atom interactions, van der Waals forces.
Language models as orchestration layer: ingest literature/experimental data, direct experiments, use specialized neural nets (symmetry-aware for atoms, low latency) as tools/reward functions.
Work closely with scientists, Periodic as customer zero. Opportunities in industries bottlenecked by materials/process engineering. Universal for data questions, aberrations, debugging, formulations. Position as intelligence layer, system of record/control plane. Some breakthroughs akin to biotech discovery model, but start as software business.
Reference to Neal Stephenson's The Diamond Age: AI tutors, matter compilers, 3D printers, nanobots. Vision: AI generating matter for semiconductors, aerospace, energy; accelerate physical development like digital realm. Atoms hard, but order-of-magnitude speedup possible via data sense-making.
Grant humanity agency for atomic rearrangement/synthesis; physical world keeping pace with digital changes life dramatically. Materials equivalent of agricultural revolution, removing bottlenecks.
Multidisciplinary iteration: physicists, chemists, top AI researchers, engineers. Seeing fields transform. Analogy to ML evolution: few GPUs to industrialized scaling laws, massive capital. Physical sciences need scaling mindset, larger experiment sets via intelligent automation.
Capital-intensive, primarily compute (GPUs expensive). Physical infrastructure lower cost but long lead times, calibration challenges. Scientists undercompensated vs. societal value (e.g., Stanford postdoc vs. ML engineer).
Built interdisciplinary team of world-class scientists.
Decomposed into bits (mid/pre-training AI, infrastructure) and atoms (control/system engineering, project engineering, roles).
Intelligence not a scalar; systems show spikiness (world-class in math, degrades with perturbations like bad high school student). Possible to architect genius in one domain, weak elsewhere.
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.
Fields like software engineering and AI research can generalize in non-intuitive ways. Recursive self-improvement is akin to neural architecture search from about 10 years ago. In software engineering, systems have become impressive due to huge amounts of data, cheap verifiable environments like unit tests that go from failing to passing with just a few CPUs—essentially instantaneous. There's no domain expertise gap between AI researchers and software engineers. This will become a larger contributor to the next generation of systems.
When will it flip to fully machine self-improvement versus human-directed or needing human intervention? Two years? Five? Ten?
There's a domain caveat. Software engineering self-improvement will enable systems to write complete repositories, identify bugs, refactor code—but it won't suddenly understand biology due to the domain gap in knowledge. Software strategies differ from scientific or engineering strategies; software is highly verifiable, not decision-making under uncertainty to the same degree. In software, it's happening nowish.
AI research has a slower outer loop: experiments check scaling properties, model convergence, generalization—requiring GPUs and many hours. These evals are used today on existing models, providing a utility function and feedback loop for self-learning.
Connection to the physical world is critical. Software and AI research are trained in closed loops against their domains. Periodic's premise is closed loops for doing science and engineering. These domains will follow with some delay as foundational technology.
Do you need sufficiently good robotic systems for closed loops at Periodic, like PI or Skills, to hit escape velocity?
No, but it's a huge accelerator.
Periodic's goal is to generate high quantity, high-quality, diverse data. Automation assists that. Currently, they employ people and have autonomous parts that are very reliable. A dexterous humanoid that could wander into an unstructured lab, make sense, and follow instructions reliably would be a huge accelerator. Current physical automation requires careful design and is slow, but improvements in robotics will accelerate it. Hybrid systems already produce huge amounts of reliable data.
Speaker owned Color company, built custom liquid handling robotic systems. Bought robots but adjusted them dramatically: added cameras using ML to monitor and adjust, 3D printed parts to decrease vibrations for small liquid volumes. Firmware was awful, writing against it painful—versus a modern system that just works. For high-throughput experiments, need systems for liquid handling, titration, etc.
Periodic uses off-the-shelf robotics: simple, commoditized, not innovating much there. As general robotic systems hit reliability thresholds, it will massively accelerate spinning up new labs.
Over the past decade since work at Google—birth of transformer model, ChatGPT—what's most exciting outside Periodic?
Of course, robotics.
Excited about AI-physical world interface, approaching via science and engineering, needing data for advances. Agency and control of physical world via robotics will be transformative. Interface layers are a massive opportunity: far more people build the physical world than software engineers, with labor shortages everywhere. It will be a very interesting decade.
Thank you for joining. Find us on Twitter at No Prior Pod. Subscribe to YouTube. Follow on Apple Podcasts, Spotify, or wherever. New episode weekly. Sign up for emails or transcripts at no-priors.com.
Keep No Priors: AI, Machine Learning, Tech, & Startups in your library
Save the videos and channels worth coming back to, and find them again in one place.
