Stanford OnlineStanford CS153 Frontier Systems | Teaching AI to Touch Atoms
Stanford Online
Introducing Periodic Labs 0:10
The guests are Liam Fedus and Dogu Kilic, co-founders of Periodic Labs, a company about eleven months old at the time of this talk. Fedus helped set up the post-training team at OpenAI and was involved in the ChatGPT moment, while Kilic worked on Gemini at DeepMind. Their original plan was to spend a first year building purely computational, in-silico material design before switching on a large automated lab. That plan changed quickly. Instead they built small, semi-manual, semi-autonomous labs right away, which let them direct the research and close the feedback loop faster than expected. Both founders say the results of pairing AI with real physical experiments, not just code, have been more dramatic than they anticipated, producing better materials synthesis and computational progress.
How the lab actually works 3:01
Periodic Labs runs a roughly 40,000 square foot facility in Menlo Park, staffed half by machine learning people from places like OpenAI and DeepMind and half by physicists and chemists from schools like Stanford, MIT, and Caltech. An AI system predicts candidate materials, robots synthesize them into forms like powders, and other machines verify whether the materials actually have the predicted properties, with that verification feeding back into training. Much of the AI's real value turns out to be in mundane tasks: mixing powders correctly, catching sample mixups, or making sense of X-ray diffraction patterns used to identify what was actually made. Scientific progress, they say, is mostly a chain of these small, unglamorous steps.
Why semiconductors and superconductivity 6:03
Following advice to chase a specific, state-of-the-art evaluation before generalizing, the team chose superconductors and semiconductors as their proving ground, since both are governed by similar atom-electron physics and thermodynamics. They reject the idea that AGI will be a magical, fully general capability, betting instead on deep specialization in how atoms and electrons interact. Superconductors carry current with zero resistance, unlike ordinary materials such as copper, where resistance turns energy into heat, though the founders decline to put a firm percentage on how much energy chips lose this way. This physics was first observed in 1910 after helium was liquefied, yet superconductors remain confined to niche uses like MRI machines, despite potential in fusion energy, quantum computers, maglev, and lossless transmission. Their AI system is named Onnes, after the physicist who discovered superconductivity and pioneered industrial-scale scientific research.
Combining models and active learning 15:32
A questioner asks how to combine different models to push scientific discovery. The speaker notes that physicists moving into machine learning often fall for Bayesian optimization because of its theoretical elegance, but in practice it struggles because uncertainty estimates generalize even worse than point predictions. Active learning works better, much like how Waymo had to keep training on the specific situations it handled poorly rather than uniform data. Since their models understand some science well and other parts not at all, they run active learning daily, using new experiments to push into the unknown areas and expand what the models generalize to.
Defining new material discovery 20:31
Discovery can mean several things. It might mean finding a genuinely new crystal structure, something rare among the roughly 250,000 crystals already catalogued in the ICSD database. It can mean finding a material with new properties, such as ambient pressure superconductivity beyond the current 133 Kelvin benchmark. It can also mean finding better synthesis recipes for known materials, making them cheaper or more optimized to produce. The underlying goal is expanding human control over how atoms are arranged.
Origin story of the company 24:00
One founder describes growing frustrated watching AI teams claim physical-world progress that never matched their marketing, then spending time at a Stanford applied physics group benchmarking frontier language models on condensed matter physics, only to find the results were terrible. That paper led to Jason Quan at OpenAI, who connected them with people who had just left DeepMind, and the founders say they went from first meeting to signed term sheet in about four or five days, driven by shared conviction that materials remain civilization's real bottleneck.
Tools, limits, and catalysis 27:31
On methods, density functional theory is strong for ground-state properties like formation enthalpy but weak for band gaps or excited states. Catalysis remains especially hard because it depends on messy atomic-scale details like surface steps and individual defects that are difficult to model, so it stays a more empirical challenge. Machine learning has transformed force fields, building on old empirical approximation ideas, including an early Einstein paper on surface-molecule interactions, and the company has built new graph neural network architectures for this, though full catalysis modeling still eludes them.
Science as an endless frontier 29:01
The chicken-and-egg problem of science, understanding well only what has already been discovered, applies to any civilization doing research, which is why iterative approaches like active learning matter. Large shared datasets like Meta's UMA, with about 100 million density functional theory calculations, help the way ImageNet helped vision, but no dataset can ever be fully comprehensive, because if it were, it would just become settled textbook knowledge. Unlike automating accounting, which has a finite scope, automating science has no end, which is part of what excites the founders about the field.
Sample efficiency in physical experiments 30:31
You cannot scale physical experiments the way frontier labs scale rollouts for a language model, since you can't just run a million trials in the real world the way you might run a million rollouts on a math problem. Because of that limit, Periodic focuses heavily on sample efficiency, especially through model based reinforcement learning, trying to squeeze the most learning out of a limited amount of real data.
Intentional synthesis and messy data 32:00
One open problem is learning to synthesize brand new materials on purpose rather than through trial and error, since labs are good at making things similar to what already exists but struggle with truly novel materials. Making synthesis intentional matters because it's how new technology actually gets built. Related to this is the challenge of making sense of noisy, inconsistent data pulled from across the internet, and building systems that can flag which evidence is reliable. Another gap is automating the analysis of lab instrument readings, which are often not directly interpretable, using an LLM equipped with the same tools human scientists use.
Hypotheses, world models, and staying motivated 34:31
A deeper frontier problem is teaching a model to generate a genuinely good or interesting hypothesis, something much harder to grade than checking if code compiles. Models often hold relevant knowledge in their weights without surfacing it at the right moment, so real progress means getting them to combine knowledge the way a human scientist would, including building world models that use tools to predict future outcomes from a timestamped snapshot of known data. Motivation, the team says, comes from tracking steady technical progress rather than waiting for one big win like a room temperature superconductor, and from a sense of responsibility to steer powerful AI technology toward something positive. Barriers to physical world experimentation are expected to fall as AI and robotics improve, though some domains, like particle accelerators, remain far more capital intensive than others, such as finance or camera based experiments, where the bar to apply AI usefully is still fairly low.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
