Periodic Labs Deploys Wet-Lab AI to Accelerate Superconductor Hunt
Stanford Online
Large language models cannot deduce novel physics from computational data alone, requiring closed physical laboratory loops to supply the empirical feedback digital training lacks.
AI-assisted summary of Stanford CS153 Frontier Systems | Teaching AI to Touch Atoms — original source. Check the original for full context.
Source published:
Key ideas
- Periodic Labs cofounders abandoned their original roadmap of spending a full year purely in silico, finding that launching early semi-manual wet labs immediately was critical for guiding hardware selection and grounding model predictions in physical reality. (1:20)
- Sample efficiency in model-based reinforcement learning has emerged as the critical bottleneck for materials AI, because physical wet-lab synthesis cycles cannot be arbitrarily scaled up 100x the way digital or software rollouts can. (30:33)
- The lab bounds its scientific focus strictly around the few-electronvolt (eV) energy regime—the foundation of chemical bonding and Moore's Law—bypassing high-energy particle physics and organic polymers to concentrate on semiconductor and superconductor bottlenecks. (9:06)
- With the Inorganic Crystal Structure Database containing only roughly 250,000 known inorganic materials and frontier labs heavily focused on pure software, the founders argue the competitive barrier to achieving meaningful impact with physical AI is surprisingly low. (16:48)
Detailed notes
## Background Published on YouTube by Stanford Online on September 29, 2026, this Stanford CS153 Frontier Systems lecture features Periodic Labs co-founders Liam Fedus and Dogus Cubuk. According to the page description, the talk covers their eleven-month-old company's mission to merge frontier AI with autonomous physical laboratories in Menlo Park to accelerate the discovery of materials such as high-temperature superconductors.
## TL;DR The founders describe shifting from purely digital AI scaling at OpenAI and DeepMind to operating an end-to-end automated materials loop that unites machine learning with robotic chemistry and physical characterization. They explain why in-silico simulations alone cannot replace empirical contact with reality, detailing how their AI agent, Onnes, navigates noisy experimental data and partial observability using sample-efficient active learning. Addressing civilizational bottlenecks from Moore's law to energy grids, the speakers show that controlling matter at the atomic level yields vastly higher leverage than software benchmarks alone. Ultimately, they urge students not to fear career displacement from AGI, emphasizing that most physical domains remain untouched by frontier models and offer a uniquely open, impactful frontier.
## Atmosphere & style The lecture combines the intellectual rigor of an advanced university seminar with the candid, grounded clarity of founders working directly on physical hardware and wet labs. The dynamic between the host and speakers is conversational and pragmatic, puncturing Silicon Valley marketing hype about AGI while treating students' career anxieties with thoughtful empathy. It feels both technically demanding and deeply motivating, transitioning from dense discussions of quantum mechanics and active learning to an inspiring call to direct machine learning toward tangible civilizational impact.
## Bulletin - [01:20](https://www.youtube.com/watch?v=8cAQdELWYuo&t=80s) Periodic Labs cofounders abandoned their original roadmap of spending a full year purely in silico, finding that launching early semi-manual wet labs immediately was critical for guiding hardware selection and grounding model predictions in physical reality. - [30:33](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1833s) Sample efficiency in model-based reinforcement learning has emerged as the critical bottleneck for materials AI, because physical wet-lab synthesis cycles cannot be arbitrarily scaled up 100x the way digital or software rollouts can. - [09:06](https://www.youtube.com/watch?v=8cAQdELWYuo&t=546s) The lab bounds its scientific focus strictly around the few-electronvolt (eV) energy regime—the foundation of chemical bonding and Moore's Law—bypassing high-energy particle physics and organic polymers to concentrate on semiconductor and superconductor bottlenecks. - [16:48](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1008s) With the Inorganic Crystal Structure Database containing only roughly 250,000 known inorganic materials and frontier labs heavily focused on pure software, the founders argue the competitive barrier to achieving meaningful impact with physical AI is surprisingly low.
## Key takeaways - [01:20](https://www.youtube.com/watch?v=8cAQdELWYuo&t=80s) Periodic Labs abandoned a planned one-year pure in-silico simulation phase after realizing that small, semi-manual physical experiments were required immediately to steer the research program and determine what hardware to scale. - [03:00](https://www.youtube.com/watch?v=8cAQdELWYuo&t=180s) The company's 40,000-square-foot facility couples frontier machine learning researchers directly with physical scientists, deploying their AI agent Onnes to automate both high-level design and mundane laboratory tasks like recipe formulation, powder milling, and XRD pattern interpretation. - [05:46](https://www.youtube.com/watch?v=8cAQdELWYuo&t=346s) Unlike pure software and mathematics where verification is deterministic and binary, physical science requires models to make decisions under genuine uncertainty, partial observability, and noisy instrumentation. - [09:06](https://www.youtube.com/watch?v=8cAQdELWYuo&t=546s) Periodic scopes its work to the few-electronvolt regime governing electron-phonon coupling and solid-state thermodynamics, explicitly avoiding both subatomic particle physics and organic polymer chemistry. - [10:30](https://www.youtube.com/watch?v=8cAQdELWYuo&t=630s) High-temperature superconductivity represents a transformative civilizational bottleneck because eliminating resistive loss could transform data center power consumption, fusion energy containment, and quantum computing interconnects. - [14:14](https://www.youtube.com/watch?v=8cAQdELWYuo&t=854s) Language models cannot discover novel physical materials purely through textbook pretraining; models must be tethered to continuous empirical verification loops inside real laboratories to achieve physical grounding. - [16:48](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1008s) Active learning outperforms classical Bayesian optimization in frontier chemistry because neural network uncertainty estimates compound error aggressively out-of-distribution, making directed sampling of unexplained experimental anomalies essential. - [20:31](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1231s) Materials discovery spans three distinct levels: identifying unprecedented crystal prototypes, achieving record bulk physical properties (such as raising the ambient-pressure superconducting limit from 133 K to 160 K), and uncovering viable synthesis recipes. - [25:02](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1502s) Student anxiety over AGI rendering technical careers obsolete is misplaced because LLMs have fundamentally altered very few physical domains, making the current era an outsized window of opportunity for specialists working on physical atoms. - [27:24](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1644s) While Density Functional Theory reliably computes ground-state formation enthalpies, it struggles with band gaps, excited states, and surface defects in heterogeneous catalysis, highlighting why automating the scientific discovery loop has no fixed ceiling. - [30:33](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1833s) Physical synthesis rollouts carry rigid real-world cycle times and cannot be arbitrarily scaled up like digital code generation, making sample-efficient, model-based reinforcement learning the core ML objective. - [31:35](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1895s) Intentional synthesis of genuinely novel materials remains largely unguided trial and error, but constructing denser thermodynamic convex hulls at zero Kelvin dramatically improves model predictions of whether a candidate compound can be physically synthesized. - [33:28](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2008s) Pairing LLMs with domain-specific software tools unlocks automated interpretation of complex characterization data, while benchmarking against historical science timestamps allows researchers to test an AI's hypothesis generation capabilities. - [38:21](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2301s) Because mainstream frontier AI labs remain fixated on abstract reasoning, mathematics, and pure code, the competitive barrier to entry for AI applied to physical instrumentation remains remarkably low. - [41:27](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2487s) Software benchmarks operate in closed worlds with complete context and deterministic compilers, whereas physical wet labs exhibit irreducible partial observability, making learning under unrecorded environmental variables machine learning's true frontier.
## Detailed walkthrough ### [00:10](https://www.youtube.com/watch?v=8cAQdELWYuo&t=10s) Welcome and Founding Periodic Labs
The lecture opens with an enthusiastic welcome for Liam Fedus and Dogus Cubuk, the co-founders of Periodic Labs. Gauging the room, the host asks how many students have heard of the company, eliciting an encouraging show of hands that prompts the host to joke, "The kids are paying attention. Not bad."
Both founders bring distinguished lineages from frontier AI labs. Liam Fedus helped establish the post-training team at OpenAI and was there during the launch of ChatGPT, helping shape the continuous post-training loop that defines modern LLM refinement. Dogus Cubuk previously worked at DeepMind on the GNoME (Graph Networks for Materials Exploration) project. Periodic Labs emerged out of those parallel lineages roughly eleven months prior, starting around June. The host asks what priors they have had to update most aggressively since founding the company compared to their original expectations.
[01:20](https://www.youtube.com/watch?v=8cAQdELWYuo&t=80s) Updating Priors: In Silico Illusions vs. Physical Reality
Liam points to a fundamental shift in how they conceptualized building a lab. Coming from the spirit of AI and machine learning scaling laws, they initially expected that new opportunities in science would arrive strictly at scale, leading to what he calls a "cartoon version" of the roadmap. Under that initial model, because standing up a high-throughput autonomous physical facility takes time, the company planned to spend its entire first year operating purely *in silico*—designing novel materials computationally—before walking up to a finished high-throughput lab on day 365, flipping a switch, and starting real-world science.
Instead, they abandoned that plan early on. Setting up smaller, semi-manual, semi-autonomous labs immediately allowed the team to steer the research program, determine what hardware and equipment actually made sense to scale, and close the feedback loop between model predictions and experimental reality as fast as possible. Dogus adds his own updated prior: while they were excited to see large models manipulate real atoms rather than just write Python scripts, the magnitude of the results surprised them. "I think the results are crazier than we expected," Dogus explains, noting that automated intelligence affecting physical matter has dramatically outpaced their expectations in both computational speed and materials synthesis quality.
[03:00](https://www.youtube.com/watch?v=8cAQdELWYuo&t=180s) Inside the Menlo Park Lab: Architecture of an Autonomous Materials Loop
To clarify how the facility works, the host lays out a baseline walkthrough of their Menlo Park operations, red-teamed live by the founders:
```mermaid flowchart TD accTitle: Periodic Labs Material Discovery Loop accDescr: End-to-end feedback loop from AI model prediction through automated synthesis, characterization, and model retraining.
subgraph Computation["In Silico Reasoning & Design"] A["Frontier Model (Onnes)<br/>Pre-, mid- & post-trained"] -->|"Predicts candidates"| B["Material Formulation<br/>& Synthesis Recipe"] end
subgraph Lab["Robotic Synthesis & Physical Handling"] B --> C["Robotic Mixing & Powder Prep"] C --> D["Synthesis & Oven Processing"] D --> E["Handling Mundane Errors<br/>(sample mix-ups, impurities)"] end
subgraph Verification["Characterization & Evaluation"] E --> F["X-ray Diffraction (XRD)<br/>Structural Characterization"] F --> G["Physical Property Verification<br/>(Superconductivity checks)"] end
G -->|"Empirical ground truth feedback"| A ```
Operating out of a 40,000-square-foot facility, Periodic divides its physical space down the middle. Half the square footage hosts the interdisciplinary team—machine learning researchers from institutions like OpenAI and DeepMind sitting directly alongside physicists and chemists from Stanford, MIT, and Caltech. The other half houses the automated laboratory containing synthesis robots and analytical instruments.
Dogus and Liam emphasize that the AI's role is not limited to grandiose high-level predictions about high-temperature superconductors. In practice, scientific progress is built out of hundreds of connected, mundane tasks: - Formulating recipes and determining how to blend and mill precursor powders correctly. - Detecting and flagging sample mix-ups on the floor before bad measurements corrupt results. - Catching and eliminating chemical impurities early in the preparation stages. - Characterizing materials after they come out of the furnace, because "materials don't come out of the oven with labels on them." - Interpreting ambiguous analytical signals, such as firing X-rays at samples and parsing the resulting X-ray diffraction (XRD) patterns to confirm what crystal structure was actually synthesized.
[05:46](https://www.youtube.com/watch?v=8cAQdELWYuo&t=346s) Navigating Scientific Uncertainty and Atomistic Physics
The host draws a connection to a previous lecture by Rob Riachi from DeepMind's pre-training team, who advised students working on their "one-person frontier lab" final projects to ground their work in specific domain benchmarks before generalizing. When asked why Periodic chose materials—and specifically semiconductors and superconductors—as their core target domain, Liam pushes back slightly on the standard machine learning benchmark paradigm. In pure math or software, verification is clean and binary (for instance, evaluating whether `2 + 2 = 4`). Science, by contrast, is defined by decision-making under genuine uncertainty: instruments provide noisy, partial evidence, adjacent experiments offer conflicting clues, and the ground truth is fundamentally murkier. Consequently, teams must spend substantial time learning the nuances and shape of the physical problem before blindly hill-climbing an arbitrary benchmark.
Liam and Dogus explain that their conviction comes from deep atomistic physics: - Every frontier AI lab is desperately consuming hardware, yet advanced semiconductor scaling and memory nodes face acute materials engineering bottlenecks at atomic interfaces. - The physics governing semiconductors and superconductivity overlap heavily; both are dictated by interactions between atoms, electrons, and lattice vibrations (phonons). - While conventional superconductivity is well understood as interactions between phonons and electrons, high-temperature superconductivity remains an open theoretical problem likely tied to complex couplings between phonons, electrons, and magnetic properties. - Rejecting the assumption that AGI will spontaneously generalize to everything without grounding, the team believes that deliberately directing intelligence toward atomic and electronic interactions provides the highest leverage for human technology.
[09:06](https://www.youtube.com/watch?v=8cAQdELWYuo&t=546s) From Fundamental Physics to Moore's Law
Following Dogus's technical breakdown of phonons and electrons, the host asks for a show of hands to see how many physics majors are in the audience; only about five hands go up.
Dogus contextualizes why they bound their research scope the way they do: - Periodic deliberately avoids studying polymers or high-molecular-weight chemistry, meaning their models do not currently generalize to organic macromolecules. - They also bypass the extreme, ultra-high energy limits of physics like string theory or subatomic quarks, which do not tangibly alter everyday human technology. - Instead, they anchor their efforts at the energy scale that directly shapes human existence—typically around a few electronvolts (eV). - This regime governs thermodynamics, chemical bonds, and electron-lattice mechanics—the exact rules underlying Moore's law, which Dogus highlights as arguably the most impressive technological achievement in human history.
[10:30](https://www.youtube.com/watch?v=8cAQdELWYuo&t=630s) Quantum Mechanics, Resistive Loss, and the Superconductor Frontier
At the energy scales governing chemistry, biology, batteries, and semiconductors, the physics reduces cleanly to non-relativistic quantum mechanics. The researchers explain that at this level, one does not need quarks or string theory; it is strictly an introductory solid-state physics regime consisting of ions and electrons interacting with one another. To ground this for the systems-focused class, an analogy is proposed: viewing electrical conductors as pipes through which matter and charge flow. In standard conductors like copper, electron-ion collisions generate scattering, causing resistance that dissipates substantial amounts of energy as heat. In high-density settings like data center computing clusters, these resistive bottlenecks mean double-digit percentages of total power are wasted as heat loss that must subsequently be cooled and dissipated.
```mermaid flowchart TD accTitle: The Physical Discovery and Active Learning Loop accDescr: Feedback loop connecting computational AI models, robotic synthesis, characterization, and active learning updates.
subgraph Digital ["In Silico Reasoning"] A["Onnes Agent / LLM"] -->|"Orchestrates & predicts"| B["Specialized Neural Nets<br/>(Synthesis & Stability)"] end
subgraph Physical ["Laboratory Execution"] B -->|"Formulates recipe"| C["Robotic Synthesis & Lab Trials"] C -->|"Physical characterization"| D["Spectroscopy / XRD Analysis"] end
subgraph Feedback ["Discovery & Adaptation"] D -->|"Target match or anomalous peak"| E["Active Learning Data Engine"] E -->|"Refines world model"| A end ```
The historical counterpoint occurred around 1908 to 1910, when the liquefaction of helium enabled the observation that certain materials lose all electrical resistivity, permitting current to flow indefinitely without dissipation. Despite superconductivity being one of modern physics' most striking discoveries, its real-world implementation remains constrained; commercial applications are largely frozen in older paradigms, such as liquid-helium-cooled MRI machines. Achieving modern materials discovery would transform foundational infrastructure:
- **Clean and Advanced Energy:** Enabling high-field magnetic confinement for nuclear fusion reactors and lossless long-distance electrical transmission grids. - **Computing and Transit:** Powering quantum computing substrates, lower-loss interconnects inside semiconductor stacks, and frictionless magnetic levitation (maglev) transport. - **Cross-Industry Generality:** Engineering matter at the atomic layer provides direct leverage across semiconductors, aerospace, automotive systems, and advanced power generation.
---
[14:14](https://www.youtube.com/watch?v=8cAQdELWYuo&t=854s) The Onnes Agent: Grounding AI in Atomic Reality
The conversation shifts to Periodic Labs' AI agent, christened "Onnes" by team researcher Manseesh. The system is named after Heike Kamerlingh Onnes, the Nobel laureate who liquefied helium and discovered superconductivity. Crucially, Onnes is celebrated not just for the discovery itself, but because he was an early pioneer of running scientific discovery at industrial scale—employing dedicated technicians and professional engineers in 1908 to transform laboratory physics into an organized, high-throughput enterprise. The founders view this industrial mindset as vital for accelerating physical science with urgency today.
When asked to contrast Onnes with the digital language systems Liam worked on during his time co-creating ChatGPT at OpenAI, Liam explains that physical discovery demands an entirely different operational foundation:
> "We don't think that you can just start thinking your way to a new room temp superconductor by reading a textbook, shutting it, and thinking super hard. You have to make contact with reality. You have to carry out experiments."
While standard generative models generalize reasonably well across digital code and natural text, their ability to reason about the physical world hits hard limits without empirical data. Predicting whether a hypothetical crystal is chemically stable and whether it can actually be synthesized in a physical crucible requires continuous validation. Rather than focusing resources on general code generation—where conventional frontier model training already excels—Onnes is purpose-built to predict synthesis pathways, evaluate material stability, and interpret physical outcomes generated inside the lab.
---
[16:48](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1008s) Active Learning vs. Bayesian Optimization in Real Laboratories
Opening the session up to student questions, the founders address optimization methodologies and how distinct modeling techniques should be combined. Dorje outlines a persistent trap common among physicists transitioning into machine learning: an instinctive preference for Bayesian optimization due to its theoretical elegance and explicit modeling of uncertainty. However, because all neural networks exhibit generalization errors outside their training distributions, their estimates of their *own* uncertainty compound even more aggressively than their point predictions. Consequently, classical Bayesian optimization frequently underperforms when searching unmapped chemical territory.
In contrast, active learning is indispensable in physical experimentation, even if it developed a poor reputation in academic computer vision: - **Academic Benchmarks vs. Real Frontiers:** Splitting an existing academic benchmark (like ImageNet) uniformly between train and test sets rarely shows benefit from active learning because the entire domain is already uniformly sampled. - **The Autonomous Vehicle Parallel:** Much like training an autonomous driving platform such as Waymo—where the system handles common scenarios well but must actively gather real-world data on edge cases—scientific discovery requires directing collection specifically toward regions where the model has zero predictive confidence. - **Daily Physical Rollouts:** Every daily laboratory batch at Periodic Labs acts as an active learning rollout, pushing into unknown materials spaces to iteratively broaden the system's predictive boundaries.
Liam adds that anomalous experimental findings frequently provide the highest leverage. When automated characterization tools register an unaccounted-for diffraction peak indicating an unidentified phase that does not match existing crystal databases, the anomaly is flagged, isolated, and fed into the active learning loop. When asked to clarify whether Onnes functions as an "AI scientist" or an "AI agent," the founders clarify that it is fundamentally an orchestrating language model equipped with tools—invoking external software as well as specialized, domain-specific neural nets to formulate recipes and test physical hypotheses. Dorje closes by pointing out the scale of the target domain: the Inorganic Crystal Structure Database (ICSD) contains only about 250,000 known inorganic crystals, leaving vast swathes of potential atomic configurations completely unexplored.
[20:31](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1231s) Redefining Discovery and the Conviction Behind Periodic Labs
Dorje begins by laying out what constitutes a "discovery" in materials science. Materials often fall into recognized subsets of crystal prototypes, and finding a fundamentally novel crystal structure that has never been documented before is exceptionally rare. Beyond novel crystal lattices, discovery can also mean achieving unprecedented material properties. In superconductivity, for instance, the ambient-pressure state of the art hovers around 133 Kelvin; finding a material that operates at 160 Kelvin would represent a monumental breakthrough. Equally critical is the discovery of synthesis recipes: even when a chemical compound is already known, uncovering a way to synthesize it more efficiently or with subtly optimized target properties pushes the frontier of human control over atomic arrangements. Dorje acknowledges to the students that materials science can often seem unexciting or intimidatingly opaque from the outside because it fundamentally deals with the complex physics of atoms. Yet, viewed through the lens of civilization gaining deliberate control over matter, it is one of the most exciting endeavors imaginable.
The course host jumps in to redirect the discussion toward how the company came to be, asking what gave the investors and founders their absolute conviction in Periodic Labs. Recalling the diverse roster of tech leaders the class has hosted—from Jensen Huang at the silicon chip level to Scott Nolan at the energy layer—the host explains his growing frustration over how long the tech community had been hyping up AI's application to the physical world without actual results matching the marketing rhetoric. After seeing numerous startups fail to deliver real physical progress while at Andreessen Horowitz (a16z), he spent time following the previous year's class working with an applied physics group on campus to benchmark frontier LLMs on their ability to reason about condensed matter physics.
Drafting a paper intended for NeurIPS, they found that last-generation frontier models performed terribly on real physical reasoning. Seeking rigorous peer review in the scientific tradition, the paper was shared with Jason Kwon, Chief Strategy Officer at OpenAI. Kwon instantly recognized the shared thesis and told him he needed to meet Liam Ferriss and Dorje Chubak, who were concurrently leaving OpenAI and DeepMind with the exact same ambitions. The mutual alignment was so immediate and profound—a true "meeting of the minds"—that going from their first conversation to signing a term sheet and commencing operations took just four or five days.
- **Current benchmark:** Ambient-pressure superconductivity sits at approximately 133 Kelvin; reaching 160 Kelvin would mark a massive milestone. - **Three tiers of discovery:** New crystal prototypes, unprecedented bulk physical properties, and optimized chemical synthesis recipes. - **Foundational bottlenecks:** Civilizational eras are historically categorized by dominant materials (Stone Age, Bronze Age, Copper Age, Silicon Age); physical hardware like flying cars remains blocked by materials breakthroughs. - **Origin speed:** Periodic Labs transitioned from initial introduction via Jason Kwon to a signed term sheet in four to five days.
---
[25:02](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1502s) Physical Atoms vs. Textbooks: Tackling Student Career Anxiety Around AGI
Liam reflects on why they are so driven by this mission, pointing out that much of the physical world has been built on accidental discovery and agonizing trial and error. While the digital landscape transforms exponentially, pulling forward a true "sci-fi future" requires endowing machine learning systems with genuine intelligence about ground-truth physical reality. Models cannot achieve this merely by ingesting textbooks; they must be actively tethered to the physical world.
Dorje then poses a question directly to the Stanford auditorium, asking if students are feeling acute anxiety about their education and future careers in light of rapid advancements in LLMs and AGI. After a noticeable wave of hands goes up in the room, Dorje shares that he encountered the exact same dread when delivering a physics colloquium next door, where physicists asked what the point of their work was if AGI would soon solve all of physics. Dorje considers this existential panic misplaced: outside of self-driving cars (Waymo and Tesla), automated coding, language translation, and quantum mechanical force field approximations, LLMs have fundamentally revolutionized very few physical domains.
When asked by the host what he would pursue if he were currently an undergraduate working on a senior capstone, Dorje recalls using machine learning to design analog circuits during his own undergraduate days and urges students to take a similar route. Whatever domain a student is genuinely passionate about has almost certainly not yet been transformed by LLMs. Because models cannot autonomously revolutionize these fields on their own, the present moment represents a brief, high-leverage "golden era" where human initiative can yield outsized historical impact.
```mermaid flowchart TD accTitle: The Physical AI Frontier accDescr: Contrasting digital AI capabilities with the unaddressed physical world and the need for empirical experimentation.
subgraph Digital["Current LLM Domain (Digital Saturation)"] A["Textbook Pretraining & Code"] --> B["Translation & LLM Reasoning"] B --> C["Rapid Benchmarking Diminishing Impact"] end
subgraph Physical["Physical Reality Frontier (Untouched)"] D["Physical Atoms & Experiments"] E["Materials Synthesis Recipes"] F["Condensed Matter Physics"] end
subgraph Bridge["Periodic Labs Bridge"] G["Active Empirical Loop"] end
Digital -.->|"Fails to touch"| Physical Physical --> G G -->|"Closes feedback gap"| Digital ```
---
[27:24](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1644s) Density Functional Theory, Force Fields, and the Infinite Frontier of Science
A student asks a deeply technical question regarding how Periodic Labs predicts atomistic properties: whether they rely heavily on Density Functional Theory (DFT), and how they tackle complex chemical challenges like heterogeneous catalysis. Dorje clarifies the utility and severe limits of standard computational physics tools. DFT is fundamentally derived around the ground state; its first Hohenberg-Kohn theorem makes it exceptionally reliable for predicting formation enthalpy and ground-state energies. However, DFT struggles considerably with band gaps and excited states. Heterogeneous catalysis poses an even harsher challenge: catalytic reactions depend intimately on atomistic surface morphology—including atomic steps and localized lattice defects—where a single point defect can radically alter the reaction pathway, rendering simulation unreliable. Consequently, the team avoids purely theoretical modeling for catalysis in favor of more empirical methods.
Dorje notes that machine learning has made its most profound scientific mark in replacing classical empirical force fields with Graph Neural Networks (GNNs). Tracing the intellectual history back to Albert Einstein's early papers on empirical interactions between molecules and surfaces, he emphasizes that Periodic Labs combines top-tier LLM capabilities with advanced custom GNN architectures to learn force fields directly.
Addressing the classic chicken-and-egg dilemma of scientific discovery—that models must be trained on the very science they are meant to discover—Dorje argues that this is not unique to machine learning; it is a fundamental challenge of discovery that would confront any scientific enterprise, human or extraterrestrial. Because one can only understand existing science and must push into the unknown, active learning loops are mandatory. While large-scale datasets exist—such as Meta's open-sourced Universal Materials Accelerator (UMA) dataset containing roughly 100 million DFT calculations, serving as an "ImageNet for materials"—there will never be a complete, static dataset for science. If all data could be gathered into a textbook, scientific exploration would cease. Unlike automating finite white-collar workflows like accounting, automating the loop of science has no theoretical ceiling, making it a permanently expansive frontier.
- **DFT core capabilities:** Highly effective for ground-state calculations and formation enthalpy; unreliable for band gaps, excited states, and catalytic surface defects. - **Historical force fields:** Traced from Einstein's early surface-interaction papers to modern Graph Neural Networks (GNNs) approximating quantum mechanics. - **Shared foundational data:** Datasets like Meta's open-source UMA project (~100 million DFT calculations) act as pretraining anchors, analogous to ImageNet in computer vision. - **The endless frontier:** "If you automate accountants, at some point, there's only so much accounting to do. But if you can automate science, there's no end to it."
[30:33](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1833s) Sample Efficiency and the Constraints of Physical Rollouts
Liam explains that sample efficiency is one of the highest technical priorities at Periodic Labs because physical reality imposes hard bottlenecks that do not exist in digital environments. In standard frontier machine learning labs, training policies can leverage virtually unlimited rollouts: if an agent is solving a mathematics problem or writing code, researchers can trivially scale up rollouts from a thousand to a million. In the physical realm of atoms and chemical synthesis, however, it is impossible to expand experimental throughput by a factor of 10 or 100 on a whim.
Because scaling physical experiments carries real hardware constraints and tangible cycle times, Periodic leans heavily into model-based reinforcement learning. This framework enables the system to extract maximum value from sparse data points, allowing their algorithms to establish traction in new chemical compositions and search spaces with comparatively modest datasets.
[31:35](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1895s) Intentional Synthesis and Denser Convex Hulls
Responding to an audience question about the major open problems for academic AI-for-science labs and whether Periodic still relies on the active learning pipelines from their 2022 research, Dorje identifies "intentional synthesis" as one of the most critical challenges in material science. While modern laboratories excel at synthesizing materials that have already been created or closely resemble known compounds, attempting to synthesize an entirely novel material typically degrades into unguided trial and error. Brute-force experimentation dominates current practices, whereas discovering how to intentionally guide chemical synthesis would radically accelerate human technological capability.
Addressing their historical computational pipelines, Dorje confirms that Periodic still actively employs density functional theory (DFT) and active learning. A central conclusion of their earlier work remains true: constructing a larger, more densely sampled zero-Kelvin convex hull directly yields superior predictions for material synthesizability. To maximize accuracy, Periodic incorporates the largest possible convex hulls available, aggregating both proprietary and publicly accessible scientific data.
```mermaid flowchart TD accTitle: Synthesis Feasibility and Active Learning Loop accDescr: Diagram illustrating how dense convex hulls and DFT inform intentional synthesis of materials. subgraph DataEngine ["Hull Construction"] A["Public & Proprietary Data"] --> B["Densely Sampled 0K Convex Hull"] B --> C["DFT Energetics Calculation"] end subgraph SynthesisEngine ["Physical Loop"] C --> D["Synthesizability Prediction"] D --> E["Intentional Synthesis Experiment"] E -->|"Physical Outcome"| A end ```
- **Novel vs. Known Synthesis**: Labs can reliably recreate documented crystals or slight derivatives, but novel chemical architectures lack predictive recipe roadmaps. - **Convex Hull Sampling**: The denser the thermodynamic reference hull at 0 Kelvin, the more reliable the model becomes at forecasting whether a candidate crystal can actually be physically synthesized.
[33:28](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2008s) Characterization Analysis, Latent Weights, and Timestamping the Universe
Liam notes that managing noisy data and decision-making under uncertainty remain major open hurdles across scientific disciplines. Laboratory datasets are inevitably littered with contradictory or flawed findings, making systems that can parse conflicting evidence and flag data consistency extremely valuable. Dorje adds that characterization analysis is ripe for automation: analytical instruments generate raw data streams that require extensive manual analysis by human specialists, but pairing large language models with the native computational tools used by bench scientists enables automated, high-throughput interpretation.
Liam highlights that hypothesis generation represents an unsolved frontier because, unlike evaluating whether code compiles or a math proof holds, evaluating whether a hypothesis is novel and promising lacks a simple ground truth. Furthermore, models often possess relevant scientific knowledge buried in their weights, yet fail to surface or synthesize it appropriately during reasoning tasks unless prompted with direct, narrow queries. To bridge this gap, Liam outlines a world-modeling approach that "timestamps the universe":
> "You can basically timestamp the universe. So, OK, here's the set of known... data we have experimentally, computationally, at some timestamp, $t$, and produce a world-modeling task where you're trying to, then, predict outcomes of future things... The world model is not just in weights. It's not just inference time. But it's using these tools in order to do that."
- **Tool-Augmented Characterization**: Equipping LLMs with human domain software unlocks autonomous interpretation of complex analytical instrument readings. - **Latent Reasoning Gap**: LLMs reliably recall individual factual premises when queried directly, yet routinely fail to assemble those premises into complex physical insights during unconstrained scientific reasoning. - **Historical Benchmarking**: Backtesting agents against historical science timestamps ($t$) allows researchers to benchmark an AI's ability to predict downstream scientific discoveries using computational tools.
[36:35](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2195s) Internal Motivation and Steering AI Toward Tangible Impact
When an audience member asks how the founders stay motivated when experiments stall and hard problems resist solutions, Liam emphasizes that motivation cannot depend solely on an extrinsic endpoint like achieving a room-temperature superconductor. Because that discovery has not been made—and may or may not ultimately prove physically possible—tying daily morale to that single target would lead to burnout. Instead, Liam and Dorje track granular technical milestones across their end-to-end discovery loop, celebrating steady, measurable gains in the individual components of the system.
Dorje expands on the broader ethical and philosophical mission driving the company. He views their work as a moral obligation to direct frontier AI technologies into constructive, real-world utility. Rather than leaving machine learning confined to purely digital or commercial domains, applying it to fundamental physical challenges offers a direct path toward steering technological progress into something unambiguously positive and beneficial for humanity.
[38:21](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2301s) Physical Barriers, Smart Domain Choice, and the "Low Bar" for Physical AI
Addressing the barriers to operating in the physical world and whether those obstacles will recede, Liam remarks that while standing up wet labs and automated hardware is initially difficult, these barriers should decline as underlying AI and robotic systems mature. The labs Periodic is constructing today are likely the most difficult iterations they will ever have to build. Furthermore, founders can manage these barriers through deliberate domain selection: targeting high-energy physics, for example, demands building tens-of-billions-of-dollars particle accelerators staffed by thousands of specialists, whereas inorganic crystal synthesis permits rapid iteration and manageable capital expenditures.
Dorje points out that because mainstream frontier AI labs remain preoccupied with abstract domains like mathematics, combinatorics, and pure code, the competitive barrier to entry for physical AI remains shockingly low. Simply connecting an AI model to an optical camera or a telescope already places a project ahead of most digital-only efforts:
> "The bar is so low in terms of how physical and relevant you have to make AI to make progress. You probably can just be slightly above it... I think the bar is pretty low right now, but it may not stay like that in a few years. I think a lot of young, excited people will realize LLMs can do way more than coding."
- **Infrastructure Maturation**: Early automated lab facilities bear heavy setup costs, but accelerating robotics and AI will lower the cost of deploying future physical setups. - **Domain Leverage**: Inorganic materials provide a sweet spot of rapid physical feedback loops without the multi-billion-dollar infrastructure required by subatomic particle physics.
[40:35](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2435s) Verifiable Domains and Why Periodic Won't Pivot to Finance
An audience member asks about alternative domains that offer the high degree of verifiability required by frontier AI systems, and whether Periodic Labs might ever expand beyond materials. Liam points to quantitative finance and asset pricing—predicting the future price of equities, bonds, or yields—as an example of an environment with clean, immediate empirical verification. However, Liam clarifies that financial prediction is plagued by extreme latency constraints, quickly prompting laughter from the room as he assures the audience that Periodic has no intention of pivoting into Wall Street trading. Dorje interjects to steer the discussion back toward physical verification, setting up the broader role of verifiable rewards in training frontier systems.
[41:27](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2487s) Partial Observability in the Physical World and Closing Remarks
The discussion wraps up with a fundamental contrast between pure software domains and physical experiments: the challenge of partial observability. In traditional AI domains like coding, the environment provides closed-world, fully verifiable context—the model has access to the entire codebase, deterministic compiler outputs, and direct execution feedback. In physical reality, however, that complete picture vanishes. When running real-world experiments in a wet lab, it is practically impossible to record every variable or capture all environmental nuances of what actually occurred during synthesis or measurement.
The speaker highlights this gap as one of the most exciting technical frontiers in machine learning: figuring out how models can learn, adapt, and improve under pervasive uncertainty and incomplete context. This challenge extends far beyond material science and condensed matter physics into virtually every real-world domain, including complex dynamical systems like financial asset pricing. With that open research question on bridging the gap between digital models and messy reality, the session concludes with thanks to the audience.
```mermaid flowchart TD accTitle: Full Context Digital Domains vs Partial Observability in the Real World accDescr: Diagram contrasting digital environments where full context is available with physical environments where context is incomplete.
subgraph Digital["Closed Digital Worlds"] A["Coding & Compilers"] --> B["Full Context Present"] B --> C["Completely Verifiable Outcomes"] end
subgraph Physical["Physical & Real-World Systems"] D["Lab Experiments & Physics"] --> E["Partial Observability"] E --> F["Unrecorded Variables & Uncertainty"] G["Asset Prices & Complex Dynamics"] --> E end
C -.->|"Contrasts with"| F ```
- **Digital vs. Physical Context:** In software engineering benchmarks, models benefit from complete context (the full codebase, runtime diagnostics, and compiler feedback), whereas physical experimental setups can never capture every parameter or subtle environmental factor. - **Broader Applicability:** Tackling sample efficiency and learning under irreducible observation error is not exclusive to physics—it applies broadly across all complex real-world domains, from materials synthesis to modeling asset prices.
## Notable quotes > [00:10](https://www.youtube.com/watch?v=8cAQdELWYuo&t=10s) The kids are paying attention. Not bad. — Host
> [01:20](https://www.youtube.com/watch?v=8cAQdELWYuo&t=80s) I think the results are crazier than we expected — Dogus Cubuk
> [03:00](https://www.youtube.com/watch?v=8cAQdELWYuo&t=180s) materials don't come out of the oven with labels on them. — Dogus Cubuk
> [14:14](https://www.youtube.com/watch?v=8cAQdELWYuo&t=854s) We don't think that you can just start thinking your way to a new room temp superconductor by reading a textbook, shutting it, and thinking super hard. You have to make contact with reality. You have to carry out experiments. — Liam Fedus
> [27:24](https://www.youtube.com/watch?v=8cAQdELWYuo&t=1644s) If you automate accountants, at some point, there's only so much accounting to do. But if you can automate science, there's no end to it. — Dogus Cubuk
> [33:28](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2008s) You can basically timestamp the universe. So, OK, here's the set of known... data we have experimentally, computationally, at some timestamp, t, and produce a world-modeling task where you're trying to, then, predict outcomes of future things... The world model is not just in weights. It's not just inference time. But it's using these tools in order to do that. — Liam Fedus
> [38:21](https://www.youtube.com/watch?v=8cAQdELWYuo&t=2301s) The bar is so low in terms of how physical and relevant you have to make AI to make progress. You probably can just be slightly above it... I think the bar is pretty low right now, but it may not stay like that in a few years. I think a lot of young, excited people will realize LLMs can do way more than coding. — Dogus Cubuk
## Mentioned Albert Einstein, Andreessen Horowitz (a16z), Caltech, ChatGPT, DeepMind, Density Functional Theory (DFT), Dogus Cubuk (Dorje Chubak), GNoME (Graph Networks for Materials Exploration), Graph Neural Networks (GNNs), Heike Kamerlingh Onnes, Hohenberg-Kohn theorem, ImageNet, Inorganic Crystal Structure Database (ICSD), Jason Kwon, Jensen Huang, Liam Fedus (Liam Ferriss), Manseesh, Meta, MIT, NeurIPS, Onnes, OpenAI, Periodic Labs, Rob Riachi, Scott Nolan, Stanford University, Tesla, Universal Materials Accelerator (UMA), Waymo, X-ray Diffraction (XRD)
## Video description For more information about Stanford's online Artificial Intelligence programs, visit: https://stanford.io/ai
Follow along with the course schedule and syllabus; visit: https://cs153.stanford.edu/
In this CS 153 Frontier Systems lecture, Periodic Labs cofounders Liam Ferriss and Dorje Chubak — veterans of OpenAI's post-training team and DeepMind's Genome project, respectively — walked students through their eleven-month-old mission to apply AI to the physical world, specifically using autonomous labs to accelerate the discovery of new materials including high-temperature superconductors.
Operating out of a 40,000-square-foot Menlo Park facility where machine learning researchers work alongside physicists and chemists, their AI system (named Onnes, after the scientist who discovered superconductivity in 1908) runs a continuous loop of computational prediction, robotic synthesis, and experimental verification — closing the feedback cycle between digital intelligence and physical reality in a way no purely in-silico approach can. A key early lesson was abandoning their original plan of spending the first year purely computational: building smaller, semi-manual labs first allowed them to direct the research program and understand what to scale, faster.
Dorje emphasized that the results of applying LLMs to actual atoms have exceeded even their own expectations, while Liam stressed that sample efficiency in reinforcement learning — not benchmark climbing — is the core technical frontier when physical experiments can't be arbitrarily scaled up the way digital rollouts can. They closed by pushing back on student anxiety about AGI displacing their careers, arguing that most scientific domains remain largely untouched by LLMs, that the bar to make meaningful AI-physical world progress is still surprisingly low, and that the history of civilization is essentially a materials story — making their work, in their view, among the highest-leverage bets anyone can make right now.
[Source](https://www.youtube.com/watch?v=8cAQdELWYuo)