Twenty-atom crystals taught an AI the physics of a heartbeat

Nathan Labenz, host of The Cognitive Revolution, brings on Jonathan Godwin, founder and CEO of Orbital Materials, and Tim Duignan, a returning guest who has since joined the company as a researcher. Most listeners, Labenz figures, are plugged into the AI moment and almost none into materials. Godwin starts with semiconductors. You etch circuits into a silicon wafer; the trenches are so thin that a speck of dust can kill the chip. You therefore need a material that does not conduct electricity and is still small enough to squeeze between those features and keep debris out. The industry calls them low-K materials. Without yearly improvements in their performance, he says, you do not get more transistors, and you do not get AI.

The whole Silicon Valley stack, in his telling, sits on that kind of work. Materials science is "the original tech startup," and the 21st century will be even more so than the 20th: more chips, ships to Mars, optical materials for virtual reality. Materials are the missing piece of the science-fiction future, and AI is the tool he wants pointed at it. Duignan, who has been in the field longer, adds climate. Batteries and solar panels — the most successful carbon-cutting tools built so far — came from materials breakthroughs. He likes to play the inverse game: name a problem, and he will show you how it is a materials problem. Bell Labs and the transistor, he notes, were materials science: trying different elements until something worked.

The experimental life still looks a lot like luck. Hypotheses exist, but so do trial and error and a gut built over years. Duignan analogizes it to bridge-building a century ago. The Brooklyn Bridge was massively overdesigned because no one could compute the loads; the Tacoma Narrows Bridge later collapsed in a resonance no one had modeled. Today every bolt is calculated before steel is cut. Orbital's motto is "engineer in silico, materialize in reality": know in advance that a material can be made, that it will have the properties you want, and, when a yield tanks, see what is going wrong at the molecular scale instead of guessing.

Godwin was leading a DeepMind team building large-scale materials models when the field's pace convinced him the work would not wait a decade. Simulations that were impossible on the world's best supercomputers ten years ago now run on laptops. The company exists, he says, because neither Google nor a traditional chemistry firm was going to put a frontier AI group next to a wet lab of experimentalists. Traditional companies have been saying no to computational tools for 25 years; they will not suddenly start building frontier tech. A new company had to close that loop.

Labenz has been chewing on a worldview he wants to shout from the rooftops, if the guests will sign off. Today's architectures seem able to learn an intuitive physics in almost any space you give them. People catch a ball in sub-second time; they do not integrate Newton's laws on the way. The same shortcut, he says, keeps showing up in protein folding, materials, weather, and single-cell transcriptomes. For him it is the clearest reason to believe superhuman intelligence is not just possible but increasingly likely. Duignan agrees, and offers a theory: dimensionality reduction, which physicists call coarse-graining. You do not need every part of a system. Electrolyte simulations can ignore electrons; much of biology can ignore solvent; a fly ball is one object. Mori-Zwanzig theory and Langevin dynamics are, in his view, intimately related to what the algorithms are doing.

The sentence he keeps returning to is that finding efficient ways of keeping the important information and losing the rest is a central problem in physical modeling, and that machine learning is very good at it. Godwin adds evidence from scale. DeepMind's Genie 2 and similar world models — diffusion networks with only a few billion parameters — already simulate water, reflections in puddles, and gravity, effects that recently demanded expensive CGI. Diffusion models are not parameter-heavy; they are data-hungry, and they use far less inference compute than a large language model. The implicit physics gets packed into a surprisingly small space. He compares the networks to elite athletes: they score extraordinary goals and do not know the math.

Human materials scientists already live on a version of that gut. Godwin borrows a semiconductor story: people spend 30 years on one tiny step of a roughly 10,000-step chip process — which gas cleans a wafer after an etch, how a slight change hits the silicon. The knowledge is not in textbooks and is not easily deduced from first principles. It passes from master to apprentice. That, he argues, is why money and espionage have not reproduced TSMC: the intuition sits inside a cohort of chemists and materials scientists. One goal of AI is to break that monopoly. It is still physics. If the model gets the physics right, it should be able to democratize a kind of prophetic craft knowledge. Duignan finds the loss sad — it is a beautiful body of know-how — but the human brain has limits, and the interesting future is human intuition and the model correcting each other.

Chemistry's experimental data, Duignan says, used to make him pessimistic about AI. For electrolyte solutions he cared about, a database might hold a few parameters per salt. The click was that first-principles computation can mint training sets: a large simulation is millions of points, and active learning can generate more where the model is uncertain. Godwin's further claim is about generalization. A catalytic reactor that turns CO2 into sustainable aviation fuel may need thousands or millions of atoms over long timescales. You can instead simulate a scrap of that world — maybe a hundred atoms — learn the physical rules, and apply them at inference to systems that classical methods cannot touch. The result that blew his mind: models trained on 20-atom inorganic crystals simulating a protein out of the box. That, he says, means something fundamental was learned at the small scale. There may simply be a finite amount of interaction complexity to capture.

Orbital's stack has two layers. Linus, the foundation model, is a diffusion, or score-based, generator of materials. Duignan traces the math to nonequilibrium statistical mechanics: the score vector field that atoms follow while a structure is generated is analogous to the force field that pushes them in a molecular simulation. A paper by Cecilia Clementi shows that if you train a diffusion model on a simulation "movie," the score is mathematically identical to the force field. Godwin's operational reading is that diffusion is a superb pre-training step because noise corruptions give you essentially infinite samples of a simplified physics. Fine-tuning on forces from quantum calculations then produces ORB, which computes quantum-level interactions between atoms. Get those forces right, he says, and you have physics right: nuclear reactions, the body, everyday materials.

Noising a crystal is closer to images than to proteins: jiggle the atom positions until the structure is incoherent, with extra care because a crystal in a box wraps around — an atom that leaves one face reappears on the opposite. Discrete atom types, integers from 0 to 117, are treated as a continuous target and cast back to categories, a 2022 Hinton-linked trick Godwin calls analog bits. The architectural difference that actually matters is distance. When two atoms approach, forces rise; when they are far, forces vanish; in between sits a Goldilocks well of attraction. Distance is a first-class citizen in a 3D point cloud and a non-event for pixels or tokens, which sit on a grid. Orbital therefore uses a message-passing neural network — "a Transformer that takes distances as inputs." Because there is no positional encoding, simulation size is not capped by a context window; it scales linearly with chips.

A scientist shows up to Linus with a number, not a paragraph — a band gap of X, ten candidates — because materials want a sweet spot, not a caption. The public ORB models take coordinates and atom types and return forces and energies, which unlocks diffusivity, equilibria, and the rest of the classical simulation toolkit. Duignan describes three uses: a computational microscope at nanometers and nanoseconds, where bonds form and break; in silico experiments, where you change an atom type in a file instead of ordering chemicals; and a bedrock for coarser models, including training data for the next generator. In the design loop, a generative model proposes ten to twenty porous, ligand-ready frameworks; AI-accelerated simulation then answers the questions a prompt cannot — kinetics, blocked pores — until, Godwin says, 90% of what you need to know is settled before anyone synthesizes. The lab success rate is higher, and the candidates sit further outside human chemical imagination.

The mountain Duignan had wanted to climb for years is the potassium ion channel. The body is an electrical device. Computers move current with electrons in metal; biology uses ions — sodium, chloride, and, critically, potassium. Cells store a high concentration of potassium; when a signal is needed, channels open, ions flow, voltage changes, and the next process fires. Muscle contraction, the heartbeat, and every neuron in the brain ride that current. Mutations in the channel sit behind heart diseases, including long QT syndrome and palpitations that run in Duignan's own family. He calls it a keystone problem: even basic questions are open, including whether potassium ions go through the selectivity filter alone or with water. Papers in top journals land on opposite sides every few months. A tool that could settle it would demonstrate a capability the field is missing.

ORB had not yet hired him when the model dropped. It was accurate and, more exciting, fast. He tried it on electrolyte solutions, got stable results, and began to wonder whether this was the universal force field computational chemists had dreamed about for decades. If it was, the potassium channel was the problem he would try. He set up the simulation expecting it to crash; his own earlier models had gone unphysical the moment they left distribution. It ran. He started seeing interesting things and, he says, literally jumped up and down. The training data looked nothing like the target: small, highly periodic crystals, never a wonky membrane protein. Reasonable dynamics from that mismatch was the Eureka.

The work is not peer-reviewed. Potassium-channel experts have told him it looks interesting, plausible, and consistent with experiment; he still wants direct experimental confirmation. Two observations stand out. Water did enter the selectivity filter, unlike classical simulations that tend to send potassium through alone. And a hydroxy group on the protein appeared to reach out, grab water, and pull it into the channel — a qualitative step he has not seen reported. Mutation studies already show that deleting that hydroxy group drops conductivity by almost an order of magnitude; no one had a mechanism. If he had to bet, he would bet this is why. Practically, understanding the healthy channel is a step toward simulating disease mutations and, further out, designing drugs that restore function. That sounds sci-fi, he notes, except Vertex already has small molecules that restore chloride-channel function.

The engineering was frugal. At the time he had one V100, a 2017 NVIDIA card. To keep the selectivity filter stable he froze atoms about 20 angstroms out — four or five layers — a hack borrowed from earlier simulation practice, cutting off an infinite regress of "you also need the atoms around those." A full run with the cell membrane would need a much larger system and more GPUs. There is no Transformer-style context window: Godwin notes that without positional encodings, size scales with chips. The real bottlenecks are long-range electrostatics, which decay as 1/r the way gravity does, and time. Huge systems get slow; timescales stay short. Duignan's path to whole-cell work is more coarse-graining: run all-atom trajectories, then train a model that ignores solvent and predicts a free energy over the pieces you care about. The theory, he says, is there. It is a matter of investing in it.

Labenz asks the obligatory question: when do we get a room-temperature superconductor? "It'll happen. Give us at least five years," Godwin says, "but it may happen before that." Carbon removal and the energy cost of scaling AGI come first. Duignan explains the technical brake. Superconductivity is a highly correlated electron system; it does not have the local behavior their models exploit. Density functional theory, the current training target, is not really good enough to describe it. AI can still help hunt new crystal structures once you know which electronic properties you want to induce.

The near-term focus is data-center materials. Scaling AGI, Godwin argues, is among the most important things they can work on, and new materials are how you make that buildout cost-efficient and sustainable as rack density and chip power rise. Decarbonization is the twin track. One product already public is a carbon-removal material, designed to capture CO2 under data-center conditions and due to pilot with AWS later in the year. It is sponge-like, with reactive sites meant to prefer CO2 over the rest of the air, using waste heat and waste airflow. Hyperscalers, Labenz suspects, will not stop building AGI clusters; something has to sit between them and a pile of broken net-zero promises.

The cost lever is regeneration, not the one-off grab. If desorption can be tuned to ride a data center's waste heat, energy cost falls; if the material lasts a decade, you amortize a making process that is itself energy-intensive. Orbital does not plan to stand up a plant in New Jersey. The analogy is NVIDIA: no fab, still in the chip business, via licensing and suppliers. They need more physical capital than a pure software firm, Godwin says, but they are not raising for heavy capex. The lab is cheaper than the GPU fleet an AI company needs, and having experimentalists in-house is what makes the feedback loop fast enough to matter. World-class models and a materials pipeline on less than $40 million raised is possible, he adds, because the networks look more like image diffusion than like large language models. MosaicML once got Stable Diffusion training under $50,000; there is room to throw far more compute at higher-quality first-principles data.

Labenz cites a then-circulating MIT paper by Aidan Toner-Rodgers on AI in a materials-science setting. Headline numbers: 44% more materials discovered, 39% more patent filings, 17% more downstream product innovation. The models did a lot of idea generation; scientists moved toward validation. Output rose. Eighty-two percent of the scientists reported lower job satisfaction, because they had wanted to be the ideas people. Labenz is not precious about it — if the shift helps on climate, he will take some lost satisfaction — but he wants the guests' read on throughput and on the future of work.

Duignan puts himself in the happy minority. Trained as a physicist and mathematician, not a chemist, he finds the tools empowering: they give him the intuition he does not have from a career of flasks. He can run a simulation, watch bonds form, and see why a hydroxy group matters without having read the potassium-channel literature. He has sympathy for people who spent a lifetime building that gut and now watch something else do it. His offset is blunt. Getting materials into the world faster matters more than the spark of solving the puzzle yourself. "You can do crosswords for that." The work is too important to optimize for the scientist's happiness.

Godwin's circle of AI researchers has already lived a milder version. Job satisfaction has dropped even as the work got more prestigious and better paid. People used to do creative research; now they engineer, and compute is too expensive to test your own idea properly. Creativity, he thinks, is among the first things AI takes; robotics and manual tasks come later, after the creative industries have been accelerated. He is not optimistic that people will enjoy AI. He is convinced the only happy long-term path runs through it, because the world needs more stuff for less money, which means a faster burst of scientific discovery than humanity has ever had. Short-term pain, in his phrasing, for a future in which a large portion of economic activity is automated. Meaning will have to be found outside the job.

Labenz adds a safety cut. A central worry about superintelligence is how it gets experimental grounding. Pair a language model with a reliable neural network potential, and the system has a physical checker: it can test hypotheses on the computer instead of in a lab, the same way reasoning models need a way to validate answers on math benchmarks. Shared embedding spaces across language and atomistics strike both guests as likely; Godwin would like atomistic data in the training of a scientific AGI, and assumes someone is already trying. Labenz worries that reasoning in latent space — after a season of deception discourse — means we stop being able to read the chain of thought. Godwin's dry closer, quoting someone he had seen: you just hope moral realism is true, and that the system does the right thing anyway.