A hundred experiments a day, and the valley of death in materials

Nathan Labenz spent a year as an undergraduate research assistant weighing fine powders and running tedious experiments. He was told automation of that work probably would not happen in his lifetime. On The Cognitive Revolution he sits down with Joseph Krause and Jorge Colindres, co-founders of Radical AI, whose lab is already targeting a hundred experiments a day. That throughput, Labenz notes, roughly doubles what he did in a full undergraduate year and what Krause did in a full year at the Army Research Lab.

Advanced materials, they say, are the foundation of semiconductors, aerospace, energy, defense, automotive, climate, and electronics. The process of finding new ones has not caught up. Cost typically runs north of a hundred million dollars from a computational framework through scale-up. Time runs ten-plus years to commercialization, sometimes spanning decades. Fragmentation between the people who discover and the people who manufacture is the root of both.

Colindres splits the field into two paradigms. Academic research drives understanding and is not aimed at a product. Corporate R&D is aimed at commercialization and at one, two, or five percent performance gains that can move a margin and a Wall Street slide. In the middle sits a white space: novel fundamental discovery that is still pointed at a commercial application. That gap, they argue, is where Radical AI intends to live, and where enabling materials, not slightly better paint, would actually show up.

Labenz reaches for the drug-discovery analogy: low-hanging fruit picked, clinical trials ever more expensive. Colindres grants the similarities and the differences. Most materials never enter a human body. Search spaces are still enormous, smaller than proteins but still, in his word, impossible to walk through. The analog of a clinical trial is scale-up and processing: take a lab reaction the size of a quarter, a “button,” and turn it into a 400-pound piece without losing the properties you cared about. That is a different kind of hard than a Phase III study, and it is where a lot of discoveries go to die.

Labenz wonders why materials scale-up should not be getting faster with accumulated engineering, the way a Warp Speed moment sped vaccines. Krause’s answer is that speed requires different tools, not more of the same. Self-driving labs are not a thousand copies of one protocol. They use active learning, so every run incorporates the last one, the way a human scientist would. At the Army Research Lab he ran about fifty experiments a year. Directed government programs, the Mach program among them, might push four or five hundred. That, at Radical AI, is a week.

The experimental data, Colindres adds, have to thread from the button all the way into manufacturing. If properties vanish on the way up, the question is why: what changed environmentally or chemically, and how do you stop it next time. Capture that chain and they think discovery can move hundreds of times faster. Fail to capture it and you have another orphan: a material that existed at coin scale and never left the bench.

Orphan materials, Krause says, are probably the bulk of the work. Silicon in semiconductors took a long time to push through and is still tightening node sizes. Room-temperature superconductors have been talked about for decades, with applications that would be extraordinary, and the field still struggles to understand synthesis. LK-99 is only the latest reminder. Plenty of ideas die between a computational hit, a lab that cannot reproduce it, and a factory that never starts.

The story they love is Corning. In the 1950s the company developed Chemcor. It sat on a shelf for half a century. Around 2006 or 2007, Steve Jobs walked into the upstate New York offices and said he was building a phone that would change the world, that every phone had a plastic screen, and that Apple cared about design. Corning had a product that had been sitting around. Chemcor became Gorilla Glass, now in nearly every Apple device of the last two decades, plus Samsung and LG, and one of Corning’s largest revenue drivers.

Two morals, Colindres says. First, we often do not really understand the materials we make, or which properties are the good ones. Second, you need a customer to tell you what they are looking for before those properties map onto value. Radical AI therefore tries to pull customer expectations and value drivers in at the start, so design, experiment, scale-up, and manufacturing are always pointed at an end use rather than at a pretty sample.

Labenz had filed LK-99 as a measurement error, a team that never actually had the thing. Krause agrees with the popular reading: an accidental copper-sulfide impurity is the recent explanation. He adds a methodological layer. Even the recipe was not reproducible. Temperature ranges, what the material was made in, the container that would of course affect the result, went undisclosed. Synthesis, not the press cycle, is the hard part.

He uses bread. Give two people the same ingredients. One scores the loaf, lets the starter rise, hits the right temperature and bake time, and gets bread. The other rushes it in thirty minutes at a thousand degrees and gets a rock. Same ingredients, a process that can still be described as “heat it, score it,” and a completely different output once time and temperature move. A new material, on his standard, exists when you can make it in a lab. A new commercial material exists when you can scale what you made. Room-temperature superconductors have done neither yet. Time will tell whether the material is out there. The field may have moved on. Radical AI, he says, thinks about the problem frequently, and RTAP superconductors sit on the long road map.

The company was not founded because materials looked like a large market, Colindres says, though they are. The founders wanted floating trains, interplanetary travel, a Tesla from New York to Boston without a recharge, things they think should be available in their generation and currently are not because materials keep getting in the way. A superconductor is the epitome of an enabler: second- and third-order products nobody can name yet, the way nobody named Uber the day the screen became glass you could touch.

Asked for the canonical unlocks, Krause lists room-temperature ambient-pressure superconductors, high-entropy alloys, ultra-high-temperature ceramics, ceramic-matrix composites, semiconductor materials. None of them, at Radical AI, get made in a vacuum. Academic groups and national labs can chase understanding for its own sake. This company wants materials that leave a mark by solving a huge industrial problem or enabling a new industry. The application comes first, then the composition.

High-entropy alloys, a core focus, combine mechanical and thermal performance instead of trading one for the other: high strength at high temperature, plus oxidation resistance. The same family is a candidate for hypersonic flight above Mach 5, where heat, pressure, and atmospheric oxidation have been a national unsolved, and for nuclear fusion reactors where tungsten cannot take the radiation and erosion. These are not five percent games. Krause talks about fifty times, a hundred times better than what exists. Amorphous metals, a subcategory, are another option in the same neighborhood.

Zoom out and every industry can name a materials problem. Lululemon wants silver nanoparticles that wick sweat and cut odor. Hypersonic aircraft and space travel want heat, radiation, and pressure blocked in a corrosive atmosphere. Customers rarely ask for one number. They want mechanical performance in a temperature and pressure environment, at a cost, with a supply chain for the raw elements. Optimize thermal expansion alone and you have not finished the job. The search space is endless, from a workout shirt to a radiation shield on Mars, which is why they think it is one of the largest markets in the world.

A customer arrives with a Mach 5 aircraft in mind: no heavier, stronger, able to take heat, maybe radiation. The flywheel has two buckets, an AI engine and a fully robotic self-driving lab, locked together by data. Human scientists already look at microscope images, read papers, lean on intuition from past runs, and do some computational modeling until they are confident enough to try an experiment. The engine is an attempt to put all of those modalities, architectures, and that hard-to-name scientific intuition into one system.

The method is property-driven optimization. Start from the properties the customer named, the allowed elements, the oxidation target, the strength, and let the engine design experiments it thinks will hit them. The lab’s operating system then receives a design of experiment: composition, how to make it, what to look for. It synthesizes, characterizes microstructure, and runs property tests, stress-strain, creep, hardness. Machine-learning submodels analyze the streams in real time, the way a scientist would learn to watch grain growth in an SEM if the target is strength, and feed the insight back as active learning. Experiment two is experiment one, updated.

A human did that serially: one XRD, one SEM, one tensile test, maybe a paper remembered, maybe not. The AI scientist, Krause says, compares those results to millions of XRD scans from publications, prior lab runs, and other datasets, and forms a hypothesis in a dimension a human brain cannot enter. The power is the combination, not the engine or the robots alone.

The family of models is mixed: graph neural networks, which materials people call MLIPs, for atomistic modeling; generative models; language models, including the GPT family, for literature, extraction, and recommendations; computer vision in the lab. Computational GNNs are the part they are most confident will work. The bet they want to place is on language: reasoning over scientific knowledge, calling a GNN when uncertainty needs a calculation, calling vision when SEM images should update microstructure intuition. Nail one thing, Colindres says, and it is the replication of scientific intuition and the design of the next experiment. GNNs matter. They are a tool the language model taps.

Labenz asks what the ceiling of human materials intuition looks like, a von Neumann of alloys. Colindres points to their third co-founder, whom he will call one of the world’s leading materials scientists, top three if not the best, a person who connects dots others miss. Probe him and the magic reduces. He has been doing this for forty years, running experiments, watching students run them, reading colleagues’ papers, updating what works. Some of it is talent. A lot of it is learned. Krause’s own year at the Army Research Lab, after a PhD stretch on 2D optoelectronic materials, made him unusually fluent in neuromorphic semiconductors and ion-gated channels. That fluency came from the bench.

Of that year, he says, 90 percent of the work did not work. There is no dataset of the EBAM and PVD and ALD runs. There is one resulting material the team pushed out. The negative results live in a skull. That is why a forty-year academic can only really own one or two areas, and why the founders think a system that taps the exascale of uncaptured experimental knowledge could shrink a forty-to-fifty-year materials problem under a decade. Language is part of it and not all of it. Krause tells a 2D-semiconductor story: better conductivity tracked with film transparency, which later turned out to mean monolayer rather than multilayer TMDs. An expert had to walk in and name what they were already looking at. Accidents and negative results are a data type of their own. A human can look past them. An engine, they hope, will not.

Models trained on a distribution tend to treat the weird corner as out of distribution. Science’s biggest finds often come from the ugly neighborhood everyone skipped, so the stack has to prize novelty: Bayesian diversification, Monte Carlo, noise sampling, active learning, reinforcement learning, and, early on, humans in the loop to capture the uncapturable. The best eval, Colindres says, is the lab. There is no cleaner benchmark than whether the synthesis and the test did what the model claimed, without sacrificing the other constraints. Updates cannot just be a retrain of weights. Something agentic has to sit on top and adapt between retrains. They are also investing in mechanistic interpretability, in conversation with Eric Ho at Goodfire, because a black box you cannot steer will not stay pointed at a customer’s problem.

Labenz floats paying graduate students to digitize notebooks. Krause would take every lab notebook in North America and then hire a team to sit with the authors. XRD peaks mean different things in different subfields; for amorphous materials they used XRD to confirm the absence of crystal structure, data a crystalline specialist would call empty. He used to log a timestamp halfway through a reaction because he killed the heat there and could not control the cool-down. Without that context the numbers are noise. Computational data are cleaner and already heavily funded. Colindres’s line is that they solve about 10 percent of the problem. The highest-value data are experimental, expensive, and messy. Build the lab so that the data arrive structured, labeled, and in context, and you finally control the mess instead of inheriting it.

The search space, Krause says, is the stuff of Eddington’s number, on the order of 10 to the 80, the count of observable atoms they use as a stand-in for combinations. Alloy work draws on roughly sixty elements. High-entropy, multicomponent systems use five or six. Five elements in equal 20 percent shares already give about five million combinations, and no real alloy is equally balanced: each element runs from zero to a hundred percent, with as many decimal places as you like. Finding a new alloy is a particular grain of sand on a beach the size of the Earth. Proteins, at something like 10 to the 130, are a bigger beach still. That smaller beach, they argue, is an advantage. Materials are more likely to yield something in a foreseeable time.

They do not combinatorial-screen. Combinatorial science, in their view, has not produced enough value. Property-driven optimization plus active learning indexes the space in the AI engine first. Some constraints delete whole regions: care about weight and aluminum may simply leave the set. What remains is still huge, so several models have to succeed in series so that what reaches the lab is already decent. Explore-versus-exploit repeats at every layer, atomistic forces and lab recipes being related questions that still need different techniques. Offload what you can to simulation, and never forget that science is allowed to be counterintuitive. Systems have to be able to break their own rules.

The hard automation problem is not the rover track that moves samples. It is the scientific instruments, some of them twenty-five, fifty, seventy-five, a hundred years old, designed around a human hand. Software interfaces are missing or can read outputs without actuating a turbo pump, a heat source, or chamber pressure. Radical AI spends engineering time retrofitting those tools: custom sample loading, vacuum control, laser on and off, contamination control with multiple samples in a chamber. They do not redesign how X-ray diffraction works and they do not want to sell diffractometers. They want to sell materials. Scientists still orchestrate some steps today. Every tool is on a path from a human demonstrating it to fully autonomous operation, or it will be replaced by a custom build. Some mechanical switch-and-dial boxes, they think, will simply be retired over the next decade.

On interpretability, Colindres will not claim demonstrable grokking. They see the inverse, memorization and hallucination, which at least suggests an opposite pole might exist. Krause is excited for a more scientific reason: the monolayer-transparency heuristic is the sort of thing he would never have known to look for. Weights might hide correlations a human would not think to inspect. Strip the human bias and let the math speak, and experimental science has almost never had that option. They expect interpretability to matter for science at large over the next two to five years. A materials Move 37, an LK-99-scale shock the whole world feels, has not happened. Computational optimization is happening. Experimental flywheels, they think, are still the missing half.

Krause retells AlphaGo: millions of games, then a one-in-ten-thousand move the room thought was a mistake, then later versions that did not need the million-game diet. Radical AI is trying to build the first version of that stack for the lab. If they hit the roadmap, he thinks something like that arrives in 12 to 24 months. After it becomes normal, the hope is transfer: experimental data from alloys informing a room-temperature superconductor without millions of labeled examples in that exact space. They want the largest experimental dataset ever built. The slogan is a world limited by imagination and the laws of physics, not by materials, a world in which the next Uber and Nvidia are downstream of an enabling substance.

Labenz asks whether the zoo of submodels collapses into one joint embedding, language plus an intuitive physics of materials, the way recent image models stopped routing everything through a lossy caption. Colindres does not think they will ever need only one model for everything. Internally they are building something natively multimodal, with encoders for the modalities they care about. He expects to be using it in-house within six months, and to publish or release something in six to twelve. Interpretability on that object is already a conversation with Goodfire.

Software-only, they decided, never makes a real material; several companies have already failed that way. Pure licensing demands multiple new discoveries a year for decades, then runs into reverse engineering and patent expiry. So they looked at 3M, Dow, Applied Materials, BASF, Materion: companies that sell materials at scale. Those firms, Krause says bluntly, often serve markets that do not move the species, a new Sherwin-Williams color his wife will love. Radical AI wants interplanetary travel, lost nuclear energy, floating trains. Composition patents exist. The meat is processing: how you make it at scale and at lowest cost. Hand someone the exact high-entropy composition and they still may not produce it. Some models may be released, some will not. The company, Colindres says, will not stop going after the hardest needle-movers.

The day before the taping they announced a U.S. Air Force collaboration, Direct to Phase 2, on high-entropy alloys for hypersonic applications. China and Russia already field hypersonic systems, Krause says, because they never stopped investing in materials. When China finds a new material it stands up an entire manufacturing hub just to figure out scale. The United States has struggled with that nexus. They spend time in Washington with OSTP, NSF, the Department of Defense, the Department of Energy, and the Hill, arguing that public-private partnership on materials is not optional if you care about hypersonics, fusion, or the datasets that will train the next models. From stone to bronze to silicon, every technological wave rode a material. The next one will too.

Culture is the last pitch. They want people with ridiculous curiosity and first-principles stubbornness who are also relentless, because they fail every day on purpose and expect to keep doing so for a decade. Unlocks, they say, sit on those failures. If you want a job, do not apply. If you want a mission that might put a stamp on the species, reach out. They are recruiting for life’s work, not a role. Labenz closes by wishing them breakthrough after breakthrough. The materials, if the flywheel works, are supposed to change what counts as physically possible.