ChatGPT's Co-Creator Left to Put Nature in the Reward Loop

The ambient-pressure record for high-temperature superconductivity still sits around 135 Kelvin, roughly −138°C, and it has barely moved in years. In October 2025, Periodic Labs announced a $300 million seed round led by Andreessen Horowitz. The company is about 30 people. Its founders are Liam Fedus, co-creator of ChatGPT and former VP of post-training research at OpenAI, and Ekin Dogus Cubuk, who ran materials science and chemistry research at Google DeepMind.

a16z general partner Anjney Midha hosts the conversation, which Cognitive Revolution cross-posted. The interesting question is not the round. It is why frontier models that already look strong on math and code still fall short on physics and chemistry, and what it actually costs to close that gap. Periodic's north star is to push that 135K number toward 200K or beyond, on the theory that getting there will force a stack of independently valuable sub-goals along the way.

Eight years ago at Google Brain, Cubuk was trying to flip a giant tire in the company gym and could not do it alone. He went looking for the strongest person he could find — it was either Fedus or Barret — and the two of them flipped it. That is how the co-founders met. Over the following years they kept catching up, and the conversation kept sliding into quantum mechanics and superconductivity, even though Cubuk never expected they would end up working on physics together.

Cubuk was doing materials research at DeepMind and noticing that language models had become unavoidable in his workflow. He used chatbots to recall chemistry he had forgotten, and he used them to write simulation code. The question became whether an LLM could be a first-class citizen in physics research, not just a side tool. Fedus, watching reasoning gains and high-compute reinforcement learning on the language-model side, saw similar scaling patterns in materials simulations and experiments.

Both of them, and many people in the field, treat accelerating science and physical R&D as the point of the technology. "Chatbots was a great milestone along the way, but we really want to see technology out in the world," Fedus says of leaving OpenAI. Physics looked like the right place to begin: it is verifiable, the reward is fairly fast, and simulators already exist for large classes of systems. They built that conviction and founded Periodic.

Fedus walks through how ChatGPT was first trained. You start with a pretrained autocompletion model, add supervised input-output pairs that teach it to act like an assistant, then run RLHF against a reward model built from human preferences. Early reward functions encoded being a friendly helper and had almost no sense of whether the math was correct or the code was valid. Later progress came from making those graders more precise: math has a right answer, code has unit tests, both are programmatically checkable. All of that is still digital, built from the internet, textbooks, and papers.

Periodic's move is to replace math and code graders with a physical reward function. The model proposes hypotheses, calls simulation tools, then actually synthesizes materials and measures properties, and those measurements come back as the learning signal. If a simulator is wrong, experiment error-corrects it. Fedus's phrase is that nature is their RL environment. Agents still get tools the way a coding agent gets Python and a browser; now those tools include quantum-mechanical simulation, and the loop ends in a lab.

In AI, "lab" usually means a research org. Periodic means a room with powders, furnaces, and measurement gear. The first facility is powder synthesis: mix existing powders, heat them to a set temperature, get a new material. Cubuk compares the robotics to the coffee-making robot at San Francisco Airport — that level of robot can mix powders and load a furnace. The method is cheap and simple, and it still covers superconductors, magnets, and other technology-critical materials. Underneath all of it is quantum mechanics at the Schrödinger-equation energy scale, where biology, chemistry, and materials actually happen.

Why can today's deployed models not already do this? Cubuk's answer is that science is iterative. An LLM can be very smart and still fail to discover science if it never iterates, just as a human locked in a room with no experiments will not discover anything important. What the models lack is the method of inquiry: simulate, calculate, experiment, get a result that is probably wrong, iterate. Fedus adds three reasons they have to own the lab. Literature values for a single property can span many orders of magnitude, so a model trained on them at best reproduces the noise. Negative results are almost never published, even though a valid negative is a precious learning signal and is highly context-dependent. And science requires the ability to act, not only to read papers.

Midha asks the investor's question: are superconductivity and magnetism first pipelines because they matter, or because they are stops on the way to a general AI physicist — and could they become off-ramps? Cubuk answers in layers. High-temperature superconductivity contains countless sub-goals. To get there you have to get good at autonomous synthesis, autonomous characterization, and using an LLM to run simulations correctly. It is like DeepMind and OpenAI saying they were going after AGI when what they actually had to do was a pile of independently cool results first. The north star's job is to force those sub-goals into existence.

The scientific stake is separate from the product. If they found a 200K superconductor, Cubuk says, even before making anything with it, seeing quantum effects at that temperature would update how people see the universe. Superconductivity is also a phase transition, so the transition temperature is dominated by fundamental crystal properties more than by defects and microstructure that simulators still cannot capture. Other material properties can be present in the crystal and still be masked by unsimulatable factors. That makes superconductivity a relatively clean verification target. It also rallies the whole team: forty-year physics professors and people who never studied physics both get excited about it.

Progress is easy to score. Today's ambient-pressure record is about 135K; they will know they are doing well if they go past it. Cubuk notes that the signal is hard to hack: unlike many LLM training tricks, what you measure in the real world is what the model sees. Fedus adds that the tactic is to complete the full loop somewhere rather than talking in vague terms and sliding back onto arXiv papers. Then they want a repeatable process for moving from subdomain to subdomain, including the open question of how well a system trained on superconductivity generalizes to magnetism versus fluid mechanics.

About half the company comes from machine learning and half from physics or chemistry. Cubuk's method for joining the cultures is to keep them pressed together. Physicists and chemists have to decide which steps belong in mid-training and RL so the model learns to reason about quantum mechanics. LLM researchers, meanwhile, are learning physics, simulation tools, and the actual scientific goals. Weekly teaching sessions run both ways: ML people explain RL loops and data cleaning; physical scientists teach the science, including the history of science.

Fedus notices that computer scientists think in APIs. A scientist says something, and they automatically map input, output, target, and how to wire it back. That translation is itself useful. The team also hires people who sit inside the simplex of pure ML, pure experiment, and pure simulation, as bridges; Cubuk is the example they name. Culture is "no stupid questions." The dumbest physics question and the dumbest ML question are both allowed, and a few faculty inside the company make the teaching sessions unusually good.

Do you need an advanced degree in physics or chemistry to join? Cubuk's answer is absolutely not. He inverts an NBA line about being closer to LeBron than you are to him. Even the best physicist on the team is ignorant of far more physics than they know, so a candidate with no physics background has roughly the same mountain to climb. Nineteenth-century physicists could work at the frontier of several things at once; today a leading thinker usually advances only a narrow field. Discovering a great superconductor requires chemistry, physics, synthesis, and characterization, and no human holds all of it. Collaboration is not a nice-to-have. The company is a miniature of that bind.

Mid-training, Fedus explains, showed up when pretrained models had a knowledge cutoff and search was not yet good enough. You take data the model does not have and continue pretraining. That is different from post-training, which is mostly reinforcement learning and supervised learning. The point of mid-training is to pour in a lot of knowledge that was not there before. At Periodic that means low-level crystal structures, high-level semantic descriptions of how a material was made, simulation data, and experimental data, all driven into the same model.

Mixing distributions A, B, and C does not guarantee generalization. What they want to see is that adding one dataset improves performance on the others. That is the ML problem to solve, and it is how you turn a model that was deficient in physics and chemistry into an expert. Many advanced-manufacturing companies currently lean on retrieval, a lightweight pitch in which someone arrives with a neural net and offers to search the corpus. Pretraining and mid-training encode knowledge in the weights, so the model actually understands the material instead of fetching documents.

Permissions make that harder. Retrieval can act as the user and inherit their access. If you pretrain on data only the CEO can see, you have to bucket knowledge and build different systems. After a morning customer call, Fedus's read was that they still have no good way to distill knowledge into a single model or a set of models. Their supervised training also looks like early ChatGPT: a few input-output examples. They need to switch to high-compute reinforcement learning, effective tool use, and optimization against the thing they actually care about.

Other frontier labs aimed at an AI researcher and, along the way, hit AI programming, the first domain that made people update how useful these models are at work. Midha asks what Periodic's equivalent is. Fedus's answer: copilots for engineers and researchers in advanced industries. Silicon Valley thinks in bits. Space, defense, and semiconductors still iterate on materials and physics, often without good data or good tools. Those R&D budgets are enormous. That is the opportunity.

High-temperature superconductivity remains a good north star, but technology and capital are intertwined. Fedus is direct: they will accelerate science most if the company is wildly commercially successful. The path is to become the intelligence layer for those industries, shorten iteration time, and get researchers and engineers to better solutions faster. Scaling laws, in their view, still hold empirically. The fight is over the y-axis. A coding model with a unit-test reward can take off as a software engineer and still will not cure cancer, because the knowledge does not exist in that environment. You optimize against the distribution you care about.

Cubuk's vision paper with Fedus, and the CLIP literature, showed that in-domain and out-of-domain generalization are monotonically correlated but not linear. Improve the model and in-domain performance rises as a power law; out-of-domain performance also rises as a power law, but if you are far from the training distribution the slope can be so small it is useless — centuries, not years. Formation-enthalpy labels in the synthesis literature are so noisy that models trained on them are not predictive. Superconductivity datasets exist, but the noise floor is too high to help. Periodic's strategy is to manufacture high-quality data and close the loop until the target sits on the training distribution.

Deployment is land-and-expand, not a day-one rewrite of the fab line. They look for a well-scoped critical problem, co-draft evaluations with the customer, and prove how strong the technology is when you optimize against what that team actually cares about. Then they find the strongest internal champions and match them to where Periodic's technical capability is highest. On a pipeline call that morning, Cubuk heard simulation training as an urgent pain: people spend huge amounts of time learning critical simulators, plus format-matching, feeding results into design pipelines, and keeping data in one place. Those companies are also losing senior engineers and looking for an AI strategy that is not changing as fast as they think it should.

Most fundamental work in the physical sciences has historically happened in university labs. Cubuk says academia absolutely still has a place. Many of the simulation tools Periodic uses were developed there, a lot of them in Europe, along with novel synthesis methods. Some of those tools are complicated Fortran that the company would not develop efficiently itself. Large-scale simulations now often run at Microsoft, DeepMind, and Meta, but the tools usually started in academia and then moved. Periodic also benefits from improving base LLMs and from open-sourced machine-learning property predictors; they take a pretrained model and mid-train it with high-compute RL rather than starting from scratch.

Midha had run evals at a Stanford physics lab and found models terrible at scientific analysis. Fedus's reply is that nobody trained them to do it. Academia can tell industry which tasks matter, how the analysis should go, and which skills to put in the model — a full analysis or a smaller primitive inside one. One Periodic physicist looked at a model's reasoning and said it was all wrong: it should be thinking at a higher level, in terms of symmetries, the way a particular book teaches. RL environments need to reward those strategies, because premier scientists use them. Industry is blind to that style of thinking without the partnership.

Language is not the whole brain. Cubuk notes that humans also do visual and geometric processing, so LLMs may benefit from separate geometric reasoners they can call — equivariant graph neural networks, diffusion models that are geometric by construction. Language is good for a synthesis recipe; geometry is good for representing atoms and designed structures. Two concrete academic programs are already in motion. An advisory board includes Stanford experimentalist ZX Shen, theorist Steve Kivelson, Northwestern synthesist Mercouri Kanatzidis, and Chris Wolverton on high-throughput DFT. A grant program will take proposals on LLMs, agents in synthesis, materials discovery, and physics modeling, including work that is a bad fit for industry.

What they want in hires overlaps heavily with a frontier lab, with one extra filter: do you actually care about accelerating science? Fedus says people who want to improve a mega-corp product should go iterate there; people who care about discovery should come to Periodic. They look for deep curiosity, contact with reality, pragmatism, careful process, solution orientation, speed to a goal, and being world-class along some dimension — ML, experiment, or simulation. Cubuk adds urgency. They do not want LLMs to start improving science in ten years. They want it ASAP.

The practical test Periodic ran on scientific literature is one any vertical AI team can copy. Check whether labels span orders of magnitude, whether negative results are missing, and how far the target distribution sits from the data you have. If any of those fail, more compute will not save you. Cubuk's power-law observation is the second test: a strong in-domain benchmark does not mean your real use case will climb at a usable slope. If the benchmark is far from the job, you may need to make the data yourself and close the loop.

The third choice is retrieval versus training. Retrieval is light and ships quickly, but the knowledge never enters the weights. If the work requires a model that understands the domain, mid-training is the more serious path, at the cost of messier permissions and data governance. Three hundred million dollars, thirty people, and a 135K-to-200K target is a bet that will take years to grade. The question it poses is already live for anyone building applied AI: how far is your reward function from the real world?