Swyx, a co-host of the Latent Space podcast, said he was worried about taking the conversation off the rails, but he asked anyway. He came from engineering and machine learning rather than physics, he said, and he owned plenty of physics books. His question: don't we have most of the laws of physics worked out, and how come we don't have perfect simulators already?
His guests were Liam Fedus and Ekin Doğuş Çubuk of Periodic Labs, a company building AI systems meant to discover new materials. Periodic's approach combines AI models, computer simulations of the physical world and its own high-throughput lab, where many experiments can run quickly and in parallel. Çubuk opened his answer this way: "we're definitely not done with any of the laws of physics."
The answer, in an episode published October 8, 2026, explained why knowing the rules of physics is not the same as calculating what a real piece of matter will do. It also explained why, in the discussion's telling, even a much better simulator would still leave a materials company running a lab.
"The rest is chemistry"
The discussion started with a story about Paul Dirac, one of the founders of quantum mechanics. According to the story, Dirac treated the theory as essentially finished in the late 1920s and said that "the rest is chemistry." Physicists could solve the hydrogen atom. Harder cases, such as nitrogen interacting with oxygen, were supposedly just chemistry. The story came with a caveat in the conversation itself: "I'm not sure actually how accurate it is."
The conversation drew the opposite lesson from the century since. Moving from hydrogen to real materials is not a simple extension, and "the rest is not just chemistry. There's actually a lot of physics there."
High-temperature superconductivity was the main example. Superconductors carry electric current with no resistance below a critical temperature. The discussion explained that conventional superconductors are understood. Their behavior comes mainly from electron-phonon coupling, the interaction between electrons and vibrations in the crystal lattice. One sign is the isotope effect. If you make the same material with a heavier version of one element, the superconducting temperature drops by the amount the theory predicts. Cuprates are copper-oxide superconductors that work above 77 kelvin, some around 93 kelvin. They do not follow that physics. "But we don't know what physics they obey."
The gaps go beyond that famous puzzle. Even some simple quantum systems with strongly correlated electrons, whose behavior is tightly linked, still cannot be simulated. The main simulation tool, density functional theory, was described as "incredibly accurate on some things," but it misses some effects once strong electron correlation appears.
The physicist Phil Anderson's phrase "more is different" came up as the deeper reason. You can understand the basic laws governing individual pieces, the argument went, yet large collections of those pieces can behave in qualitatively different ways. The conversation drew a parallel to deep learning, where power laws show up everywhere and are not understood, much like the emergence and universality physicists see in matter.
What density functional theory does
Later the discussion turned to how Periodic uses computer simulations in its decisions, especially when simulations and experiments disagree. The answer began with a division of labor. Some quantities are easier and more accurate to get from simulation. Others are better measured in the lab. Experiments are generally more accurate, but some measurements are so hard to make that the experimental data ends up worse than a calculation.
Density functional theory, or DFT, was called "probably the most commonly used simulation method for materials." The explanation went like this. Exact quantum mechanics is very expensive, because the mathematical description of the electrons grows exponentially as the system gets bigger. Walter Kohn and Hohenberg showed that, for a system's lowest-energy state, or ground state, you can sidestep that problem. All the ground-state properties are determined by the charge density, the distribution of electrons across ordinary three-dimensional space. Kohn later published a practical method with Sham, and he received a Nobel Prize for the work. Periodic uses DFT, as does the wider field.
Periodic uses DFT to estimate a material's formation enthalpy, essentially how much energy the material holds. That number shows whether a candidate is likely to exist at all. If the atoms could settle into other compounds with lower energy, the material probably cannot be made. In the conversation's terms, a stable material sits "on the convex hull" of competing materials. Picture plotting the energy of every possible compound made from the same elements and stretching a floor under the lowest points. A material on that floor cannot lower its energy by breaking apart into its neighbors.
The conversation then summarized the trade-off. DFT turns an exponentially hard problem into an approximate one whose cost grows roughly with the cube of the system's size. That makes quantities computable that otherwise would not be, but it introduces some error. "If you could do it perfectly," the remark went, "we wouldn't need a lab." The conversation also recalled Heather Kulik, a guest on Latent Space's second episode, who said there is no equivalent of AlphaFold, DeepMind's protein-structure predictor, for materials, because reliable ground truth about crystal structures does not exist.
The reply drew a sharper line. DFT, it said, "doesn't have to be an approximation"; the Hohenberg-Kohn theorem shows it can be exact. The trouble is that nobody has access to the exact form of a key term, the exchange-correlation functional. For models built only on charge density, nobody knows the kinetic energy functional either. Then the bigger point: "even if we had perfect DFT, I still think we'd need a lab. Because even if we have perfect DFT, we cannot fit 10 to the 23 atoms in the computer." Real samples contain roughly that many atoms.
That drew teasing and pushback ("you're very obsessed with that"). The counterargument was that computing power keeps growing exponentially while DFT's cost grows only with the cube of system size. If you waited long enough, you could in principle calculate such systems. The exchange ended without a resolution. Fedus noted that even the charge density is "still quite a big file."
Millions of calculations, a few hundred experiments
The next question was about scale. Computing power is easier to bring online than new labs. DFT could run millions or hundreds of millions of calculations, while even a high-throughput lab might manage hundreds or thousands of experiments a day. How do you stop AI models from leaning too heavily on the simulations? Fedus replied that Periodic is scaling up its lab as well.
The fuller answer was that both people and the language models Periodic uses know where DFT falls short. One gap is microstructure. DFT usually simulates a perfect crystal, but real crystals are imperfect, and their structure at intermediate scales can strongly affect their properties. Another gap is that some properties, such as a superconductor's critical temperature, cannot be easily simulated by DFT. Even 100 million DFT runs would not be enough, the answer went, so the team relies on heuristics and experiments as well. "There's a huge filter from all those calculations to actually what gets executed in the lab," Fedus said.
That filtering can be a mix of people and language models. Calibration is part of it. The discussion pointed to the Materials Project, described as for a long time the best open-source DFT database, which used experimental data to calibrate its DFT results. Periodic does the same internally: for each chemical system it cares about, it gathers both experiments and simulations and calibrates one against the other.
The Materials Project's documentation supports that account. It describes corrections fitted against experimentally measured formation enthalpies. Those corrections depend on chemical environment, so oxygen in an oxide is treated differently from oxygen in a peroxide. The documentation also notes a mismatch: the calculations generally describe zero-temperature, zero-pressure energies, while reference experiments are often reported at room temperature. A 2021 study, whose corrections entered the database that year, shows calibration has limits of its own. The researchers fitted 22 corrections using 222 compounds. Even after correction, the mean absolute error was 51 milli-electronvolts per atom on that same set. When the researchers fed the uncertainty in the corrections into phase diagrams of scandium, tungsten and oxygen compounds, borderline compounds could switch between being labeled stable and metastable.
That fits how DFT's place at Periodic was described in the conversation. "I personally don't know of a better method to predict the stability of a new material. It's definitely not perfect," the reply went. Simulations "will never be enough by themselves," but a loop of simulations, AI and experiments could make progress "much faster than before."
Would a quantum computer change this?
Near the end, the conversation turned to whether quantum computing, described as "massively, embarrassingly parallel compute," would change things. The first answer veered into reversible computing, which in principle needs no energy because it never deletes information. It came with a frank "I'm not an expert."
The conversation then recalled an exchange with Elad Gil, who was described as very skeptical about quantum computing: "even if you had it today, there's no applications." Breaking encryption was offered as the obvious exception.
A reply agreed with the skepticism and applied the same argument used against perfect DFT. People say a quantum computer would simulate materials much better. Grant that, the reply went, and ask what you would simulate. "You're still simulating perfect crystal." A quantum computer would still have the problems DFT would have even if DFT were perfect. Because computing in quantum logic is so exciting, the argument went, people are "glossing over what it would actually do when it's made." The reply then hedged: "Maybe that's fine because maybe once it's made, it will do amazing things we can't even imagine."