Room-temperature superconductors can't be computed by AI in one shot; they require hard lab testing
Even perfect density functional theory can't fit 10²³ atoms into a computer; scientific discovery by definition lies beyond training data—this is why even next-generation models must rely on repeated lab experimentation.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
When experimental environments change, standard reward functions no longer transfer
Periodic's reinforcement learning comes directly from physical experiments, where data lacks ground truth, filled with noise, instrument error, and incomplete telemetry—for instance, the furnace reads one temperature but is actually off by a significant margin. While mathematical or code optimization achieves high precision, materials discovery offers no such luxury: samples leave the furnace unlabeled, and even the labeling process itself introduces randomness. This demands a different reasoning strategy than digital environments: making decisions under extreme uncertainty while using limited data efficiently, because you can't arbitrarily increase rollouts like you can in simulation.
— Liam FedusPhysics remains far from solved, especially for high-temperature superconductors
Dirac famously said after finishing his quantum mechanics textbook that ‘the rest is chemistry,’ but the past century proved otherwise—the interactions between nitrogen and oxygen are far from simple extensions of hydrogen. Large swaths of physics remain unsolved: copper oxide superconductors, for instance, don't follow conventional electron-phonon coupling theory (the framework you'd verify via isotope effects), yet no one knows what physics they obey. Density functional theory fails whenever strong electronic correlations appear. This is precisely why Periodic bets on AI entering materials science: theory, computation, and experiment all still have vast gaps to fill.
— Ekin Doğuş ÇubukEven perfect density functional theory cannot replace the experimental lab
DFT compresses the exponentially growing Hilbert space problem in quantum mechanics into a manageable calculation of charge density as a three-dimensional object, making the previously incomputable tractable. But practical DFT remains an approximation—the exchange-correlation functional and kinetic energy functional still lack exact forms. More critically, even with theoretically perfect DFT someday, you couldn't fit 10²³ atoms into a computer, so the lab remains unavoidable.
— Ekin Doğuş ÇubukEvery instrument needs intelligent decision-making embedded in real-time operation
As experiments scale up, technicians watching electron microscope images to assess morphology quickly fall behind pace, and data from simple automated image-capture programs is ‘too dumb’ to be useful. Periodic embeds AI directly into each instrument, giving it knowledge of the experiment's intent, what's being synthesized, and what evidence exists so far—so it makes smarter decisions while collecting data in real time. This produces higher-value data for downstream AI training and computational prediction. Latency is a hard constraint: if model inference can't keep pace with the physics timescale itself, the system fails—like self-driving cars.
— Liam FedusFull automation isn't the goal; humanoid robots would actually slow progress
Periodic believes deploying humanoid robots to solve lab automation would actually delay reaching the real objective. The more realistic path is human-machine partnership: skip the months-long automation engineering for high-difficulty manual tasks humans do well; prioritize routine operations that consume scientists' and technicians' time and are easy to automate. Labs actually need high-volume, high-quality, diverse data; autonomous operation itself was never the goal, only a means to that data objective.
— Liam FedusLiterature systematically publishes only positive results; accumulating negative samples is essential
Training a classifier requires negative samples—you can't train on positives alone. Materials science literature carries systematic bias: people publish crystals they synthesized, almost never their failures, and you can never confirm whether a crystal is ‘impossible to make’ or simply ‘no one found the right method yet.’ Periodic runs its own experiments and collects large numbers of negative results with full context, training classifiers directly on them. More importantly, they're making the entire scientific process—including the failed iterations leading to success—into training data.
— Ekin Doğuş ÇubukNo model can derive room-temperature superconductors in zero shots
Even if future frontier models were vastly stronger than today's, they'd still need to run experiments to get results, because machine learning excels at what it trained on, and scientific discovery by definition lies beyond the training distribution. This is why Periodic is building a lab: to let open-source and proprietary models actually ‘manipulate the universe.’ The zero-shot leaps you see in math and theoretical computer science don't directly translate to the physical world.
— Ekin Doğuş ÇubukSynthesis, not theory, is the actual bottleneck in materials discovery
Take magnesium diboride (MgB₂): Japanese researchers found it's the highest-temperature superconductor at normal pressure—not through theory prediction, but by trying roughly 30,000 materials, with 30 showing interesting superconducting signals. Yet BCS theory existed since 1957, and the precursor materials had sat on shelves for decades. Periodic's judgment: it's not hard to think of which chemical spaces might harbor superconductors; hard is actually synthesizing them. Scaled-up lab automation compresses the experiments one scientist might do in a lifetime into a month—essentially expanding the area where ‘luck’ can strike.
— Ekin Doğuş ÇubukIn their own words · checked verbatim
No amount of rereading that textbook or paper, thinking, or disappearing into a room is going to allow you to think through all possible experimental outcomes, expected and unexpected, and you really need this iterative process.
Liam Fedus0:00
It’s not enough just to do optimization against, some answers that were known in some papers or textbooks because we’re going beyond that, and I think that’s one of the biggest differences.
Liam Fedus2:49
But then higher-temperature superconductors like cuprates, like the ones that are above 77 kelvin, like 93 kelvin, turns out don’t obey that physics. But we don’t know what physics they obey. They’re just incredible superconductors.
Ekin Doğuş Çubuk14:50
And then even if we had perfect DFT, I still think we’d need a lab because we. Even if we had perfect DFT, we cannot fit ten to the twenty-three atoms in a computer.
Ekin Doğuş Çubuk33:42
But ultimately what we want from the lab is a huge quantity of data, high-quality data, diverse data, and those are our goals. And full autonomy is a non-goal. It’s sort of in the service that we use automation in the service of achieving the goals on the data.
Liam Fedus51:39
machine learning is really good at what it’s been trained on but scientific discovery is almost by definition what you haven’t been trained on.
Ekin Doğuş Çubuk1:00:20
There’s not going to be like. No one’s going to zero shot the room-temperature superconductor.
Liam Fedus1:00:20
I think they tried 30,000 different things. Thirty of them seemed to host interesting superconductivity, and MgB₂ was one of them.
Ekin Doğuş Çubuk1:21:59
Figures
| Copper oxide superconductor critical temperature | approximately 93 Kelvin (above 77 Kelvin) | 14:50 |
| Atoms difficult to store in a single lab | 10²³ atoms | 33:42 |
| Materials tried before MgB₂ discovery | approximately 30000 materials, 30 showed superconducting properties | 1:21:59 |
Glossary
- DFT (Density Functional Theory)
- Approximation method using charge density instead of wave functions to compute ground-state properties of quantum systems
- XRD (X-ray Diffraction)
- X-ray technique for measuring atomic spacing and generating a structural fingerprint of crystals
- Matter Compiler
- System metaphor that compiles target material properties into concrete synthesis pathways
- Landauer limit
- Theoretical minimum thermodynamic energy required to delete one bit of information
- Synthesis Superintelligence
- Periodic's mission vision, emphasizing synthesis capability over pure theoretical computation
- Phase transition
- Abrupt change in material structure or state under specific conditions, such as ice melting
How to listen
Engineers and investors tracking AI for Science, materials discovery, automated labs, and embodied AI who want to understand where LLMs will deploy next.
13:27–16:41 Physics theory background (DFT approximations, strongly correlated electrons, Kolmogorov complexity analogy) is technical; engineering-background readers may skip.