AI FOR SCIENCE · INTERVIEWS
Fedus says Periodic's focus right now is the semiconductor industry: engineers and researchers placed on site, deploying inside partners' secure environments and training models on their data. Cubuk says the hard part of finding superconductors is synthesis, not ideas, and the measurement that decides whether a sample superconducts has no high-throughput method, at about an hour each.
The two earlier interviews left how Periodic earns unsettled and spoke of scaling labs; this one gives what it does now (on-site deployment, partners owning their own models, outcome pricing later), how RL rewards are set on single steps of the loop, that 'full autonomy is a non-goal', and that its lab lead is Joe Checkelsky, an MIT professor on leave to work there full time.
No semiconductor partner is named, no superconductor candidate is given, and there are no new throughput or cost figures; the first paper from its academic grants is 'about to come out' and has not appeared.
Fedus says starting an experiment, waiting days and updating on whether a room-temperature superconductor turned up is 'completely infeasible': one rollout spends days moving between instruments, and there are too few agents to bring the variance down. So they build environments around the data they have: characterization is rewarded for identifying the phases actually present and penalised for spurious or chemically implausible ones; evidence can be cut at a date to ask what the scientist did next; or one instrument, or a few, can be taken alone. Viewed these ways, a finite basket of data becomes more training tasks.
As the lab scaled, the scanning electron microscope (SEM) fell behind: a simple program that took a field of view, zoomed and captured again produced data that was 'not all that useful'. They moved to AI controlling the instrument directly, with the experiment's intent and the evidence so far in its context. Fedus says the goal is a large quantity of high-quality, diverse data, and automation serves that; a dexterity task people do well and a machine would take months to learn is left alone, and humanoids would be slower. Where an off-the-shelf instrument becomes the bottleneck they build their own: a weighing machine, or a robot arm whose resting position sat above mixed samples and risked contaminating them.
“Full autonomy is a non-goal.”Liam FedusCubuk says thinking of chemical spaces that might superconduct is not hard; making them is, which is why they speak of 'synthesis superintelligence'. The measurement that decides whether a sample superconducts has no high-throughput method, unlike XRD, and takes about an hour; pooling candidates into one sample and screening them like COVID tests is something they have thought about and not done. His example is magnesium diboride: by his account, a Japanese group tried about 30,000 materials, thirty showed interesting superconductivity and MgB2 was one, after it had sat on shelves as a precursor for decades. His picture is a group's career of trials compressed into a month.
“We just really increased the surface area for luck.”Ekin Dogus CubukFedus says they built the tools for themselves first and are now taking them to different industries, with a large focus right now on semiconductors. These companies are highly private, so Periodic's forward-deployed engineers and researchers work on site, integrating its systems in the most secure environments, running inference locally and training on the partner's data, so the partner can 'own their own intelligence' rather than call an API. His comparison is coding: from Copilot to Codex, and once automation is strong enough, outcomes can be priced.
Asked whether a future model tuned on these traces could do without the lab, Cubuk says scientific discovery is almost by definition what a model has not been trained on, which is why they build labs that open or closed models can use to 'tinker with the universe'. Fedus says results in mathematics and theoretical computer science have to be kept apart from the physical world, and compressing experience into their own weights still matters. Otherwise frontier labs would have stopped training at GPT-4.
“No one's going to zero shot the room-temperature superconductor.”Liam FedusPeriodic's site says it is helping a semiconductor manufacturer with heat dissipation on its chips, training custom agents for its engineers and researchers to read their experimental data; The Generalist interview (29 September) records how it earns as unsettled: custom systems, or outcome pricing.
Semiconductors are 'a huge focus right now'; the method is engineers and researchers on site, local inference and training on the partner's data, with the partner owning the result; outcome pricing comes once automation is stronger.
The Periodic Neon timeline entry (15 September) records a model post-trained on Periodic's own XRD data and its in-house scores, and says nothing of how the RL reward is set.
Characterization is rewarded for identifying the phases present and penalised for spurious or chemically implausible ones; treating a whole experiment as one rollout they judge infeasible.
The Mercor interview (1 October) gives the superconductor bar, a critical temperature well above liquid nitrogen, and says nothing of what limits the search.
The limit is synthesis, and a superconductivity measurement of about an hour with no high-throughput method; pooled screening is only an idea.
The record names no one at Periodic beyond its two founders.
The lab lead is Joe Checkelsky, an MIT professor on leave to work there full time; Dzmitry Bahdanau leads much of the AI and LLM work; and the company gives academic gift grants to university groups.
On Fedus' account, Periodic's commercial work now is installing tools that read experimental data inside other companies' plants and training on their data, not selling materials it discovered. Its rounds are better read against whether this service line shows named customers and results than against a superconductor that has not appeared. That is our reading.
The one-year goal of about 1,000 experiments a day and XRD runs of 40 samples count synthesis and structural characterization; the measurement that decides superconductivity takes about an hour and has no high-throughput method yet. So any later Periodic throughput figure needs reading for which step it counts.
Fedus says full autonomy is not the goal and automation serves only to get more, cleaner data. Set beside Radical or Lila, Periodic's labs are better judged by data volume, quality and noise than by how close they are to running unattended.
A large focus right now is the semiconductor industry; forward-deployed engineers and researchers work on site, run inference locally and train on the partner's data so the partner owns its model.
Company-reported
Fedus' description; no partner or result is named. The company's site mentions one chipmaker with a heat-dissipation problem.
Once autonomy and capability grow, outcomes can be priced.
Target
No timeline, and no outcome-priced contract is mentioned.
Characterization RL rewards identifying the phases present and penalises spurious or chemically implausible ones.
Company-reported
A description of method; no scores for models trained this way are given.
The SEM is controlled directly by AI, with the experiment's intent and prior evidence in context.
Current automation
Fedus' description; how many instruments this covers is not said.
Reading a batch of data, the AI found that a machine had been loaded wrongly, with results cyclically permuted; reversing the permutation made everything consistent, and the loading mistake was then found on the hardware.
Company-reported
One anecdote, no data; The Generalist interview tells a similar story of samples shifted two places in a holder, possibly the same incident.
With its own data, it can go beyond frontier models at their highest reasoning effort on some tasks, and more compute-efficiently.
Company-reported
Neither the tasks nor any figures are given.
Superconductivity measurement has no high-throughput method and takes about an hour; pooled screening has been considered and not done.
Not yet closed
Cubuk says plainly it has not been done.
Experimental data in the literature usually has a noise floor worse than DFT accuracy; they hope standardised workflows will lower it.
Guest's judgment
Cubuk's judgment, with no figures.
Lab lead Joe Checkelsky is an MIT professor on leave, full time at Periodic; Dzmitry Bahdanau leads much of the AI and LLM work.
Company-reported
The roles are the founders' words; both are on the team list on Periodic's site.
The company gives academic gift grants to university groups, and the first paper from this funding is 'about to come out'.
Company-reported
No group or paper is named.
Even with a quantum computer one would still simulate a perfect crystal, and DFT's problems would remain.
Guest's judgment
Cubuk's view; he says he is not an expert in quantum computing.
No superconductor or new-material result has been published.
Not yet closed
No candidate or measurement appears in the interview.
Whether a semiconductor partner is named and says what on-site deployment produced.
Would show up as The partner's own statement, or a case study naming the task and the result.Whether superconductivity screening becomes high-throughput, by pooling or otherwise, and the time per sample.
Would show up as A screening method and time per sample published by Periodic.Whether the claim of beating frontier models at their highest reasoning effort is shown with a named task and numbers.
Would show up as Scores beside frontier models on the same task, ideally scored outside.What the first paper from the academic grants reports.
Would show up as The paper itself, and whether it used Periodic's data or labs.Read 2026.10.09 The full transcript on Latent Space's episode page, labelled by speaker, read in full; the video was not watched. The transcript misspells several colleagues' names, so this page follows the team list on Periodic's own site. Open the interview ↗Full interview on YouTube ↗