AI FOR SCIENCE · INTERVIEWS
Fedus says science at machine-learning scale needs AI: Periodic's first small lab produced more data than its scientists could read, before the larger lab was even online. He is also clear about today's scope: serious alloy and polymer work will need new labs.
The interview two days earlier said Periodic must build many kinds of lab; this one gives today's limit (no serious alloy or polymer work), 'easily dozens' of labs in five years, how scientists' reasoning is captured (GoPros tried and dropped for voice recordings and lab-notebook uploads), and the superconductor bar: well above liquid nitrogen is enough, room temperature is not needed.
No new figures or results: no lab count, throughput or superconductor candidate is given; 'dozens of labs' is a five-year estimate, and the fusion timeline is a judgment he prefaces by saying he is not an expert.
Periodic first built a small lab to test its earliest predictions and decide which machines to scale. Fedus says what surprised them was how fast it became intelligence bottlenecked: the scientists and engineers said they could not make sense of all the data and needed help, before the larger lab was online. His conclusion is that doing science at machine-learning scale needs AI.
“One thing that was really surprising to us in the construction of our small lab to begin was just how quickly we became intelligence bottlenecked.”Liam FedusAsked how many labs Periodic wants, Fedus says more than it has, because 'our labs today are incapable of doing serious alloys research or polymers research', and different physics needs different lab settings. What carries across labs is how to orchestrate them, how to detect bad data and where to put intelligence across the machines. Asked how many in five years, he says 'easily dozens'.
Fedus says science data online is noisy: two groups running the same experiment reach different conclusions, papers leave out the degrees of freedom that mattered in a synthesis, and negative results barely exist. Simulation only gives a set of hypotheses; collapsing several possibilities into the true one means carrying the experiment out. Periodic built its labs in Menlo Park to get the negative results and check them against simulation and the literature.
“We think this is a non-negotiable.”Liam FedusFedus says giving the model the full context (what the hypothesis was, how it was carried out, what came of it) has consistently been an advantage at Periodic; a twenty-year-old simulation run with no stated intent is hard to use. Early on scientists wore GoPros around the lab, which they disliked; Periodic moved to voice recordings and lab-notebook uploads. It is also putting models directly on the instruments, for monitoring and steering, so the data captured is richer.
Fedus says room-temperature superconductors are not known to be physically possible; Periodic is looking, but does not need to reach room temperature. A critical temperature well above liquid nitrogen, plus manufacturability (ductile, easy to make, less dependent on rare earths), would already matter a great deal, for fusion magnets among other things. He says commercially viable fusion 'feels really plausible in the next 7 years or so', while saying he is not an expert.
“We don't have to hit room temperature.”Liam FedusPeriodic's post of 15 September, Building Labs that Learn, says it has built high-throughput labs in Menlo Park running 24/7, aimed at superconductors and magnets; it does not say what those labs cannot do.
Fedus says today's labs cannot do serious alloy or polymer research, which needs other lab settings.
In The Generalist interview of 29 September, Cubuk says a model from one lab generalises only to similar labs, so Periodic must build many kinds; no number is given.
Fedus says 'easily dozens' in five years, and that what carries across labs is orchestration, detecting bad data and placing intelligence on the machines.
The same interview puts the training unit as the full experiment trace (hypothesis, prediction, execution, result); the public record does not say how scientists' reasoning gets captured.
Scientists wearing GoPros was tried and dropped; voice recordings and lab-notebook uploads are used instead.
The record's goal is 'better superconductors and magnets' (the 15 September post) or 'a new high-temperature superconductor' (the 29 September interview), with no bar stated.
The bar is a critical temperature well above liquid nitrogen, plus manufacturability; room temperature is not needed.
On Fedus' account, the first small lab lacked not experimental throughput but someone able to read the results. That matches the order in which Periodic trained Neon to read XRD first: interpretation was the link filled first. For a company building its own labs, the earliest spending may go to analysis rather than more instruments; that is our inference.
This is about Periodic's own labs, not a claim that self-driving labs cannot do alloys. In The Generalist interview Cubuk describes Periodic's loop as powder solid-state synthesis (mix precursor powders, fire them in a furnace, read the result by XRD), aimed at superconductors and magnets. Radical's lab was built for alloys: it makes button-sized samples, cuts them into test pieces by EDM, and tests strength and oxidation resistance. What a lab can discover is set by its synthesis and test equipment. Any later Periodic result should be read first for whether it falls inside what the current labs can make, or comes from a new lab.
'A new high-temperature superconductor' can be stretched; above liquid nitrogen and manufacturable is a line outsiders can measure. When Periodic puts forward a candidate, its critical temperature and manufacturability can be checked against it directly.
The first small lab was intelligence bottlenecked before the larger lab was online.
Company-reported
Fedus' account; no data volume or timing is given.
Today's labs cannot do serious alloy or polymer research.
Current system
Fedus on the limits of his own labs.
Scientists' reasoning is captured through voice recordings and lab-notebook uploads; GoPros were tried and dropped.
Current automation
Fedus' description; no data on how much this context helps the model.
Models sit directly on the instruments for monitoring and steering.
Company-reported
Consistent with the 15 September post's 'AI embeds in our equipment'; no instrument or case is named.
'Easily dozens' of labs in five years.
Target
An answer to the host's question; today's lab count is not given.
The superconductor target is a critical temperature well above liquid nitrogen and manufacturable, not room temperature.
Target
No candidate material or measurement is mentioned.
RLHF reward models are not sensitive to correctness and encode style and preference; science has no data that ships with tests, as GitHub does for code.
Guest's judgment
His reading from working on ChatGPT, offered as the reason to build labs.
Commercially viable fusion 'feels really plausible in the next 7 years or so'.
Guest's judgment
He prefaces it by saying he is not an expert; it is about others' progress, not a Periodic plan.
No new material result and no outside validation.
Not yet closed
No candidate material or third-party data appears in the interview.
Whether Periodic opens a lab that can do alloy or polymer work.
Would show up as A company post or paper naming the new lab's material systems and equipment.Whether the lab count reaches 'dozens' within five years.
Would show up as A count and list of lab types the company publishes.Checked on Said: 'easily dozens' in five yearsWhether a first superconductor candidate clears Fedus' own bar: a critical temperature well above liquid nitrogen.
Would show up as Resistance and magnetisation measurements for a named material, ideally reproduced outside.Whether context such as voice notes and lab notebooks actually makes the model better.
Would show up as A same-task comparison with and without that context, for example on an evaluation like FrontierXRD.Read 2026.10.05 The full YouTube auto-captions, read in full and checked against the chapters in the video description; names follow the description. The host, Brendan Foody, is Mercor's CEO and the show is on Mercor's own channel; the first half is about OpenAI and ChatGPT, and what has no bearing on science is left off this page. Open the interview ↗