When Periodic Labs began scaling up its materials lab, the bottleneck was the machines themselves, or rather the people needed to run them. Technicians and scientists sat at instruments, looking for particular shapes and textures in what the lab had made and checking whether the result matched what they had intended to make. As the lab grew, Liam Fedus said, "it just wasn't keeping up."
Fedus, of Periodic Labs, explained the company's approach to lab automation on the Latent Space podcast, in an episode published October 8, 2026, with hosts swyx and Brandon. His colleague Ekin Doğuş Çubuk also took part. Periodic's goal is to discover new materials by letting AI systems learn from real experiments. That means making things, measuring them and deciding what to try next. Fedus's account of how the company automates that work is pragmatic. It does not start with a robot that does everything. It looks for the step that is holding things up, automates it, and then looks for the next one.
A microscope routine that was "a little too dumb"
The first fix was ordinary scripting. Periodic wrote a program to capture images on its scanning electron microscope (SEM), an instrument that scans a sample with a focused electron beam to image its surface at very high magnification. The routine was simple: take a field of view, capture an image, zoom in, capture again.
The data it produced, Fedus said, "was just not all that useful. It was a little too dumb for what we really wanted to get." A real SEM session involves many choices. According to Thermo Fisher's explanation of the technique, magnification, field of view and the choice of detector signal all change what you learn. One signal mainly shows surface shape, another gives contrast linked to chemical composition, and X-rays given off by the sample can support elemental analysis. The scripted loop captured images, but it did not take into account what the experiment was meant to show.
So Periodic put AI systems directly on the machines. Those systems now control the instrument with "the full context" of the experiment, Fedus said: what the experiment was for, what the team was trying to synthesize and what other experimental evidence existed. Then they decide what to capture.
This is the idea behind the line the conversation picked up, that every piece of equipment in the lab will have "140 IQ." The point is not clever hardware for its own sake. Experiments are full of hidden variables that nobody writes down, and an instrument that captures data intelligently while the experiment is happening leaves a better record. That record matters for "future AI systems and future computational predictions," Fedus said. "You'll have just a richer set of data."
The limits of putting a model on every machine
The conversation turned to the state of the art in putting intelligence on every device, and Fedus started with latency, the delay between asking a model something and getting an answer. Physical processes have their own timescales. If calling out to a model through an online interface is too slow, he said, so that "the amount of reasoning or tokens or tool calls is just not matched to the latency of that actual process, then that's infeasible." The discussion noted that the delay comes from two places: the trip to a cloud server and the computing power available on the device.
Someone in the conversation was surprised, since many of Periodic's experiments take hours. Fedus agreed that experiments "do take, you know, many hours, days, but there could be particular steps where latency would matter." The discussion also turned to cost, since the same slow, heavy reasoning that delays a model's analysis also makes it expensive, and to human patience. Waiting two hours for an analysis of an X-ray diffraction (XRD) pattern, the fingerprint a crystal leaves when X-rays bounce off its atoms, is a very different experience from getting it in two minutes.
Fedus added that reasoning here is rarely a single question and answer. A model may run simulations or do deep research in the middle of its work. "It's not a single tool call. It's not a single inference thread," he said. "So then the latency can blow up quite a bit." How much of that time goes to tools rather than to the model's own reasoning, he said, is "really process dependent." Periodic's own infrastructure write-up describes the same pressure from the training side. Scientific agent runs can last hours, so the company runs training and inference on separate pools of GPUs, which keeps a slow, tool-heavy run from holding up every training update.
Why not just buy a humanoid?
The conversation then turned to the strategy itself: whether automating a human-run lab piece by piece is a trap, a local optimum or "short horizon optimization," when "actually, you should get a humanoid and just put them in there."
Fedus disagreed. "I think we're in the opinion that solving humanoids would actually be slower to kind of getting to some of our goals," he said. He described a different process. A mix of humans and machines shows quickly where the experimental process gets stuck, and those bottlenecks can then be cleared one at a time. If a fiddly dexterity task is something people do well but a machine would take months to learn, "maybe don't spend a ton of time there." The better target is "something very routine that takes up a huge amount of scientists and technicians time and it's easy to automate."
That reasoning follows from what the lab is for. "Ultimately what we want from the lab is a huge quantity of data, high quality data, diverse data," Fedus said. "Full autonomy is a non-goal." Automation serves the data goals, not the other way around.
Characterization came first
The step Periodic targeted early was characterization: working out what a batch of material actually turned out to be. In the conversation, the starting hypothesis was laid out simply. Mixing powders to try things is not that hard. But a lab gets little from mixing powders at random unless it can analyze the results and decide intelligently what to do next. So Periodic focused on automatic characterization and on connecting it to simulations, so the lab could try many things, make an informed decision and choose the next day's work.
Fedus said that in the early days characterization was human-driven and "the scientists were very much overwhelmed." Since then, "the balance has shifted significantly towards AI driven," which has let scientists who once spent all their time on these analyses move on to other work. Guiding whole research campaigns, the series of experiments aimed at one goal, was once "fully human driven," he said. Now it is "a mix of AI driven and human driven." Asked where humans fit in the loop, he said the split is "always changing."
The discussion also sketched the levels at which this work happens. At the bottom are models of individual atoms; above them are continuous models of a material as a whole; there are thermodynamics, which asks where a reaction would end up if left forever, and kinetics, which asks how fast it actually gets there. Fedus added another level: a theoretical advance that could shape several campaigns at once.
This matches what the company said publicly. In its September 15, 2026 announcement of Neon, a model Periodic says has a trillion parameters and was post-trained on laboratory data to analyze experiments, the company said it is extending the same approach to directing campaigns, developing synthesis recipes and choosing the next experiments.
An AI that caught a loading mistake
Fedus said robots may simply make fewer mistakes than people. AI that can see all of the lab's data helps catch the mistakes that remain. At one point, one of the lab's steps had a cyclic error: a machine had been loaded incorrectly, so the patterns were inconsistent with what had been run. Reading through the data over time, the AI system flagged that, "given what was run, this is not expected." It then found that if it applied a cyclic permutation and reversed it, "everything is consistent."
The team went back to the physical setup and confirmed that a mistake had been made in the loading. Fixing the operating procedure was one answer. Fedus saw something else in it: "this is another great opportunity for automation. How do you make this just so reliable, so durable that we never have those types of mistakes again?" Managing data quality, he said, "is so foundational to doing AI in the physical world."
Building its own hardware
The same habit of hunting for bottlenecks has pushed Periodic into building its own instruments. Early on, the company bought off-the-shelf equipment for speed. As it pushed the lab's scale, Fedus said, "we have to make our own." In one instrument, he said, the components were pretty quick, but the weighing machine became the bottleneck.
Other problems were about data quality, not speed. In one design, the robotic arm's resting position sat above the plate of previously mixed materials, which created a contamination risk. Periodic's redesign moves the resting position away from them. "These are the kind of subtle details that allow us to kind of push the noise floor down for our experimental campaign," Fedus said.
He said the company has been proving out this loop in its initial labs in Menlo Park, California. With new resources, it plans to scale those labs up, add computing power and build new types of labs that repeat the same process.