Ekin Dogus Cubuk
Related people
Episodes
Notes
Latent Space: The AI Engineer Podcast · 🔬 Synthesis Superintelligence: from Semiconductors to…
-
What Does It Mean for Lab Equipment to Have 140 IQ?
Liam Fedus says lab equipment with a 140 IQ is an instrument that runs its own AI, because people checking every sample by eye cannot keep up once the lab scales. Technicians and scientists were looking at morphology to see whether a run matched the intention. That review became the bottleneck. A simple program that grabbed a field of view and zoomed in was too crude to be useful. The useful step was an AI system on the machine, controlling the capture instead of dumping frames for someone to interpret later.
The limit is latency. A full experiment can take many hours or days, but one step on the instrument can be much faster. If a call to an API, or a run of tokens and tool calls, is slowe
-
What Is a Matter Compiler?
Liam Fedus says a matter compiler is a lab that takes a requested material and works out a process that can make it, the way a chip fab builds a design it did not invent. The property you want and the process that produces it can be split. One side asks what the atoms should do. The other asks how to make that configuration. A perfect reverse of every process is unlikely, because a material contains more atoms than anyone can record. The practical aim is to reverse enough of the process to improve the material.
Given a set of requirements, the question is whether that arrangement of atoms can exist, and which tools, experiments, and earlier papers get there. The check is replication. Did th
-
What Is an AI Scientist Lab?
Liam Fedus says an AI scientist lab joins AI systems, simulations of the physical world, and a high-throughput experiment, because thinking alone cannot create new knowledge. The universe is too complicated for a person to reread the textbook, disappear into a room, and think through every experimental outcome, the expected ones and the unexpected ones. A conjecture has to be checked against what the experiment actually does.
Periodic Labs was built around that loop. The model, the simulation, and the physical lab were in the plan from the start. A different kind of intelligence emerges when those three sit in one place, instead of a digital setup that only optimizes against answers already
-
Why Did Alex Müller Win a Nobel Prize for Cuprates?
Ekin Dogus Cubuk says Alex Müller won a Nobel Prize because cuprates worked after his nickelates did not. Müller thought transition metal oxides might be a path to superconductivity. He tried nickelates, nickel and oxygen plus some cations, and that did not work. He tried cuprates, copper and oxygen plus some cations, and that worked.
Nickel and copper sit next to each other on the periodic table, and there are only so many 3D transition metals. Nickelates later turned out to be a good idea. Harold Hwang, a Stanford professor working with the lab, made nickelates as a thin film and showed the superconductivity. Chemical spaces that might host superconductivity are not hard to think of. What
-
How Does Periodic Labs Work With Semiconductor Companies?
Liam Fedus says Periodic Labs works with semiconductor companies by sending its own engineers and researchers on site, into very private and secure labs, to integrate the company's AI and computation with the partner's process. The lab has been its own first customer, building the tools for its own materials discovery and materials engineering. It is now taking that know-how into other industries, with a huge focus on semiconductors.
Those forward-deployed engineers run inference locally and can also train on the partner's data, so the partner owns the intelligence built from that data. Fedus says that beats handing over an API or an untrained model and leaving the partner to figure it out.
-
Why Is There No Fast Test for Superconductivity?
Ekin Dogus Cubuk says there is no fast test for superconductivity. Unlike X-ray diffraction, there is no high-throughput superconductivity measurement, and each one can take one hour.
Combinatorial sputtering is the shortcut people have already tried. Sputtering targets lay down a compositional gradient, different crystals form in different places, and one run tries hundreds or thousands of crystals. As far as he knows, it has not worked very well, because diffusivity lets the material move around after it is deposited.
Pooling candidates into one sample is still only an idea. The team has thought about mixing materials and then chasing which ingredient produced a superconducting signal, a
-
What Is a Null Result in Materials?
Liam Fedus says a null result in materials can mean the team intended to produce a structure and every measurement says they did not. Negative controls are part of the experiment. Sometimes the point of the run is to show that a structure was not made. An impurity phase can kill the property being chased, and it can spoil the result.
A negative run can also be a different finding. The team aimed at one structure, and the data showed a new structure that had not been identified before. The result is negative against the plan and still a discovery. A string of failed attempts that finally reaches a known material is process data, the record of how the synthesis was changed until it worked. Ev
-
How Did Periodic Labs Catch a Bad Machine Load?
Liam Fedus says Periodic Labs caught a bad machine load because an AI system with a full view of the lab data noticed a pattern no single run made obvious. One step had a cyclic error. The patterns were inconsistent with what had been run. The system read through the runs, judged the result unexpected, and looked at the basket of data over time. Reversing a cyclic permutation made the runs consistent. The team went back to the physical equipment and confirmed that the machine had been loaded incorrectly.
He wants that lab to produce a huge quantity of data, and for the data to be high quality and diverse. Robots may make fewer mistakes than people on the routine steps. Managing data quality
-
Why Can't Density Functional Theory Replace Experiments?
Ekin Dogus Cubuk says density functional theory cannot replace experiments, even at a hundred million trials. The method usually simulates a perfect crystal. Real crystals are not perfect. Microstructure, the arrangement at medium range, can change the properties, and those trials never test it. Superconducting temperature is another property the method cannot simulate easily. At that scale of calculation, the work still falls back on heuristics and on the lab.
People and the models already know those limits, so a filter decides which calculations ever run for real. For a long time the Materials Project, which he calls the best open-source density functional theory database, used experiment
-
Why Is Materials Data Never Deleted?
Ekin Dogus Cubuk says materials data is never deleted because any measurement the team can collect is precious. A person cannot watch every atom. If every position were visible, the experiment would hold more data than a computer can store, but that full record is not available. What the instruments return is already a thin slice, so throwing a reading away means losing something the lab cannot recapture.
Brandon Anderson compares the measurement to linear probing of a foundation model. The material is the large model. The experiment is a tiny probe, about eight bits of output, and a human or an agent has to rebuild what is happening inside from that coarse reading.
-
What Does More Is Different Mean?
Brandon Anderson says Phil Anderson's line, more is different, means someone can know the basic laws and still be wrong about what happens when many parts are put together. The collective behavior is qualitatively different from what the simple laws suggest. The starting laws look simple, so the behavior of many atoms looks like it should be simple too. It is not. Collective behavior is hard to model.
Ekin Dogus Cubuk adds that the behavior is emergent and universal. The same pattern shows up in deep learning, where power laws appear across models and are not understood, yet they resemble the emergence physicists already study. He wants AI applied to materials because theory, computation, a
-
Why Can't Reinforcement Learning Wait on a Superconductor?
Liam Fedus says reinforcement learning cannot wait on a superconductor, because an update that sits for a couple of days to see whether a run found a room-temperature superconductor is completely infeasible. The materials loop has three jobs. First, decide what to make: the atoms have to stay together, and the arrangement has to show the property you want, not merely be a new configuration. Next is synthesis. Even when a structure should hold, the processing conditions that bring it into being are highly non-trivial. Then characterization, because the sample does not come off the furnace labeled.
A reward that waits out the furnace is too slow and too noisy. There are not enough agents to c