Inside a $300 million bet on AI for physical R&D

The sober view a year earlier rested on training data, and that problem is worth restating precisely because it has not gone away. Machine learning works best on the distribution it was trained on, while science and technology care almost exclusively about the opposite: predictions about things unlike anything in the training set. Language models have the internet to learn from; materials discovery has something on the order of thousands of data points. Cubuk not only says this is still true, he speculates that there may be an undiscovered law bounding how hard out-of-domain discovery is, in the family of thermodynamics or Landauer’s limit on the energy cost of erasing information. He labels that a guess.

What changed is an opening rather than a fix. Reasoning models demonstrated that spending compute at inference time improves results, which was the first way he had seen to invest resources in a direction other than more training data. Those models then went on to win gold medals in math olympiads and to do well on coding and physics problems. But Cubuk draws the limit himself, and it is the sharpest line in the episode: you can practice for a math olympiad by studying previous years’ problems, and you cannot practice discovering the next big theory. He says plainly that he is not claiming the models are good enough yet. What the results show is that reasoning over complex problems is improving, which is directional evidence rather than a solution.

That is why the company is built around a physical laboratory, and why the second change matters as much as the first. If a model cannot reason its way past its training set, it has to try things and be wrong often, which is close to how solid-state chemistry and physics have always worked: many trials, a fair amount of accident, resting on deep understanding of the system. Robotic experimentation that mixes powders and liquids and passes samples on for measurement has become close to commoditized, which makes those trials affordable. The loop is simple: the model proposes a synthesis recipe or a simulation, runs it itself because models have become good at tool use, gets back results nobody has seen, and revises the next step using that plus its prior training and the literature. The weak link is automated characterization, the step that tells you what you actually made. Cubuk says that is difficult for AI tools today and that he expects to improve it quickly, which is a forecast, not a capability they have.

Superconductivity is the flagship target, and he defends the choice on grounds other than tractability. A discovery pays off before any product exists: it changes how physicists think, enables experiments that were not possible, and sits directly under fusion and quantum computing. Chasing it also forces capabilities valuable in themselves, including automated synthesis, automated characterization, and predicting high-temperature superconductivity in the absence of a theory for it; his analogy is that OpenAI and DeepMind were mocked for prioritizing artificial general intelligence and built a great deal of useful technology on the way. It has a training virtue too. A critical temperature measured in a real experiment, meaning the temperature below which a material conducts with no resistance, is very hard to reward hack in a way a simulated result is not. And the target is not only room temperature: a higher critical magnetic field may matter more than temperature for fusion magnets, and a superconductor ductile enough to be made into devices would matter as well. The original limitation survives all of this. Cubuk names hypothesis generation as the thing AI is still not better at than humans, leaving two paths, improving the models at it or having humans supply hypotheses while AI executes. Periodic is not prioritizing full automation, and which mix works is, in his words, an empirical question.