01 Labs
The autonomous laboratories where the model learns.
A model can only get smarter about the world if it can touch the world. So we build the benches, not just the software: automated facilities where the AI scientist proposes an experiment, real instruments carry it out, and the measurement comes back as something to learn from. The lab is not a service the model calls on. It is where the model is trained.
Robotic handling, precision sensing, and the model, joined in one loop and kept running. Physical sciences first, because the results are fast, dense, and verifiable.
02 Why autonomous labs
The labs are the center of the strategy, not a convenience.
It would be easier to build a model alone and hope the answers were already latent somewhere in its training. We think that road is closing. Everything below is a reason the experiment, not the archive, is where the next results have to come from.
The internet ran out
Scientific AI has leaned almost entirely on models trained on web text, and the leading systems have largely exhausted what that text can teach. There is no larger corpus waiting. The knowledge that is left to gain was never written down, because nobody has produced it yet.
Data found nowhere else
Each facility produces high-quality experimental data that lives in no dataset, no paper, and no archive. It is measured here for the first time. That is the point of running our own labs: the model trains on records that only exist because we made them.
Negative results, finally kept
Failed experiments are rarely published, yet a failure tells you exactly where the boundary is. We keep every outcome, including the ones a journal would never print. Over time that record of what does not work becomes some of the most valuable signal the model has.
Tools to act, not just predict
A language model on its own can only rearrange what it already knows. Labs give it real instruments: something to move, measure, and change in the physical world. Prediction becomes action, and action is the only thing that produces a genuinely new observation.
Gigabytes per experiment
A single run returns gigabytes of raw signal, dense with detail no summary would preserve. The model reads all of it rather than a headline number. Depth of measurement, not just the count of experiments, is what makes each cycle worth so much.
Built for scale
Automated facilities run continuously, without the queue or the idle bench that slows a human lab. Every hour adds to the record, and the record compounds. The advantage is not any single result but a dataset that keeps growing and holds signal found nowhere else.
03 The autonomous laboratory
Robotic handling, precision sensing, and the model, in one continuous loop.
This is the heart of the operation. Three capabilities that would normally sit in different rooms, run by different people, on different schedules, are joined into a single loop and kept turning.
Robotic handling prepares and moves the samples so an experiment can be set up and reset without a person at the bench. Precision sensing reads the result closely, down to the gigabyte, capturing far more than a human would think to note. And the model sits inside the loop rather than outside it, choosing what to run next the moment the last measurement lands. Because none of it waits on a queue, the laboratory runs continuously, and each turn feeds directly into the one after.
Handling
Robotic preparation and movement of samples, so an experiment can be set up, run, and reset with no one at the bench.
Sensing
Precision instruments that measure each run deeply, returning gigabytes of signal rather than a single reported figure.
The model
Positioned inside the loop, reading every result and deciding what to try next while the facility keeps running.
04 The simulation environment
Model the candidate before you make it.
Real lab time is the scarcest resource we have. Before a material is ever synthesized or a condition ever set, the simulation environment models the candidates and rules out the obvious dead ends.
Physics is well suited to this. Many results can be simulated closely enough to tell a promising direction from a hopeless one, which lets the model screen a wide search in software and narrow it to the handful of experiments worth running on hardware. Simulation does not replace the physical run, because reality still gets the final word, but it decides where that run should be spent. The bench then goes to work on candidates that already survived a first, cheaper test.
05 Programs
The work reaches past our own walls.
The same method is more useful the more people can point it at real problems. Three programs extend the labs outward: one into industry, one to independent researchers, and one back to the academic community that keeps the science honest.
The industry lab
Working with a manufacturer
We work alongside a semiconductor manufacturer on chip heat dissipation, one of the stubborn limits on modern hardware. Beyond the experiments themselves, we train custom agents that help its engineers and researchers interpret their own experimental data and iterate faster on it.
The grant program
Supporting outside researchers
We support outside researchers working at the same frontier, so the field does not have to wait on us to test ideas against reality. The aim is straightforward: put the ability to run and learn from real experiments in more hands, and let good work compound wherever it happens.
The scientific advisory board
Academic guidance
An academic board keeps the science both honest and ambitious. It pairs fast-moving engineering with the depth of the research community, a check that the questions we chase are the right ones and that our claims hold up to people who have spent careers on them.
06 Go further
See what the labs are aimed at, or start a conversation.
If you are a researcher, an industry partner, or a builder who wants ideas tested against reality, we would like to hear from you. And if you want to know where this method leads, the frontiers are where the results are meant to land.