01 Approach
The method behind an AI scientist.
A model forms a hypothesis, an autonomous laboratory runs the real experiment, and the result comes back from nature to sharpen the next attempt. This is how discovery works here: not a search across text that already exists, but a loop that produces the data as it goes.
Hypothesize and design, run and observe, learn and iterate. One engine, turned again and again, faster each time around.
02 The premise
Intelligence alone does not move science.
A model can reason with great fluency and still discover nothing new. Reasoning rearranges what is already known. Discovery needs something the model cannot supply on its own: contact with a world that is free to disagree with it.
The last decade of scientific AI leaned almost entirely on models trained on the text of the internet. That is a finite resource, and the leading systems have now largely read it through. The results that matter next are not written down anywhere yet. They have to be produced, one experiment at a time, and read back from the physical world.
So we do not stop at a smarter model. We build the place where a model can act, on real instruments and real materials, and then we let reality return the verdict. Kept in full and studied closely, that verdict becomes the teacher. New knowledge is made only when an idea is tested against something that can push back.
03 The AI scientist
What the AI scientist does.
Nine things, and none of them optional. Together they add up to a system that does not just describe the world but proposes changes to it, carries them out, and revises itself on the evidence that comes back.
Forms hypotheses
Proposes what to try
It turns a goal into a specific, testable prediction. Given a property worth chasing, it names the materials or conditions most likely to deliver it, and it says so precisely enough to be proven wrong.
Designs experiments
Plans the test
A hypothesis is only as good as the test that settles it. The model chooses the conditions, sets the controls, and decides exactly what to measure, so that the answer will be clean rather than ambiguous.
Runs them autonomously
Acts in the world
The plan becomes instructions for real instruments. The model does not stop at recommending an experiment; it commissions one, on physical materials, and waits for the world to answer.
Learns from results
Updates its beliefs
Every outcome changes what the model expects next. Data folds back in the moment it arrives, so the next hypothesis is built on the last measurement rather than on a static snapshot of the past.
Keeps the negative results
Values every outcome
Failed experiments are rarely published, yet they carry real signal about where not to look. Nothing is discarded here. A run that did not work still narrows the search for the run that will.
Makes data that exists nowhere else
Builds the missing dataset
Each experiment produces measurements found in no paper, dataset, or archive. Over many cycles this becomes a body of evidence that no amount of reading could assemble, and it belongs to the loop that made it.
Works at the gigabyte scale
Measures deeply
A single experiment can return gigabytes of signal. The model reads all of it, not a headline number, so subtle structure that a human summary would flatten stays available for learning.
Simulates before it builds
Models first
Before any hardware is touched, candidates are screened in simulation. Obvious dead ends are ruled out cheaply, so the scarce, slow resource of real lab time is spent only on tests worth running.
Closes the loop
Iterates fast
Hypothesis, experiment, measurement, revision, and then the next hypothesis. The model runs the whole cycle end to end and comes back around faster, and better informed, than the turn before.
04 The discovery loop
The full discovery loop.
Three movements and roughly twelve steps, from a question to an answer to the next question. The model designs, the lab runs continuously, and the physical world decides. Then the whole thing turns again.
Hypothesize and design
Run and observe
Learn and iterate
High signal-to-noise
Effects in physics are often large and clean, so a real result stands out from the background instead of hiding in it.
Fast turnaround
Experiments run and resolve quickly, which means more turns of the loop and more learning per unit of time.
Simulatable
Candidates can be modeled before they are built, so the loop spends real lab time only on tests worth running.
Verifiable
A physical result can be confirmed the way a proof or a program can, which keeps the model honest.
05 Why physics first
We begin where reality answers clearly.
Not every field gives clean feedback. We start in the physical sciences because it is the domain where an experiment settles a question fast, and where the answer can be trusted.
The signal is high, so a real effect is hard to mistake for noise. The work is fast, so the loop can turn many times before the trail goes cold. Results are simulatable, which lets the model rule out dead ends cheaply before committing hardware. And physics is verifiable: a result can be checked against reality the way a proof or a program can, rather than debated. Those four properties are what make a closed loop of hypothesis and experiment actually converge, and it is why we chose to start here rather than somewhere the world answers slowly or ambiguously.
06 Go further
See where the method runs.
The loop is only real because there is a laboratory underneath it. See the facilities and programs that carry it out, or tell us what you want to test against reality.