01 About
A company built to make knowledge, not just model it.
Rafati builds an AI scientist together with the autonomous laboratories where it works. The model forms hypotheses, runs real experiments, and learns from what the physical world returns. We exist for one reason: to move science forward by putting ideas in contact with reality, starting in the physical sciences.
Founded 2026 in Los Angeles, California. Backed for the long term, and built to close the distance between a question and a measurement.
02 The mission
Why we exist.
Intelligence alone does not move science. A model can reason well and still learn nothing new, because reasoning only rearranges what is already known. New knowledge appears at a single moment: when an idea is tested against physical reality and the world answers back.
That answer is the thing we are after. So we do not stop at better models. We build the places where a model can act, on real instruments and real materials, and we keep every result it produces. The model proposes, the laboratory performs, and reality decides. Rafati exists to run that exchange at scale, again and again, until the dataset holds signal found nowhere else.
03 Why now
Four things are true at once.
The idea of learning from experiments is not new. What is new is that all of the pieces have arrived in the same decade, and each one sharpens the others.
The internet ran out
A decade of scientific AI leaned almost entirely on models trained on the text of the web, a finite resource the leading systems have now largely exhausted. The next breakthroughs will not be found on a page. They have to be produced.
Physical domains are verifiable and fast
We begin in the physical sciences because the signal is high, the work moves quickly, and results can be simulated ahead of time. Best of all, physics is verifiable: a result can be confirmed the way a proof or a program can, with no room for wishful reading.
Autonomous labs have matured
Robotic handling and precision sensing are now reliable enough to run experiments continuously, around the clock, with no queue and no idle bench time. A model can finally reach into the world through instruments instead of only predicting what they might show.
Models can reason about experiments
The models are now good enough to turn a goal into a testable prediction, design the experiment that settles it, and update the moment results arrive. Paired with a lab, that ability stops being a demonstration and becomes a working method.
04 Principles
What guides the work.
These are not slogans on a wall. Each one is a decision we have already made about how the model, the laboratories, and the people around them behave.
Nature is the teacher
The feedback that counts
We treat the physical world, not a benchmark or a corpus, as the source of truth. When the model and reality disagree, reality wins, and the model changes. That is the whole point of building the lab beside it.
Keep every result
Including the failures
Failed experiments are rarely published, yet they carry real signal. We record every outcome, the successes and the negative results others discard, because the shape of what does not work tells you where to look next.
Simulate before you build
Cheap tests first
Before we touch hardware, we model the candidates and rule out the obvious dead ends. Simulation narrows the search and saves real lab time for the questions only a physical experiment can settle.
Verifiable science first
No wishful reading
We start where a result can be confirmed the way a proof or a program can. Verifiability keeps the whole loop honest, so the model learns from what actually happened rather than from a hopeful interpretation of it.
Build for the long term
Compounding, not sprinting
The dataset grows with every cycle and holds value found nowhere else. We are backed to let it compound, and we plan on the timescale that real discovery takes rather than the one a quarter would prefer.
Widen the circle
The field, not just us
We run a grant program for outside researchers, work alongside industry partners, and keep an academic advisory board close, so more people can test ideas against reality faster, and the science stays both honest and ambitious.
05 Who is building this
The people behind the loop.
Rafati was founded in 2026 and is built by people who have co-created widely used AI systems, contributed foundational machine-learning techniques and materials-science models, scaled autonomous physics laboratories, and taken part in real materials discoveries.
It is a rare combination on purpose. Closing the discovery loop takes both halves at once: the engineering to run a laboratory continuously and the scientific judgment to know which experiment is worth running. Fast-moving engineering is paired with the depth of the research community, and an academic advisory board keeps the science honest and ambitious as the work speeds up.
We are backed by prominent venture investors and individuals who have committed for the long term, which is the only horizon on which this kind of discovery pays off. That patience is not a footnote. It is what lets the dataset compound instead of being harvested early.
06 Go further
See how the method works.
If any of this resonates, there are two good next steps: read how the discovery loop actually runs, or start a conversation with us directly.