building AI capable of autonomous scientific discovery

Reinforcement learning has taken models remarkably far in math and code, where an answer can be checked and a model can learn from ten thousand attempts. Science offers no such reward function. We build the environments and datasets that supply what is missing: expert scientists working on real research problems, and a record of the judgment they exercise before any result is certain. Our long-term goal is a world where discovery is no longer bottlenecked by the number of human scientists alive to pursue it.

our work

Aristotle

We built and operate Aristotle, the world's first AI research partner. Aristotle takes on the reasoning legwork so scientists can stay on the science: forming hypotheses, weighing what the evidence supports, and deciding what to pursue next. Aristotle is trusted by thousands of researchers at some of the most demanding institutions in science, including GSK, Pfizer, AstraZeneca, Merck, Stanford, MIT, Harvard, and the FDA. You can try it here.

Columbus-1

Columbus-1 is our autonomous research system for long-horizon scientific problems. It runs on Metacognitive Steering (MCS), a sparse controller that switches a frozen 1T-parameter model between exploration, execution, and self-challenge without fine-tuning a single weight. Columbus-1 has already found 8 zero-day Bluetooth vulnerabilities and designed a rocket that lands itself. You can read the announcement.

our humans