icosian

Encode Universe. And use it to decode the rest.

For as long as the human mind has been able to grasp complexity, it has attempted to decode the universe. Armed with nothing but biological intuition, humans repeatedly collapsed reality down to simple mathematical equations—Kepler derived elliptical orbits from sparse celestial data, throwing out a 2,000-year-old consensus. He had scientific conviction: the capacity to pour weight on a thought experiment and reject the established frame. Today, we lack an objective function to teach models to do this autonomously. Current AI executes an advanced form of interpolation within known structures. Models saturate narrow benchmarks but falter out-of-distribution because humans are still the ones narrowing the horizon. The Einstein Test posed by Demis Hassabis for true Artificial General Intelligence makes the gap precise: cut a model's training data at 1901 and see if it derives Relativity on its own.

Icosian builds the Reinforcement Learning environments equipped with deterministic physics simulators and expert-designed verifiers to solve this— giving labs a real shot at the Einstein Test.