We are building superintelligence for discovery.

A meta-solver for the formal sciences.

See our research

No cyclic (465, 145, 45) difference set exists, so none exists in any abelian group of order 465. Combinatorial designs

The last invention people will need to make.

People have been inventing for 3.3 million years, and nearly all of it happened at the very end. Each line below magnifies the orange tip of the line above it.

  1. 3.3 million yearsStone tools3.3 million years agoFire1 million years ago
  2. The last 12,000 yearsFarming11,500 years agoWriting5,300 years ago
  3. The last 600 yearsPrinting press1450Telescope1608Steam engine1712Dynamo1831Powered flight1903
  4. The last 80 yearsTransistor1947Moon landing1969The web1989Deep learning2012AlphaFold2020

The last invention

Four timelines, each to scale. The first covers 3.3 million years, from the first stone tools to today, with fire one million years ago. The second magnifies its last 12,000 years: farming, then writing. The third magnifies the last 600 years: the printing press, the telescope, the steam engine, the dynamo and powered flight. The fourth magnifies the last 80 years: the transistor, the Moon landing, the web, deep learning and AlphaFold. All four end at today, where the last invention is marked.

In 1965 I. J. Good called the first ultraintelligent machine “the last invention that man need ever make”. Once a machine can discover, every discovery after it can come from the machine. That is what we are building.

Dates are approximate. Every line is drawn to scale.

Every discovery should make the next one cheaper.

Machine discovery today is bought with scale. OpenAI’s Navier–Stokes run took about ten thousand agents and 130 billion output tokens.

Most of that cost is search where no answer can be. We find a problem’s structure first and build a solver for exactly that: up to 47,000 times faster than published run times.

A result counts only when multiple separately written verifiers re-derive it. Where only one engine stands behind a claim, we say so.

Every verified result leaves its tools behind. A kind of problem the system had never seen got a verified solver in five to six minutes.

The first stepping stones.

Open questions settled in six fields so far: combinatorial designs, integer sequences, coding theory, quantum error correction, SAT and proof, physics and materials.

None of them is the goal. We began where a result is simply right or wrong, and each field left tools that made the next cheaper to enter.

A century of science, every year.