The world is not only software.

Most of what happens on Earth is never written down. It exists as signals passing through instruments, and almost none of it has ever been part of how models learn. Models have read nearly everything humans have typed and almost nothing the world has said.

We are a small lab. We are trying to teach models to decode the physical world and to operate the instruments it runs on.

We work this way because reality is the one grader that cannot be fooled. A benchmark can leak. A judge can be gamed. A measurement is simply true, and a model rewarded by measurement is learning something no text corpus can give it.

We are backed by Y Combinator and are currently in the F26 batch.

If this is the future you also see, write to us directly.