About
A lab built around one property: forward only.
We study how a machine can learn the causal structure of its world efficiently, represent it as programs it can check, and keep every gain it verifies.
The thesis
Three ideas, usually pursued apart.
Ratchet Labs works on the intersection of three ideas that are usually pursued apart. The first is sample-efficient causal learning: a system should be able to work out how its environment behaves from a few dozen interactions, the way a person forms a working theory of a new game after a handful of tries, rather than by consuming datasets the size of the internet. We take efficiency, not scale, as the measure of understanding.
The second is world models as causal simulators. We want internal models you can do more than sample from — models you can interrogate. Set a cause and read the effect; change one rule and watch the consequences propagate. A model that supports intervention, not just prediction, is a model that has captured something about how the world actually works.
The third is program synthesis. We represent each learned law as a discrete, inspectable program instead of a diffuse pattern spread across millions of weights. A program can be read by a person, tested against new observations, and either confirmed or rejected. Verification stops being a hope about a black box and becomes a concrete check you can run.
Bringing the three together yields the property we are named for. When a law is written as a program, it can be verified; when it is verified, it can be locked in. That is the ratchet: a toothed wheel and a pawl that lets motion pass in one direction and jams it in the other. Turn the wheel with every interaction, and let the pawl fall into place each time a law survives its test. Knowledge accumulates — new rules can always be added, and no proven rule can be silently lost. It is continual learning without catastrophic forgetting, by construction rather than by luck.