Reasoning belongs in a machine you can read
Most AI reasons inside a black box of numbers. We put the reasoning somewhere you can actually look.
The problem with reasoning in activations
In a conventional model, the thinking happens inside opaque token activations: billions of numbers with nothing you can grab onto. When the model is right, you can't say why. When it's wrong, you can't say where. For plenty of uses that's fine. For anything you need to trust, it isn't.
Reasoning in a symbolic machine
Cubby works differently. Its language model writes a small program, and a symbolic machine runs that program over grounded world-models. Every step is there to read — you can watch the decision being made, not just the answer that falls out. Moving the reasoning off the token axis and into that machine is also a big part of why the system is cheap to run.
Method, not moat
This is something we're happy to show, because it's a method rather than a secret. Our maze-solving agent does exactly this in the open: it reasons in a symbolic machine you can inspect, with no neural network doing the deciding. See it reason in the open →