Cubby
Cubby isn't a single model. It's a neuro-symbolic agent with three faculties: language, reasoning, and a brain that gives it character.
Three parts, each doing one job
Most AI systems are one enormous model asked to do everything at once. We built Cubby the other way, as separate faculties that each handle what they're actually suited for — the way a mind is more than any single region of it.
Cubby-LM — the language cortex
This is the part that handles language. It reads a request, writes a response, and, most importantly, emits programs for the reasoning machine to run. We built it wide rather than deep on purpose: the heavy reasoning is meant to happen off the token axis, so the language cortex never has to carry it. It's a neutral, bilingual substrate. It speaks; it doesn't decide.
The reasoning machine — decisions you can read
Cubby's decisions don't live inside opaque activations. The language cortex emits a program, a symbolic machine runs it over grounded world-models, and every step is auditable. The reasoning happens somewhere you can look — which, again, is a big part of why the system is cheap — and far greener — to run.
The brain layer — character that isn't baked into the weights
Affect and personality come from a separate brain layer: a spiking neural system with a working neurochemistry that shapes behaviour while the model runs. Character lives here, not fossilised in the language cortex's weights. That keeps the substrate underneath neutral, and lets the personality be its own, evolving thing.
Senses, and a world it learns for itself
A mind has to perceive before it can understand. Cubby's brain takes the world in through a visual cortex: a live image stream feeds the spiking brain, and what it sees drives real learning — the network rewires itself frame by frame and grows new neurons when it meets something it hasn't seen before. This part runs today, not on a slide.
From what it perceives, the agent builds its own model of the world and keeps it current as it goes. It learns which moves matter, and what they're worth, by living in an environment rather than being trained on a fixed dataset and then frozen — no policy network deciding for it, no offline retrain, just a world model it updates for itself while it runs. See it do exactly that →
An auditory cortex — letting language and sound arrive as something heard — is the natural next sense. The groundwork already runs: the brain binds what it sees, the words attached to it, and how it feels into a single grounded memory, so hearing just adds one more stream into machinery that already works. It's on the roadmap, not in the demo yet — evolution, not invention.