Architecture
Cubby is a mind assembled from parts. Here is how those parts fit — and why the arrangement is the whole point.
How it all connects
Start with the whole loop, then we'll take it apart. A prompt comes in and the language cortex comprehends it. The request is routed to the world model that best fits it — Cubby holds a mixture of them, not one monolith. That model compiles the reasoning into bytecode, checked against a values gate before anything runs, and the symbolic VM executes it — deterministically, auditably. The result comes back to the language cortex to be voiced as an answer. Through all of it, the live brain perceives and modulates — the loop has a mood.
That's the shape: the language cortex speaks, the world models and the VM do the reasoning, the brain gives it character. Each of the pieces below is one part of this loop.
Three faculties
The whole agent is Cubby. Its language cortex — Cubby-LM — reads and writes and emits programs. A reasoning machine runs those programs. A brain layer gives it character. No single part is asked to do everything; each does the job it's suited for.
Where the reasoning happens
This is the thesis in one picture. In a conventional model, the thinking is trapped inside opaque activations — when it's wrong, you can't say where. Cubby's language cortex instead emits a program that a symbolic machine runs over grounded world-models. The cyan path is legible at every step: you can read the decision, not just its output. Putting the reasoning there is also a big part of why the system is cheap to run.
The live brain
Character isn't fossilised in the cortex's weights, and it isn't frozen at training time. A visual cortex feeds a spiking central nervous system with a working neurochemistry and a hippocampal memory. It runs a wake/sleep cycle that consolidates what matters. And it is always on, always learning: the network adjusts itself frame by frame through spike-timing plasticity, and it grows new neurons as it meets things it hasn't seen — real neurogenesis, live, with no offline retrain.
It genuinely learns from what it watches. Point it at narrated video and it grounds meaning the way an infant does — not by hearing words, but by binding them to the moment: the transcript, aligned to the timestamp, is fused with what the brain sees on screen and the chemistry it feels right then. Watch the same thing again and it recognises what it knows, with a flicker of familiarity. Its taste develops with experience, an internal chemistry of dopamine, serotonin, cortisol and oxytocin shifting with what it takes in — and the changes persist from one session to the next. A brain that watches the world and is changed by it — which is what "intelligence that isn't artificial" means when the neural mechanisms are real: perception, plasticity, memory, neurochemistry, growth, running in real time.
The next cortex: hearing
The brain already grounds meaning across modalities — it binds what it sees to language and to how it feels, all at once. An auditory cortex is the natural next step, because the hard part is already proven: the machinery that fuses a second stream into grounded memory exists and works. Today language reaches the brain as timestamped transcript; a hearing cortex simply lets it arrive as sound — the same grounding, now heard rather than read.
This is on the roadmap, not in the demo yet — but it is evolution, not invention. The concept is proven; the next cortex extends it.
Seen in the wild: an agent that reasons in the open
The claim that reasoning can live in an inspectable machine isn't only a diagram. We built a small world — a 3D maze game — where the agent plays without a trained policy and without a neural network doing the deciding. It learns each move as a composable vector operation, and when a hazard walls off its goal, it discovers a new ability by composing the moves it already has — binding a step with itself to leap over the obstacle. It plans over those operations, chooses targets with a symbolic value memory, and runs a small affective brain that learns to fear the things that catch it.
Every part of that is legible: you can watch it get stuck, invent a jump, and change its mind. It is the whole thesis in miniature — reasoning, world-model, and affect, none of it hidden in a black box.
And here is where the architecture meets the economics. There is no policy network, no reinforcement-learning run, no large model burning compute on every move. The agent decides with cheap vector arithmetic in a symbolic machine — so it runs in real time on modest hardware, light enough to host on a small box or even entirely in your browser. An intelligent agent that discovers abilities, plans, and feels — running at a compute cost that is a rounding error next to a language model asked to do the same thing. That is second-generation AI in one word: cheap to run — shown, not asserted.
It is also more than a demo. The maze is a world — a place to test an agent that perceives, plans, avoids danger, and manages its own energy and time. It is where we validate reasoning-as-action before the agent has a body. The same stack is designed to run on modest physical hardware, down to a robot's microcontroller; cubby-man is the pre-robotic world we prove it in first.