Not a language model. Sparse predictive coding in pure numpy — no backpropagation, no GPU, under a megabyte of core. It forms its own concepts, and a teacher only names what is already there.
Every drawing below is pulled from a set the system has never been trained on. It looks at the raw pixels and answers with a word it was taught by coupling, not by gradient descent. Nothing is cached — press the button and watch it think.
The same brain, the same read-only path — nothing you draw changes it, so try as often as you like. One catch worth knowing up front: it has been taught ten words and can only answer with those. Draw a bicycle and it will pick the closest thing it knows.
Draw with the mouse, or on a touchscreen with your finger. Nothing leaves your browser but 784 numbers.
"75% accuracy" means nothing on its own. Each figure below carries what it was measured against — chance level, or the strongest simple baseline we could put in its way.
On 4,000 drawings held out of training. Chance is 10%.
33.8% correct across 100 classes where chance is 1% — and every one of the hundred gets named.
Against 19.3% for gradient descent on the same schedule, which forgets catastrophically.
The 20-watt principle: never everything at once. Accuracy holds at 2% activity.
One file. Copy it, and the thing that learned comes with it.
Sender and listener agree on a signal system nobody designed, and it holds on meanings never intended.
Each layer guesses what the layer below is about to show it. Only the mismatch travels upward.
Roughly two units in a hundred fire at any moment. That constraint is the design, not an optimisation.
Patterns that recur get their own high-dimensional signature — without a single label.
A teacher couples a name to a concept that already exists. Naming is the last step, not the mechanism.
A capability page without this section would be advertising. Everything demonstrated above runs at toy scale — hand-drawn 28×28 images and grid worlds.
above biology in energy per step. Only 0.9% of that is actual arithmetic — the rest is interpreter and memory traffic.
parameters. The human cortex has 1.1·10¹⁴ synapses — a factor of 2.7 billion between us and the thing we are pointing at.
shape information surviving the deep hierarchy. It destroys what it is supposed to abstract. The most important open problem, and it sits in the core.
the consciousness metric runs along, but is coupled to no measured capability. Until it is, it measures nothing.