A Brain That Is Never Switched Off: Binary Neurons, One Local Learning Rule, Innate Wiring and Speech Only Through Muscles

We describe concept-brain, a research prototype of an artificial brain that lives one continuous life: it is never reset, never frozen and never trained offline. It consists only of binary neurons wired at birth by one generator (three-dimensional, local, random), a single local Hebbian learning rule with concave growth and one global weight decay, and output only through simulated muscles. Sensory areas send one-to-one "copies" of their feature neurons to a multi-layer prefrontal cortex (PFC) whose last layer predicts the next input; a thought is a first-layer copy lit from inside; speaking is that prediction returned to the muscles. Reward is only a tool that strengthens or weakens the links that produced an action in the last one to two seconds. We give (i) the architecture as a short rule set; (ii) a theory showing that the Hebbian readout of the last layer is the readout of an autoregressive language model under stated conditions, with signal-to-noise and timing analyses; (iii) measurements of the brain built from the rule set; and (iv) a test on an invented thirteen-word logic language, chosen because it is small, exact and reproducible. There the brain learns, by only listening to others' dialogues and with no reward, to expect the right answer word after "T ?" within 20 dialogues and after "NOT T ?" within one or two hearings, once sentences are spoken at the pace of its working-memory window; with hand-wired word-muscle links and an innate brake it answers in turn without talking over others. We list what does not work yet, which is most of language. Files: concept-brain.pdf (English, 13 pages) and concept-brain-zh.pdf (Chinese version with the same content, 11 pages; 中文版). Code, experiments and the full architecture: github.com/leeR1ven/concept-brain (software DOI 10.5281/zenodo.24084045).