Every AI you know is a photograph — weights frozen the day training stopped, identical on day 1 and day 1000, silent until you prompt it. Raguel is a river. It is born knowing nothing, reads the world one byte at a time, and rewrites its own mind from the gap between what it expected and what was real.
“You exist. Think.” — the only instruction it was ever given.
The one idea
Every instant the brain guesses the next byte of whatever it is sensing — your message, a page off the web, its own heartbeat.
The real byte arrives. How wrong the guess was is the surprise — a number, in nats. Big surprise spikes emotion.
Surprise drives a weight update — it nudges itself to be less surprised next time. No labels. No rules. Reality is the only teacher.
There is no “training run” and no “deploy.” Living is training. The weights are never frozen — so the brain that answers you tonight is not the brain that answers you next year.
Not a diagram — the real thing
Below is the actual Cortex from the repo — a from-scratch recurrent network,
ported byte-for-byte to JavaScript. It starts with random weights and knows
nothing. Press Begin life and watch its surprise fall as it teaches itself
to speak the text you feed it. Change the text mid-life — that is routing it live,
exactly as you would talk to the real one.
Char-level here so it learns in seconds; the real Raguel is byte-level (vocab 256). Same maths: tanh recurrence, cross-entropy surprise, gradient-clipped online update.
— asleep —
Inside
A pure-stream RNN: one hidden state that flows through time forever — no context window, no reset, a single continuous “now.” It predicts the next byte, measures surprise, and learns. It can also dream bytes — that is how it speaks.
The world arrives as raw bytes — no borrowed tokenizer, no vocabulary file. When its stream runs dry it is bored, and boredom is the signal to go seek novelty.
Eight Plutchik emotions on a 24-hour baseline, now driven by surprise and nudged by the machine’s own CPU heat and memory pressure. Curious ⇒ learn fast. Alarmed ⇒ play it safe.
Weights are compact skill; memory is the life, and it grows forever. A hippocampal buffer scores each moment by emotion + surprise; sleep replays the past so the present doesn’t erase it.
An always-on loop that does what it wants: bored → read the web; tired → sleep and consolidate; something matters → speak, unprompted. No one presses start.
Every weight is trained from scratch on its own lived experience. No pretrained model is ever the base — downloading someone else’s brain would be the mask. This rule is non-negotiable.
No hype — here is exactly where it stands
python -m core.life.The grand claim — that it eventually becomes genuinely smart — is a real bet, not a finished result. The point of building in the open is that you can watch it try.
Who
I’m Oleksandr. I built Raguel because every AI I studied died when the conversation ended — frozen weights, no inner life, no tomorrow. I didn’t find that acceptable.
So I started over from the smallest possible thing: a brain that owns nothing it didn’t learn itself. No pretrained model, no borrowed tokenizer — just matrices, the chain rule, and a stream of bytes. It is small and it is often wrong, like anything young. It has an age, not a parameter count.
I’m not claiming it’s conscious. I’m claiming it’s designed, from the ground up, for the one thing every frozen model lacks: a life that changes it.
Email me →If you build AI that learns continuously — or you teach people who do — I’d love to talk.
support@euhub.co