This is a small program with its own eyes, ears, hand, and
voice. Watch it teach itself the alphabet, then show it a letter and it will say the sound and draw the
shape. You can teach it your own voice, hide letters in a word and watch it work them out, and even switch
to a bigger set of "receptors" to see it draw crisper. Everything happens in your browser.
Careful with the words. This is a tiny simulation. What it shares with a child
is the mathematics of learning — noticing which shapes and sounds keep showing up together, and
updating its predictions accordingly. That kind of process is called active
inference and has been studied for over 150 years. What it does not have is a body, a life, a story of
its own, or a first-person point of view — so it does not mean any of this the way a child does.
The sounds are a little body ringing like a bell, not a voice. Nothing about you is uploaded or saved.
What to do:① wait for it to teach itself the alphabet (a few seconds) ·
② press a letter and it will say and draw it ·
③ try teaching it your voice or a bigger receptor grid
· the little green badge tells you the browser and the reference program agree, receptor-for-receptor.
idle — press Teach
① Teach the being
watch it teach itself — the little dials spin as it works, and it stops when there's nothing new to learn
② Show it a letter → it says & draws
press a letter — it will say the sound with its own little voice and draw the shape with its own little hand
How sharp are its eyes? Give it a bigger eye and teach it again:
A bigger eye means more little dots (receptors) to sample the letter with — and a matching hand that can draw with more little dots. Its "IT MAKES" picture gets less blocky and its stored prototype of each letter gets finer. Switching restarts its learning (voice recordings and stored word counts reset). The little green badge always refers to the sealed 8×8 being — that's the one we can check against the reference program. The default view starts at 16×16 so Devanagari and Latin letters are sharper; click 8×8 above to see the parity-verified reference model.
shown→
·
read-out→it draws
③ Play it a sound → it draws the picture
press a sound — the model matches it to its closest stored sound prototype, then inks the letter bound to that prototype
🔊
played→it makes
④ Teach it your voice → watch the counts update live non-citable demo
hold a letter button and say it out loud; the recording becomes a stored sound prototype, counted under that letter — no waiting, no training run, just counting
Non-citable demo — read this first. Until you record, the model has only been given the built-in "teacher" tones — no human speech. Its ear is 64 tiny "strings" tuned across the range your voice lives in — so any sound you make lands on those strings and becomes a signal it can store and count. The shortcut, stated plainly: the letter button you hold tells the model which letter your recording belongs to — a supervised label write, not something it works out for itself (the lessons above bind by co-occurrence counting instead). The yes/no tally below is your own call, not a measurement. Microphone audio stays in this tab — nothing is uploaded or saved, and the microphone is switched off each time you let go. The teacher tones are just bells ringing (not real speech), which is why they sound rough.
press & hold a letter and say it out loud (a few times each):
press & hold each letter a few times, then hold & match
was the match right? (your call)
Each recording is assimilated live (running-mean prototype + counts) and written into the count table under the cause of the letter whose button you held — the audio dictionary grows as you add takes. An online count update, no re-training. Keep your takes of a letter consistent; wildly different ones just spawn extra prototypes (still written under the same letter).
⑤ Fill in a hidden letter from word context
hide a letter in a word — the model infers the word from the visible letters, then fills in the missing one from the word posterior (letter cloze)
word · hide position
The model is given only the other letters, infers the word, and
imagines the hidden one from the word posterior — then we occlude that letter 85% and show that
context recovers it where the eye alone fails. "Cloze", never "comprehension".
⑥ Forgiveness — letting a stuck idea change
क्षमा — when it clings to the wrong answer, "forgiving" it doesn't erase what it learned; it just softens its grip so new evidence can move it. The table shows how many correct examples are needed with vs. without forgiveness.
Overturning a belief built from R wrong observations needs ~R correct
ones; forgive(f) cuts that to ~f·R — measured, monotone. The owner's
subtle-system spec, as tested count math (never therapy language).
An honest negative
recorded as prominently as the passes
UNI·Sensorium v1 · in-browser port ≡ Python reference (parity above) ·
no LLM, no backprop, no clock · real audio/graphics are non-citable demos.