UNI · a developmental active-inference simulation
The Sensorium & the first alphabet
One hierarchical, no-backprop agent whose senses — reading, hearing, seeing — all land as one clean signal. Its priors are stored only over signals; it is taught the alphabet in Latin, English & Hindi at once, and it has motors that draw a picture and speak a sound. Shown a picture it says the sound; given a sound it makes the picture.
v0 · curated 5-letter lesson
29/29 tests green
picture→say read-out 5/5
sound→draw 5/5
15 prototypes → 5 causes
The single loop
read / see / hear → one signal → bind → draw / speak
Sense · readglyph → signalLatin & Devanagari letter shapes
Sense · see / hearimage · sound → signalfixed transducers (the sense organs)
→
One clean signalℝ⁶⁴ · ≥0 · Σ=1the ONLY thing priors are stored over — no typed words or symbols
→
Agent · bindcount co-occurrencerunning-mean prototypes + Dirichlet counts
→
Motor · actdraw · speakinverse plants over the same signal
One signal, three senses
the letter “A”, taught trilingually in a single “now”
Why it matters. The three inputs look nothing alike — a Latin shape, a
Devanagari shape, a sound — yet each becomes the same type of 64-channel signal (the heatmaps).
The agent binds them into one latent cause not because they resemble each other, but because they
co-occur. That is cross-modal binding by counting — not understanding.
The lesson
Latin + English + Hindi, taught at the same time · press ▶ to hear what the agent emits
| letter | Latin (seen) | Hindi (seen) | sound → signal (heard) | agent MADE | says |
|---|
Read this honestly. “MADE” is the agent’s own picture, drawn from
hearing the sound alone — a blocky 8×8 read-out of the prototype it bound, not a copy of the input.
Fidelity is bounded by the 8×8 signal, on purpose.
Two capabilities, like a body
a sensor to take the world in, a motor to put it back out
Show a picture → it says the sound
Give a sound → it makes the picture
🔊 /a/→![]()
audio signal → nearest prototype → latent cause → read out the bound vision prototype → draw motor.
Nested time, in the yuga ratio
“time” is not one clock — it is four loops at 4 : 3 : 2 : 1
L0 · Kali
×1
signal — plastic, forgets fastL1 · Dvapara
×2
cross-modal bindingL2 · Treta
×3
sequence / order (next rung)L3 · Satya
×4
slow self / genre prior — stableShallow loops update every tick and forget quickly; deep priors update rarely and
forget slowly. Class-C engineering choice inspired by the yuga proportions — not a cosmological claim.
Evidence
every claim carries a class · re-derive with pytest -q
| invariant | class | result |
|---|---|---|
| Signal uniformity — every sense → Signal(64), ≥0, Σ=1 | A | 15/15 valid |
| No backprop — AST scan, positive control bites | A | 0 autodiff imports |
| Transducers deterministic & distinctness-preserving | A | min cross-letter 0.20 |
| Nested rhythm — timescales L3..L0 = 4:3:2:1 | A | exact |
| Cross-modal binding recovers the taught letter | A/C | 5/5 (bar 0.8) |
| picture→say read-out exact (lossy speak plant, still identifiable) | A/C | 5/5 |
| sound→draw round-trip (clean 8×8 draw inverse, <1e-6) | A/C | 5/5 |
| Repetition-surprisal falls over the lesson | C | 0.197 → 0.00 |
What this is — and is not
Allowed, and shown
- ✓ a developmental active-inference simulation
- ✓ the vision / audio signal channels
- ✓ next-signal prediction & surprisal
- ✓ cross-modal binding by counting
- ✓ picture→sound & sound→picture generation
Forbidden — never claimed
- it sees / hears / understands / reads
- it knows the letter
- binding = comprehension
- conscious · human-level · AGI
- real audio/fonts are non-citable demos