\o/ SILO v1
dual-transformer silo · v1

TRANSFORMER SILO

Two floors. Downstairs an intent engine spins the preloaded user context in a centrifuge until like settles with like; upstairs a normal transformer runs over what comes out. ()\< loads the context in, <\o/> spins it, and the intents get spat >>>> upstairs. It really converges, sorts like-to-like, compresses, and transforms — a runnable, honest sketch, not a trained model.

Run the silo

Load context · spin · transform

Type a "user context" (a bag of words), pick how many intents K to spin it down to, and run. Floor 1 clusters like-to-like and compresses; floor 2 runs a real toy transformer over the intents. Everything is computed live and deterministically in your browser.

◎ run_silo — the centrifuge + a toy transformer, computed client-side (no network)
K intents
3
Floor 1 · the intent enginethe centrifuge — spin until like settles with like
()\< intake  ·  <\o/> spinning  ·  convergent similarity clustering (Lloyd's)
\o/
>>>>  the K intents rise  >>>>
Floor 2 · a normal transformera real toy forward pass over the intents
Floor 1 is Lloyd's algorithm (assign → recenter → repeat); the intra-cluster energy is monotonically non-increasing and it settles — that is the "centrifugal, like-to-like." Floor 2 is the the-forward-pass engine (RMSNorm → attention → MLP) over the K intents, mean-pooled to a prediction. Disc positions are a 2-D projection of the 4-D vectors; colour = cluster. A runnable TOY of the architecture — not a trained model, and no claim it beats a plain transformer.
The colophon

What it accomplishes — and what it doesn't

What's real

It genuinely converges (energy falls each spin, then settles), sorts like-to-like (planted groups separate cleanly), compresses N context vectors to K intents, and transforms — a real toy forward pass over the intents. python selftest.py proves all of it with no network (12 checks).

What it isn't

It is not a trained model and makes no claim to beat a standard transformer. Whether a clustering pre-pass actually helps on real data is an empirical question this doesn't answer — it only shows the mechanism, runnably and deterministically. v1: an honest sketch of the silo David described.

SOURCE · github.com/DavidWise01/transformer-silo · silo.py (centrifuge + toy transformer + run_silo) · selftest.py (converges · deterministic · like-to-like · N→K · runs). Floor 2 reuses the the-forward-pass engine.
⧉ the code (GitHub) \o/ the forward pass ↗ v2 · the honest head-to-head ⊞ the membrane map ⌂ ud0 · the atlas