MODEL OBSERVATORY / EXPERIMENT 05
What does a
model sound like?
Turn measured expert routing and recurrent head activity into a small, strange piece of music. Same capture. Same mapping. Same notes.
This is a musical mapping of real numbers, not a recording of a model's voice or thoughts.
Recorded prompt
Explain why the sky is blue in two sentences.
A COMPOSITION WITH A RECEIPT
The numbers choose the notes.
These are the same real captures available in Model Observatory. No new inference runs here. Switching blocks changes which recorded measurements you hear.
Inspect the original topology ↗Flint: six routed experts
Each selected expert becomes a sine tone. Its expert ID chooses a note in a repeating, three-octave C pentatonic scale. Its routing weight sets amplitude using 0.11 × √weight. The shared expert and dense blocks are excluded.
Qwen Open: recurrent head output
The six recurrent value heads with the largest output RMS at each token become triangle tones. Head ID chooses a note on the same scale, one octave higher. Amplitude is 0.08 × RMS / the largest head RMS across the chosen block's capture. All 48 heads remain visible.
How to read the comparison
Both use a fixed token clock selected above, not the original inference timing. Pitch repeats every 15 IDs, so different IDs can share a note. Colors show routing weight for Flint and normalized head RMS for Qwen. The two mappings have different units and are not a model-quality comparison.
Mono, 22,050 Hz. Identical settings produce identical samples. A short attack and release soften note boundaries. WAV export includes the selected volume and stays on your device.