Notes You Can’t Delete
depth
Concept surgery on small music transformersa living study

Notes you can’t delete

Take a musical idea out of everything a small model is ever trained on, and it rebuilds the idea anyway. You cannot un-teach it. This is the music, the numbers, and the whole build — from a single note up to an entire instrument.

~40models, from scratch
note → bandscales of the same test
<$30total GPU spend
26Mparams, the workhorse
The one result

Delete the concept. It comes back.

The same experiment, run at every scale we could reach. Edit an idea out of the training data, train a model that never sees it once, and measure how strongly it still expects it. The bar is how much came back.

The idea doesn’t live in its instances. It lives in the grammar around them — the key, the root, the riff, the groove — which the edit never touches. So the model infers it back. The only place deletion finally bites is a trait so novel that nothing else in the data implies it.

Hear it

A model wrote these drums for this riff

Strip the drums off a real metal riff and let a model — one that had the drums torn out of every training example — fill them back in, beat by beat. Flip between its drums and the real ones while it plays.

0:00 / 0:00

The model's kicks land on a guitar onset 100% of the time — exactly like the real drums. It learned that the kick locks to the palm-mute chug, and used it.

Rendered through a GM soundfont from symbolic MIDI. Full experiment, the double dissociation, and the honest caveats: Rebuilding the bass and drums.