An experiment with Google’s Magenta RealTime 2

Piano
to flute.

The same piano input.
Flute from the edited model.

0.032% of the model’s weights changed.
A permanent edit, calculated without backpropagation.

Inspired by Sparks of Weight-Space Programming.

Main edit by GPT-6 Astra, using Python and piano/flute examples.

Listening room Recorded audio
Text prompt12 seconds · Stereo

Prompt sent to the model

A solo honky-tonk piano plays a fast boogie-woogie, only piano.

Original and edited models get this exact prompt. Only the saved weights change.

Original model output0:12
0:00 / 0:12

Heard in this outputPiano

Generation details

These are saved model outputs. Playback restarts when you switch. Audio levels are matched for comparison; each waveform is scaled to its clip’s peak.

Inside the
edited matrix.

This adapter maps text or audio style into the generator. Compare its original weights, edited weights, and the change.

768 × 1,024
NegativeWeight change · contrast enhanced Positive

Hover to inspect. Tap or use arrow keys.

Selected block

Rows 0–15 · Columns 0–15
Before—
After—
Change—

Select a block to inspect its values.

  1. 1

    Encode the examples

    The frozen encoder turns piano, flute and control examples into style vectors.

  2. 2

    Calculate the update

    A linear solve sets where to apply the piano-to-flute shift, using the controls as constraints.

  3. 3

    Save the matrix

    The edited checkpoint loads in the original runtime. No extra prompt or activation hook.

Main patch, actual checkpoint data. Each square shows the mean of 256 weights.Inspect the data

How the weights were chosen

Try the
3.15 MB patch.

Replace the style adapter in the pinned Magenta checkpoint, then generate locally.

These are selected examples. Audio references guide the style; the model generates new music.