To test the limits of generative AI
I build with generative AI to find out how far a machine can follow a feeling. It begins somewhere unexpected: with noise.
It does not store the recordings. What it learns is the way back from noise.
The grains become 30 seconds of a song, drawn as sound: time runs left to right, pitch rises from bottom to top. A real model does this to millions of numbers at once, and for music it works on a compressed picture of the sound much like this one.
Take a real recording and add random noise, a little at a time, until only static is left.
A neural network watches that happen to millions of recordings, and each time it is asked one thing: which part of this is noise?
Then hand it fresh static and a few words. It removes what it believes is noise, step by step, and the words steer every step.








