I don’t know if gwern realized how powerful his model was. His examples are underwhelming, because you have to prompt it in a certain way to get it to generate chords. He was showing me samples and they were neat, but boring.

One day he posted something that sounded pretty amazing, and I was blown away. “More like that, please.” It had chords in it.

He didn’t pursue it past that. I did. So it’s possible that no one is aware of how crucial the input format actually is to the success of the music that I was able to produce.

(And “produce” is a fair description here — choosing the instruments was really important, and the model didn’t do it. It wasn’t as easy as press a button. It felt like I was suddenly a 15x music producer, since I made all those tracks in one night. Such is the power of ML.)