[bryceneal]

One thing that occurs to me about music which makes it different from software or film is that it's a very impulsive art form. A human with semi-modern equipment can write, produce, mix, and master a song in less than a day. They can do this in a way that gives them full creative control over the thousands of tiny conscious and unconscious decisions that ultimately feed into what it is you like (but are probably unable to accurately articulate with words) about a song.

In this case, it's difficult for me to see what exactly is gained by offloading all of those decisions to a mean regression algorithm. I agree it's a fun toy, but I can't imagine actually listening to or deeply enjoying any song made with these music models.

I could imagine it being used successfully in some targeted part of the process. For example, manipulating drums or changing the timbre of some previously recorded instrument. In that case it's not much different than traditional music creation tools like sampling.