Agreed to some extent, but I don't think you'll have too long to wait before that comes to pass. It can't possibly be that hard to build a model that will disassemble a given song into its original tracks, like a Fourier transform that yields drums, strings, keyboards, and vocals rather than sines and cosines. Equally unlikely that such a model will be too large to run locally.

We will get some very cool tools -- and some very cool remixes - when that happens.

Of course there are music models that let you edit and remix music, just add a bit of Audacity:

https://github.com/adefossez/demucs

Yes, why are deep learning models so effective relatively at writing code? It's because programmers have been making their work copyleft for decades, and continue to do so.