I just tried it out because of the discussions on this thread, and I got to say I land squarely on the side of this is neat, but it is not artistry. Every little thing I generated sounded like things I've heard before. I was trying hard to get it to create something unique, using obscure language or ideas. It didn't even get close to something interesting in my opinion, every single output was like if you combined every top 40 song ever made and then only distilled out the parts relevant to certain keywords in a prompt.

These tools will probably be great for making music for commercials. But if you want to make something interesting, unique, or experimental, I don't think these are quite suited for it.

It seems to be a very similar limitation to text-based llms. They are great at synthesizing the most likely response to your input. But never very good at coming up with something unique or unlikely.