Pre-ai:

You compose a song. It exists in your mind, or on paper, as a series of notes and chords, sometimes with words. This is not yet "a song", it's sheet music. To get to a finished product, you have to do a demo, and that involves a lot of choices, but those choices may still be only a temporary bridge to get to a later re-recording, with the demo acting only as a proof of concept that was necessary to get sign-off or funding from other members of your band, a record label, a producer, etc.

Back to the demo. You have to decide how to ARRANGE that music you've written. That means you have to decide which kinds of instruments/synthesizers play what parts. Maybe the bassline is played on a bass with a fuzz filter. Maybe it's clean. Maybe it's a slap bass. Got a guitar part? Is it acoustic or electric? Get it? Lots of choices to make.

If you have vocals, you have to decide who is going to be singing the vocals and in what style.

Now, with this new model: you can either hand-write the "sheet music" (in its proprietary format, but it's still a form of musical notation) or have an LLM do that. Either way, given the sheet music, this model can go and reliably produce new demos for you, so that you can try out different arrangements of the same exact musical structure (including lyrics). It just automates all the arrangement and renders it in one go, instead of you recording each part yourself.