So it's basically a midi predicting LLM? I don't understand how it solves the issue of representing music, i.e. actual sound waves and not simply the notes that make it up?
It's the best note mixing LLM I've seen. Someone just needs to train a sound diffusion model on the midi notes to orchestral mp3s. Give it a few weeks. "Unified" seems to mean all the notes in one place which you can edit/build on/mix so that you can put another layer on top, human or ai performs it with words & flare.
Uh, yeah. I've been looking at programmatically writing guide tone lines and motifs in jazz for my personal piano practice, and this... looks very sophisticated.