Quite interesting.
It seems to be producing some suggestions of note, including some really off colour ones, which isn’t a bad thing. This may be a nice way to break up a creativity block - at least if it does something I DON’T want out of a song, that seems to assist in clarifying what I do want, and I can go off on my own.
Can I suggest being able to resize the suggestions in the sidebar - at the moment I can’t work out a way to do so. The previews are so small so as not to really be indicative of the suggestions. This might mean more user clarity, less user clicking through them to determine which one is desirable, therefore less unnecessary resource usage.
Is Aria to use the same dataset that the original Anticipatory Music Transformer at Stanford used? If no (or only partially), can you give specifics on the training data?
Can I confirm there are no plans to include the private projects of Hookpad/Hooktheory users into training data?
Innovations in generative music are something I’m interested in personally, but data governance certainly is too. 
Thanks guys.