One is when it attempts to generate vocals without a lyrics prompt. It's gibberish but just on the edge of comprehensibility. Sometimes it'll be entirely spoken word with no accompaniment. Very uncanny.

Another is transitioning between vocals and instrument in the same melody line. Like a humanesque voice holding a steady note at the end of a verse which seamlessly transitions into a saxophone sound and proceeds into a solo. Or vice versa, an instrumental morphing into a voice.

Finally is when the generation goes wrong and it starts spitting out absolute nonsensical sounds with no rhythm or melody, in a uniquely fragmented way I can't really describe. It feels like seeing the musical matrix, the inner thoughts of the AI.

Now I've written all that out and had a think about it, I'm tempted to sample these oddities and try to make something more structured out of them.

Or, I can make a barebones recording of musical ideas - of melodies, harmonies and the overall song structure - then upload it to Suno and inspire its robotic session musicians to play it in the style I ask. Quicker to make, faster to iterate upon, and can even be used as the basis for resynthesizing the track using more traditional methods.

Half-finished tracks, ideas that were only in my mind or existed only as badly recorded piano-bashed drafts, now (almost) fully fleshed out. It's immensely satisfying, and has made me even more creative as I use this tool to understand and explore musical styles I'm less familiar with.