on using AI for art

regarding (my) 'ai generated' art stuff, i never want it to look at all 'correct' or 'professional'.  i specifically would never use AI to generate art that looked like it had been made by a human, in fact, i'm not precisely sure why but i think that's stupid.  there's already 8 billion plus humans, they are fine at doing art that looks like humans did it.  in fact, they are excellent at it.  getting a computer to do that is just dumb.  i want something that looks like, or sounds like, a human COULD NOT have made it.  right?  apparently that's crazy and literally nobody besides like 10 other IDM artists are working around that concept.

my whole point of using AI to generate stuff was just that i was bored with the limitations of my creativity which are bounded by my obsessive need for organization and boringly precise arrangement.  having studied and worked with purely deterministic systems like, every x bars pitch shift n notes, etc, i really liked that concept.  and i liked that you could introduce randomness into the equation too, letting the universe influence your work.  (letting the wind blow through your reed)

but an llm is a whole other level of procedural machine, and the results were fascinating from the start.  partly because of the way the llm would misinterpret every other thing i told it to do.  sometimes it would take my boring idea and then implement it in the stupidest, most backward way -- and, then that would produce the most fascinating bizarre output i could have never imagined.  

we went back and reverse engineered some of the bizarre music generation processes that were created in some of these applications and just stared at them like 'that's so stupid...' then we would go and listen to the midi it output and be like ... what the ever loving christ is that

one recurring theme through musical analysis by LLM is the theory of music being represented by curves.  somehow a frequent perception of the LLM is that the notes in musical melodies and rhythms seem, to it, to be curves with the notes as points on a plotted arc of some kind.  another thing an llm decided to do was, instead of 'extracting interesting patterns of musical bar lengths' (my naive instructions to it), create histograms of multiple analyzed midi files, determine the most popular, essentially, arrangement of notes in a bar, and repeat those, sometimes with a curve.  fascinating.  never would have created that with my own brain.

then i took these sequences of notes and i loaded up old musical projects i had done and replaced my bits with its bits, moved them around, shuffled them, applied various deterministic sort of pitch shifting hacks to them, etc... but never actually rearranged or significantly changed the order of the original note data.