Turn one of those into "I would like the output to be factually accurate" and we have something to talk about.

Can I tell you why I hate developers? Let me tell you why I hate developers. THIS SHIT WAS SOLVED DECADES AGO.

Native Instruments didn't come up with this; Kyma had this shit in the 80s, back when it was a grad program side project at the U of Illinois computer science program. "Here's your sound. Hit the 'mutate' button and here are six choices. Pick one and hit mutate again. Here are six more choices. Hit mutate again. here are six more choices. Congratulations, you now have a completely different sound and you didn't have to touch a knob." Kyma even used a not-very-large language model to give you patch name prompts - you'd hit the "name" button and it would give you "adjective noun" that was often eerily appropriate through sheer naked pareidolia. That would be useful for the normies to experience, too - "look how any random language result seems eerily appropriate because you want so much for it to be so."

The really dumb thing is the LLM method of "composing" is the exact same thing Synclavier and PPG and Korg did going back to 1981: they held, in memory, a "wavetable" of different waveforms and depending on the inputs of the performer (velocity, aftertouch, modwheels, envelopes, generators) the synthesis engine would trace a path over the various stored values and interpolate intermediate stages between them. Except since it was music, rather than fraudbusiness, it was assumed musicians would want to get their hands dirty so all the parametric shenanigans were accessible. Most people didn't use 'em, yeah, but you mess with Kyma or Absynth for even a few minutes and you recognize that if you make the step size too big, you get unplayable garbage. My beef with LLMs is the charlatans selling them are dead-set against letting the normies see just how much shepherding the critters need to be able to fool the rubes into thinking they're alive.