I'm not sure how to articulate this. I have synths that I can program in my head because I know the architecture really well and when I imagine a sound I can walk over and dial it in from the front panel and get more or less what I expected, plus I can then tweak the results with abandon and for musical satisfaction. Then I have others (sometimes from the same manufacturer which emit all manner of nice sounds but are far harder to program and easily veer off into sonic mush - technically impressive but not really fun to play.
I also think wavenet sample by sample generation and interpolation between latent features doesn't sound that exciting, as cool as it technically is.
We'll find some place to use machine learning in music/audio eventually :) I think perhaps more natural sounding pitch shifting could be one area (since you could learn the structure of sound of different instruments at various pitches), reverb removal, denoising, polyphonic audio to midi - things like that, where you have obvious training data.
Well, they are probably not tuning 100 instrument parameters but only the play characteristics like accent, velocity and so on.
One thing I thought would be pretty cool - I've got a friend who did a phd on physical modeling of sound in Edinburgh - http://www.ness.music.ed.ac.uk/. Often those physical models have hundreds of parameters and it's quite difficult to tune them to make meaningful/good sounds. Perhaps a neural network would be useful for tuning the parameters - you could use a real musician playing a real instrument with sensors, and use the generated sensor data and audio for your training set. And then the neural network could learn to match the physical model parameters to that.