Conversation over coffee this morning about the costs of a neural net for collapsing 30-40 parameters into 3 or so, and I strongly suspect that for most data sets I could implement a 3d parametric curve fit that's close enough and way faster. But wouldn't be "AI".