This is giving me a lot to think about. I have gotten frustrated with continuous models in 2018 because I've always had to take the result and make it discrete in a way that has heavy loss. When you have several layers of models inputting and outputting to the next, this become maddening. I didn't become a zealot or fanatic, but I cracked some wildly complicated problems by modelling discretely. I even did some regressions with objectives other than best fit with traversal tree searches.