I really don't believe exascale / GPU / neural / tensor / black box solvers are going to be the death of thoughtful code. Anything that's already being turned into a service for people who pay to just throw garbage in and get garbage out is a toy, not a tool. I remember around like 1994 when photographers were worried that Photoshop filters were going to let anyone do what they did. A few billion cameras and web apps later, there are still surprisingly few people who can craft a great photograph or distinguish one when it's perfect. An algo cooked up in a black box is just as good as the sample set, and all that's being sold out there now is a gambler's fallacy.