Indeed, optimization problems aren't all that exciting because there are usually efficient heuristics that get you 99% of the way there. A useful NP problem is

Indeed, optimization problems aren't all that exciting because there are usually efficient heuristics that get you 99% of the way there. A useful NP problem is generating code to meet some requirements. Instead of programming, you can write just some tests and get some code that passes them. Deep learning is a computationally feasible way of automatically finding a function that approximately matches some data points, but it doesn't always generalize well. Searching for the shortest program that matches all the data points should generalize better.