Agreed with all of that. I would add BDDs and such in there, as great breakthroughs. In a very real sense, they are essentially moving how the computer builds up logical circuits into a data structure. (That is, I'm asserting that the math unit on a computer is closer to a BDD than it is any other data structure in computer science.)

But, even outside of advanced modeling, I'm always amazed at how much effort we go through to not reduce our problems to the computation that we want them to be. Instead of seeing our programs as tools to translate from a high level model to a low level one, we seem to want that low level model to be completely implicit and not visible in what we build.

Even code I have seen that uses solvers, there is rarely a good separation of "convert this problem specification into a SAT problem" with a corresponding "convert this SAT solution into the problem specification."

I /think/ this is where programs that do focus on text seem to have a win. You have a "how to convert this text to a SAT problem", "how to convert this text to a problem specification", and "how to convert this solution to a problem specification." Stabilize what those text representations are, and you get to spend a ton of effort at each section without immediate impact on the other areas.