"Sufficiently Smart Compiler", like most AI, is a concept with constantly shifting goal-posts. As soon as compilers can do something, we no longer consider that thing "smart." Consider variable lifetime analysis, or stream fusion -- a decade ago, these would be considered "sufficiently smart compiler" features. Today, they're just things we expect (of actually-decent compilers), and "sufficiently smart" means something even cleverer.

And, given those optimizations, the programmers get sufficiently dumber to compensate, resulting in a constant or decreasing level of performance.

That's gotta be a law codified somewhere, right?