I recently wrote a simulation that involved some heavy numerical calculations. Without optimization it was dog slow and would have taken probably years to produce the results I needed. With compiler optimization it took a few days, which was good enough for me. Perhaps I could have improved on that with manual optimizations, but the speed was good enough so instead of spending time optimizing I did other things.

The implicit assumption in the talk (based on this summary) is that we always want the most optimized code. But in fact that is never the case -- what we really want is for the code to be fast enough (or be space-optimized enough). So the relevant question is this: in what percentage of cases do compiler optimizations take us from 'not fast enough' to 'fast enough'. In all these cases the compiler is useful, because it saves us the cost of doing our own optimizations. In my experience this percentage is far from negligible, so at least for me optimizing compilers are highly valuable.