I'm a C++ developer in an embedded system where speed counts. I still live by that badge. I know that if I find a particular tiny area of code that is a bottleneck I can drop to low level assembly and after a year of work beat the compiler by a tiny amount on ONE of the THREE CPUs we support - but it would take me an entire year of nothing else.

I know of thousands of places where our code does not use the best algorithm (they are not bottlenecks). I know of hundreds of real bottlenecks that I think could be fixed with algorithm work, but the bottleneck isn't large enough to be worth putting effort into. (this decision is subject to change)