These rules are good as far as they go. However, they are notably silent on what happens when they stop working. If you're a celebrity computer scientist and a

These rules are good as far as they go. However, they are notably silent on what happens when they stop working. If you're a celebrity computer scientist and a problem has been 'squashed' to the point where there aren't any more 80-99% "hot spots" I guess you can just parachute out of there and on to a more interesting problem. However, if you're paid to work on a particular thing, sooner or later, you will fire up a profile and say "oh, great, I don't have a hot spot anymore, just a bunch of 10-30% 'warm spots'". At that point you need to attack problems that aren't traditional bottlenecks if you still want to get speedups. Moreover, the things that you learn from repeatedly attacking those 10-30% 'warm spots' might be fundamentally different from the learning you get from, I dunno, taking some O(N^2) monstrosity out of the "99% of the profile" and Declaring Victory.