For example, say you have one hour to pick five locks by trial and error, locks with 1,2,3,4, and 5 dials of ten numbers, so that the expected time to pick each lock is .01,.1, 1, 10, and 100 hours respectively. Then just looking at those rare cases when you do pick all five locks in the hour, the average time to pick the first two locks would be .0096 and .075 hours respectively, close to the usual expected times of .01 and .1 hours. The average time to pick the third lock, however, would be .20 hours, and the average time for the other two locks, and the average time left over at the end, would be .24 hours. That is, conditional on success, all the hard steps, no matter how hard, take about the same time, while easy steps take about their usual time (see Technical Appendix). And all these step durations (and the time left over) are roughly exponentially distributed (with standard deviation at least 76% of the mean). (Models where the window closing is also random give similar results.)
The logic is actually from Hanson’s “The Great Filter” essay.