Earlier quoted context omitted.
Dan Luu also elsewhere makes the argument that most of the impactful engineering work done in large companies is performed by a small percentage of the engineering group.
That seems consistent. If there's a 1% chance of a meaningful improvement, and 99 out of 100 changes do not meaningfully improve things, than the challenge is knowing beforehand what the most impactful engineering work is going to be.
If the payoff in the 1% case is high enough, then you don't have to know beforehand! Just do all 100 changes and the one winner pays for all the rest.