When it comes to loading pig iron, you're really just doing the same thing over and over, so the amount of pig iron you can load in a day scales roughly linearl

When it comes to loading pig iron, you're really just doing the same thing over and over, so the amount of pig iron you can load in a day scales roughly linearly with the number of hours a day you spend loading pig iron, all else being equal. We know that the design of software is fundamentally a different sort of task than loading pig iron. It requires management of complex amounts of state in working memory, which must be acquired before work begins, and much of which is lost after work ends (or is interrupted), requiring reacquisition when work is restarted. I believe that because of this, Carmack's observation is that more contiguous time spent on the task yields superlinear gains in output, because you're just spending that much more time with the system state in your head. And if you're a really smart guy working on cutting-edge stuff you want to take advantage of those superlinear gains because you are messing with much more deep and complex state, in terms of your own cognition, that would take more time to reacquire when lost.