I think it's partly the regular, simple, easy to understand structure that makes the approach outlined in this paper fail so badly. Because CPUs are made up of

I think it's partly the regular, simple, easy to understand structure that makes the approach outlined in this paper fail so badly. Because CPUs are made up of a number of simple, regular pieces connected in a way that's generic and easy to reason about, each of which is general enough to be used across many kinds of programs, knocking out individual transistors and observing the effects on the output of complex programs doesn't tell you much. Which parts of the CPU are used for a particular "behaviour" depends on the whims of the programmer and the compiler's optimisation passes as much as anything else.