> was working in a space where fast inverse square root was the difference between the program working and it not working
Pentium 3 released in January 1999. That CPU, along with all newer ones, support SSE which has RSQRTPS instruction, faster, also much more precise approximation.
AMD had it before, PFRSQIT1 / PFRCPIT2 in 3DNow from 1998.
Quake 3 Arena released in December 1999. The point of using that approximation is running on older hardware, like Pentium and Pentium 2.
So, basically a difference between the program working and not working for most users.
I think the difference here is that Carmack was working in a space where fast inverse square root was the difference between the program working and it not working. At that point, it doesn't really how matter "clever" in any bad sense of the term the code is.
Carmack works in 3D engines, which even today is a space where people are trying to wring every last cycle out of them. It's perfectly appropriate when programming in a space where every last cycle counts to program as if you're in a space where every last cycle counts.
I don't work in such a space, so I don't program that way. But while I haven't got anything as "clever" as FISR, I have got some things like non-obvious database schemas for certain performance reasons that if my successors don't find and read the comments I've left about why the schemas are that way, they could well get themselves into some trouble if they try to "fix" it.