There are portable (or not) software applications, portable algorithms code, non-portable algorithm code.

Assembly based software for the Motorola 68000 /and/ the Amiga will not run on a 68000 Mac. A "polynomial based packed fastSin()" subroutine written using the XMM x64 registers will work on most x86 CPUs. The same written for the ZMM registers will not work on a number of x86 CPUs.

Surely 40 years ago we could easily assume talking about specific implementations of software applications; clearly today we see problems as being optimizable on some classes of platforms (e.g. the vectors in ARM work differently from the vectors in x64 CPUs).

If you can understand what someone means when they talk about a “small elephant”, then you can understand what they mean when they talk about “portable assembly”. In this case, the relevant point is that you can write ARM64 assembly routines that do useful work (e.g optimized matrix multiplication, or something like that) in such a way that they’ll work correctly on a number of different ARM64 platforms.