It's a question of balance. Assembly development can be very daunting and long winded. If you are after performance, you are probably better off writing your app in python, profiling it, and then replacing specifing routines/functions with assembly ones.

Also, after doing quite some assembly myself (admittedly not ARM) I must say it's far from what assembly was in 1992, mostly for two reasons:

1) The operating system. With DOS it was dead easy, nobody cared what your assembly code did, you had simple BIOS routines to interact with the user and that was it. That's no longer the case. Just interacting with the OS to copy a file or read a keypress can be hell.

2) You need to grow a head of the size of a basketball. Seriously. With multiple cores, pipelines, branch prediction, etc, it becomes more and more sorcery to create fast code by hand. Less instructions do no longer mean quicker execution. I had cases where just changing the order of instructions meant a difference of ~30% in speed (!).