It comes from the "intelligence is a form of compression" hypothesis that has been floating around in the ML space. Also, with a good compression algorithm it is a fairly direct measure of entropy, which is quite well correlated with what a developer might consider code size and/or complexity.