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.

I'm familiar with the concept[1], but I'm unsure if it's a good showcase of code complexity. I've tested some internal microservices I'm deeply familiar with and found no correlation...

[1] for the past ~15 years actually, got introduced to the concept through works of mr. Hutter, after becoming aware of his Prize, and I'm dabbling in compression to this day (right now trying to improve on Bellard's nncp)