Compressing LLMs Without Compromise (sigarch.org)

Ok, that was a bit hard to read but pretty interessting. I wonder to what degree compression already is integrated in the base models.