CPU's are not changing their instructions based on driver updates either.
IMHO, time for CPU's to offer thousands-of-floats-wide SIMD instruction sets instead of the measily 512 bits, that can do the same as GPU's but are way more stable, and get rid of those annoying, version-of-driver-dependent, install-100-python-libraries-to-use-them GPU's.
GPUs will never be outperformed by a CPU with wider SIMD. Intel learned that the hard way with Larrabee. They are on a different place on the pareto front.
When did compiling of shaders become a thing? Nvidia GeForce had shaders in the year 2001 but I don't remember ever seeing this then. The first time I saw compiling of shaders happen was in 2017 in a console emulator I think.
So how did it work before then, and what exactly has this compiling now improved compared to before?
Internal pipelines can vary between GPU architecture too. Currently AMD ships a different branch of their own Windows drivers for the new RDNA3 architecture GPUs (RX 7000 series). So precompiled shaders probably work won't between architectures.