> What, exactly, did M1 change

For instance, the unified CPU and GPU memory, which allows to run ML models on a Mac as if you have a large dedicated GPU. (Unified memory of course harks back to the 8-bit era; the key was to make it performant.)

> For instance, the unified CPU and GPU memory

Was already common and widespread with iGPUs in CPUs on x86, and was standard on all ARM mobile SoCs for a solid decade.

AMD in particular had already done a bunch here with their APUs including OpenCL support in 2012 and fully coherent shared address spaces with 2014's Kaveri

Pretty common in PC integrated GPUs, game consoles.