meanwhile there is more fixed function or special purpose hardware available (RT cores, Tensorcores) on RTX, I kind of want a system where there are no explicit shader stages, you just have "shaders" that can access fixed function and special purpose hardware (texture units, rasterizer, ROPs, RT Core, Tensor Core, etc.) and dispatch new shaders

Totally agree, this complexity is dissapointing.. if they instead had looked at the many ways people ray trace in a shader and given us more tools to make that more powerful..