There are certain features that can't be replicated with ML-based rendering, or are poorly suited to it. It doesn't have a concept of worldspace, for one, which makes it struggle with light propagation and game logic in particular.

It's more likely that they'll complement each other - tagged geometry + hard raytracing for certain features + ML style transfer/rendering. At least until convincing and fully hallucinated worlds appear (which might or might not happen).

the approach to RT is basically cramming more and more cores and next to no specialised hardware anyway... so it is a safe path to take regardless what the drivers end up doing with them later

it's all about more cores, faster cores, more VRAM, faster VRAM - I doubt any research achieving any of that will go to waste

No, AI rendering is crazy slow and will remain that way for a very long time.