Sampler feedback streaming helps the implementation (less code needed to calculate and tell the system what memory pages to stream in), but does not give you back that much memory compared to best in class virtual texturing developers used to implement in software (think Xbox 360/PS3) or using features like PRT (partially resident textures) to only keep visible textures loaded in GPU memory and caches):
The stats you quote compare a title with SFS implementing a form of virtual texturing vs a title not using virtual texturing at all. So, it might make it easier for small indies (that were not using UE4 or Unity3D but their hand made engine), but not such a big world changer for AA/AAA+ titles.
ML wise, support for lower precision integer math at higher rate helps a lot, but it uses the same shader array as the rest of the rendering does (same for mesh shaders) instead of having super powerful dedicated unit like the Tensor Cores nVIDIA uses separate from the Shader ALU’s. MS and Sony will try to flex the importance of ML, but neither has anywhere near the computational power dedicated to it that nVIDIA or Apple have in their devices (I would have expected a lot more iPhone games to leverage DLSS like solutions sooner to be fair... although their Neural Engine is a fully decoupled component in their SoC so maybe it is not as easy to use for this kind of use case at a low enough latency).