I mean, they've got that double rate matrix multiplication pipe in RDNA4, and have had that huge wave matrix multiplication register in RDNA3 for a while, making use of those makes sense.
AI is better at figuring out things like cheap upscaling than humans are, given enough guidance. I suppose they could've just reverse engineered what the AI is doing (
for those wondering how) but then again what the AI is doing is assumedly based on low bit depth matrix multiplication anyway since that's what it has access to and will be rewarded for optimizing for.
Either way, looking forward to more competition in upscalers. XESS is mostly good but has some ghosting problems on occasion.