Interesting idea. Basically they reach framerates high enough that a lot of the temporal filtering actually happens in your eye (and your brain) instead of in the GPU.
While I think the idea is a little too bold for actual products, it does point out other pardigms that open up once we actually go full RT (secondary AND primary rays). Instead of thinking of rays per frame, devs can think in terms of Rays per second. The bottleneck then is only on animation side, but rendering becomes mostly framerate independent (as long as your temporal accumulation and filtering can stay stay slim). It's an extention of what already is being done with temporally reconstructed checkerboarded buffers (or other sparse sampling techniches that are rasterization-friendly) only that when we let go of rasterization, sparse sampling becomes much cheaper, easy to implement, and more flexible. And once we are there, of course we can also start talking about temporal jittering for high quality MB and actual lense optic simulation based DoF and Bloom which will be relatively more trivial to adopt when sparse stochastic-like sampling + smart reconstruction and denoising already is implemented and a fundamental and robust part of the rendering pipeline.
