This is great.
Epic has been submerged in the Fortnite life for a very long time now, and a large part of that is probably learning first-hand how difficult it is to take a single idea, or a single set of artist produced assets, and take them through a pipeline for each and every platform to optimize along that path. The idea of mapping 1 pixel to 1 triangle (and theoretically culling everything else, through whatever virtualization this is), makes rendering seem more like a cost relative to resolution, which naturally scales up and down depending on how powerful the hardware is. This *seems* like it would take a huge load off of the shoulders of artists who have been given the difficult task of making sure their crazy high-poly models can exist in that iPhone version of their game. It's a horizontal scaling method that seems really powerful and will save a lot of pain and time on behalf of larger companies that can afford to think in the direction of many skus, and make that considering a possibility for even the smaller teams. It kind of makes sense that some of this is born out of the battle-testing grounds that is Fortnite -- it's more than just a game to them at this point, or some social experiment, it's an honest-to-god testing ground for the latest and greatest technology.
I'm a little... confused, or maybe underwhelmed that the solution here for GI is not RT. I actually don't know exactly what it is, but I am reminded that at some point there was a voxel octree GI solution (or something like that) made for UE4.x that got cut due to hardware not being there; my guess is, this is that, brought back, better and more refined. It looks great, mind, and I realize that this doesn't mean reflections or shadows are going away -- those are still options. But it does mean, I think, that Epic has chosen to steer clear of the route taken by, like, the Metro guys or Remedy. This is their solution, and they think it cuts a perfect middleground -- for consoles, they are probably right on the money, and this leans in hard to the earlier point I was trying to make.
Overall, super excited. This stuff really does seem next-gen to me -- and a huge paradigm shift in terms of how we deal with detail in games. It reminds me about the granularity MS was showing off with meshlets in DX12; being able to cull to almost perfect precision high density models bought them a lot of performance, so it stands to reason that if you can cut off all waste you can reasonably do what they have set out to do here.