I was actually kinda moved by the Global Illumination with Radiance Regression Functions clip. However, it appears scenes are static and have a huge number of vertices. What do y'all think?
It looks great! Unfortunately, giving it a quick glance over, it doesn't have anything like a constant cost and each additional light evaluated adds to the cost. But the main problem seems to be that it doesn't scale with scene complexity that well. The more of their "RRF's", which is a section of level or model, that you have to evaluate the more your runtime cost is, and the bigger the "RRF" the runtime expands even faster.
Thus a small scenes works very well, big outdoor scenes don't, as it doesn't scale and doesn't handle environment (read, skylighting) well at all. And big outdoor scenes are the major problem for GI to begin with. There are solid enough hacks for short range lightbounce, even Crytek's light propogation volumes can work well enough for that, you can raymarch for reflections and voxelize secondary geometry for secondary shadows. It's the typical huge world that needs a good solution.
I was starting look up raytracing into signed distance fields, it makes quite a bit of sense from my initial impressions. If you could combine it with voxel cone tracing, which has a constant overhead cost, scales well with scene complexity, and can easily add lights, then the only thing you'd need to handle are dynamic objects. Which could be handled with something like Sucker Punch's ambient occlusion fields and a hacked impostor for reflections, as you don't care about most dynamic objects influence beyond a small area. Not perfect, but then perfect isn't doable yet.