Very interesting. Is there a real-world application for this?
The idea is that you can model an infinite set of points generated by an infinitely-deeply-nested tiling, and then efficiently sample from it at arbitrary levels of detail. This lets you do things like zoom in and out while retaining frame-to-frame coherence, and without having to explicitly store the entire point set. (It's worth watching the full 5-minute video, which explains it better than I can.)
The implementation in that paper used squares with labeled edges to tile the plane non-periodically. This seems like it could be used to do something similar, except with a periodic but irregular tiling. No idea if that would have any practical benefits, but it's interesting, at least.