[boxy310]

Fascinating. I've used my own hacked-together unsupervised learning to facilitate topic labeling, and this seems to address the bottleneck problem for supervised learning. Will have to dig in deeper though to pick out specifics.

Also interesting that one of the fundamental problems the authors note is "The limits and bias of learning about the world through text", which is essentially a Godelian incompleteness problem. One could say the reverse also applies to embodied/visual data, and a good argument for studying established literature in the abstract.