to me, i initially thought it was just some version of the latter, but reading the blog post in detail, i think the wording leaves open the possibility of both.
people do create activation atlases to try to visualize the concepts captured in a given computer vision model. if you can do that, i don’t see why it would be impossible to associate some of the more distinctive concepts with words that describe them. that said, that seems complicated, and i would be surprised if any version of this could be created that works well, at least in the near future with the budget contemplated here.