Yeah, I think it leaves us back at square one: we don't know much about how it really works.

I don't follow it too closely, but I've seen papers on mechanistic interpretability that look promising.