That the definition of concept you're citing here uses the term "understanding" is incidental - clearly it's a definition of concept in the context of human reasoning.

Whatever terminology we use ultimately for the connectedness of neural networks pre-trained on language is fine by me. It should be as precise to the technology as possible whilst conveying effects of "intelligence" that are appropriate.

We're at the point now where GPT-4 seems to produce connections that come from a place that's difficult to find or reverse engineer - or perhaps which simply come from token selections that are surprising.

That's what I take away from a lot of the discussion at the moment - I have no personal insight into the model's design, or the many parts that are stitched together to make it work as it does (quoting Altman here talking to Lex).