Indeed, GPT-3 is far from AGI. But in this relation we have to re-consider what to understand under "comprehension". GPT-3 can summarize complex texts with short memos. It also can apply different approaches very convincingly (see the "Love letter written by a toaster"). Or it can perfectly apply an art movement method on description of a situation (advanced step in human creative writing).

Of course, if GPT-3 is writing about an "apple" it doesn't know what is an "apple", besides of it's "eatable", "round", "a fruit", "Paradise", "Mac", "Eris, Hera, Athena & Co", "Turings death" etc. Transformer does good work with associations from a huge source of training bases. GPT-3 lives within text after all, it has no other data and knowledge besides of textual dimension.

I share your concerns with issues of overhyping GPT-3 (like every other AI model) and causing another AI winter. But in my experience, GPT-3 is a new milestone in an NLP journey, whose writing ability and creative power cannot be denied or deminished.

Compared to GPT-2, which used already self-attention but could be easily detected as AI-generated contents even by human reading, GPT-3 bears new possibilities, at least for creativity.

One thing is clear, everything depends on human use of it. And we are just at the beginning.