The main thing to consider is the author choosing the sentences. The author says they only made a couple( less than 10) edits. But the human input is what is important. I remember something similar when GPT2 came out(I think it was an interview conducted by GPT2, with human tuning).

I have been following the twitter hype on GPT-3. @quasimondo has done some excellent analysis on text generation using GPT-3. For me, when reading GPT-3 text, something is off. It jumps around too much. It does not seem to care. I couldn't read more than 10 sentences if the text is completely GPT-3 generated. But if prompted properly by the human at regular intervals, it sounds ok. Generally it holds a thought upto one paragraph. So maybe the usefulness of GPT-3 is in scenarios like, I have a thought, I want to complete it, I fire up GPT-3, prompt it and generate different results, choose one which is closest to my thoughts and edit it and publish. I am sure that's what this blog author did.

GPT-3 is really good at NLG. But people seem to be extrapolating the results to NLP.