Interesting. BTW I thought it was ironic that in the first part of the post the author points out the lack of logical connections between sentences in GPT2-gerenated texts, while in the second part she changes the topic to performance on IQ and math tests, which have nothing to do with how well people can detect the failure-modes of GPT2 - after all, people with high IQ or math scores can easily be inattentive when reading texts. Maybe this post itself was generated by GPT2?!?!?!?!? (or GPT...3?!?!?)