It's the verbatim output of the forth or fifth sample I got out from GPT2 1558M after a short prompt that explained that GPT2 and GPT3 were AIs from OpenAI and that GPT3 was better, that GPT2 was asked if it was jealous and it replied ". (unfortunately I've rebooted so I don't have the prompt in my terminal buffer anymore).

GPT2 output quality is highly variable. Well crafted prompts can inspire some great output without much digging through samples. I believe I've gotten much better at writing prompts-- even using the right typography matters-- after spending some time doing it.

It sounds like GPT3 gives something much closer to best-of-gpt2 all the time. I hope I get access to it at some point (I also hope the api exposes raw probabilities).