Apparently there are parameters you can tune that cause it to produce different kinds of answers. And the answers in that linked article appear to actually be its attempts at jokes (because it’s interpreting each of the prompts as the setup for a joke). In contrast, see the more straightforward answers to similar questions in the image attached to this tweet from Gwern:
Trying it out with my favored settings (temp>0.9, top p=0.98, best-of>10), I do tend to get much more sensible GPT-3 answers with a more straightforward prompt than @JanelleCShane does. Her answers keep veering into comedy, which I think is part of the problem with her session.
Apparently there are parameters you can tune that cause it to produce different kinds of answers. And the answers in that linked article appear to actually be its attempts at jokes (because it’s interpreting each of the prompts as the setup for a joke). In contrast, see the more straightforward answers to similar questions in the image attached to this tweet from Gwern: