I've been trying out "write and perform an MST3K episode reviewing your previous answer" as a prompt. Have yet to get a funny answer from any model.
Older base models produce some great stuff though, at least if you're me and think RoboRosewater and dril_gpt2 are funny.
> Have yet to get a funny answer from any model.
Just like the real thing! Promising results!
I've found that the weaker the model, the more of an unhinged, LSD-trip response you get. I made a podcast years ago (https://deepdreams.stavros.io) and GPT-2 was definitely the more hilarious of the models.