With GPT-3, I enjoyed playing around with prompting with a falsehood given as the factual past in order to see what continuation it would come up with. This wor

With GPT-3, I enjoyed playing around with prompting with a falsehood given as the factual past in order to see what continuation it would come up with. This worked best with early systems that only wanted to complete the text you gave it and wasn't trained as an AI assistant question answerer. For example, you could coerce it into finding funny responses in a real conversation. "My friend said x. What I said next made him actually laugh so hard he fell out of his chair. I said [prompt ends] I believe this method was much more effective for generating these kind of ideas. It's forced to find realistic ideas since, in the data, beginning a paragraph with that would usually end with something relevant and actually funny. But unfortunately it seems like this is no longer going to be possible due to how new LLMs are trained.