Two prompts sent to chat:
- What do you know about github?
- Write a short poem about github
The first query times out, the second produces a poem.
(I know gpt is not an expert system, I am happy for incorrect answers to factual questions).
I get better results by reversing the order of the queries. Poem first, then “What do you know”, but still mostly times out.
I have noticed this generally. Requests for creative output (poems, limericks, sonnets…) are much more reliable than the “ask an expert” style of question.
The “completions” API with “text-davinci-003” does not have this property.