I would like to explore using GPT-3 as a general question-answering resource.
Can anyone link to any systematic study of what GPT-3 knows or how successful it’s been on some general knowledge test?
Should I assume DaVinci with low temperature is the standard choice or is there an optimal set-up for the most accurate possible answers (along with sending examples in the prompt)?
If I decide to make it more domain-specific, what options are there for giving it the information? Do I have to pass it as the prompt, or can we train GPT on new information ourselves?
Thanks very much