TEXT-DAVINCI-003 is too optimistic. It tends towards hyperbole, maximizing, and superlatives. Even when you ask it for nuance it tends towards being positive, optimistic and ignores negatives.
For instance, I asked TEXT-DAVINCI-003 to grade a resume based on fit for a role and it invariably says “Yes this candidate is perfect for this role!” even though they had none of the required skills or experience.
This is feedback for OpenAI - do not let your focus on “alignment” send you down the toxic positivity path. We need a product that adheres to TRUTH first and foremost. If I want rose-tinted feedback I’ll ask for it. As it is right now I have to spend quite a lot of time with prompt engineering and decomposing problems to get honest feedback.
Here’s an example of what I mean. This kind of prompt generally won’t work. It will always answer the question in the positive.
Is this candidate a good fit for this role? Yes or no, why or why not?
However, this kind of prompt does work:
Criticize this candidate. What are their greatest weaknesses or gaps? What's wrong with them?
You have to go completely overboard to get TEXT-DAVINCI-003 to be judgmental at all. The problem here becomes that you must explicitly ask for criticism, so that’s what it gives, whether or not it is warranted. This is problematic because then you have output that is explicitly (and always) negative, even if it just made it up.