GPT-3 is an exceptional mimic. It looks at the text input and attempts to respond with what text it thinks best completes the input. If the first line sounds like something from a romance novel it will try to continue writing in that style. If it’s a list of video games, it will try to add to that list.
If it sees the letter “Q” and a question, it will attempt to write a response…but not necessarily the correct response. Without any additional context it might assume you’re writing a list of questions and wrong answers. Or answers that a third-grader might respond with.
If you want to get the best answer, you need to let GPT-3 know what kind of answers to provide.
Instead of asking GPT-3 a question like this:
You can improve results dramatically for complex questions by using a prompt like this (or some variation):
Common sense questions Q: A:
Usually the presence of a phrase like “Common sense questions” is enough to get GPT-3 to understand that it needs to give answers that are more in line with what you’re looking for. You can also use a phrase like “Factual response:” instead of “A:” to improve the result.
Common sense questions Question: Factual answer:
This instructs GPT-3 where to begin in the world of possible answers it can provide. Just using “Q:” and “A:” isn’t always enough, especially with complex questions. GPT-3 needs to understand what kind of answers it’s supposed to be providing.
Developers exploring the capabilities of GPT-3 often make the mistake of treating it like a specially-trained question and answering model and forget that prompt design is important.
GPT-3 can do a great number of things – which is why it’s very important to be specific about the task you’re asking to perform. Otherwise it will try to mimic the first thing the algorithm hits upon, which might not be what you’re looking for.
To have a system capable of creative writing about situations and things that don’t exist and be good at answering factual questions, context is everything.
GPT-3 is just one step towards advanced AI and has limitations. It doesn’t know everything and can’t reason on a human level – but it’s highly capable when you know how to talk to it.