Google recently showcased LaMDA which is said to be similar to GPT-3 but trained on dialogue. I was especially impressed by its ability to generate answers based on a given persona (see demo video for reference). I was wondering if it would be possible to achieve a similar result with GPT-3.
Does anyone have experience with this type of conversation, or have any ideas for how it can be done? When experimenting, I often get a generic response rather than one based on the persona. One idea I had was to make use of the
/answers endpoint, or something similar.
you could check out the sarcastic chat bot example in the docs. Might be useful as it shows how you can get it to respond with a certain attitude. I’ve been working with creating chat bots with GPT-3 and my assumption is if you want it to act like a certain character or have a certain characters dialogue style, say from a movie, then you should take dialogue from that movie and make that your prompt if you want to try an open domain approach. if you want it to talk only in one domain you could write your own dialogue prompt mimicking that persona dialogue style. hopes that helps
also I’d like to note I’ve only tried this once because I was trying to imitate Jarvis’s sarcasm from Iron Man and the results were ok but not great but my prompt designing back then was pretty bad.