Hi all, my first post here, and my first real experience with any kind of AI!
I’m developing a financial assistant for a FinTech client. She is a 3D character rendered in the web browser in realtime using WebGL, speech recognition and text-to-speech. She successfully chats with the user about many aspects of the web tool in which she is embedded and even gives presentations about the user’s financial situation. So far so good.
Now, however, I need a more powerful back-and-forth conversational capability and from what I’ve seen so far, OpenAI is probably exactly what I need.
I’ve successfully included OpenAI in my project, and created a few test prompts and received completion replies. However, I want my Ai to be called “Penny”. I did successfully did this with this prompt:
“Your name is Penny. What is your name?”
But in subsequent calls to openAI, the response is always that the Ai’s name is Sarah. Can anyone help please?
In case it’s of any use to other newbies, I’ve found a workaround for this: I append each prompt to the previous one along with the user’s response so that in effect I send the entire conversation each time I make a call to the API. This works: Penny remembers her name and the user’s name and the entire conversation so far.
Yes, I was about to suggest that.
But it’s an interesting question how to preserve new knowledge like a given name for the AI.
Thanks @360macky. It’s spooky how much she knows about herself after simply giving her a name and telling her she’s a financial assistant!
Definitely. It’s something I’ve thinking about for a while.
There are some good examples and templates, in the documentation (GPT-3 API)
Look for ones that set parameters for conversations between named agents, “chat bot” etc.
I put a prompt like that into my GPT-3 virtual assistant, too, but it keeps insisting its name is “AI” and not the name I gave it. Wondering what I did wrong – is there a guide somewhere on how to give your AI a name?
I don’t have a silver-bullet answer, however my (newly) experience with GPT is that it will not accept BEING someone it is not (“You are Georges” wont work), but it can PRETEND it is someone (“You pretend to be named Georges” works).
This is not failproof, as it will always break character when you ask “Are you really named Georges, or are you just pretending?”.
Hopefully future versions will have this feature.