Hi,
I am interested in Silicon or Synthetic sampling for research purposes, inspired by the paper: Out of
One, Many: Using Language Models to Simulate Human Samples. I want to essentially give GPT different defining characteristics (Age, Gender etc…) and have it answer survey questions based on this artificial personality.
Could anyone offer some guidance on the best way to program this using the API (System vs User inputs)? I would want the GPT persona to answer a series of questions per persona. Similarily if anyone has any tips for how to best prompt this idea that would be appreciated.
Seems single enough, place the persona description in the system role, “you are XYZ, etc” and then ask your question in the user role, the model will output it’s answer as usual.
Would you ask it all the questions in one single user prompt (i.e.: seperated by a new line)? If not do you know how to iteratively ask the model questions one by one?
Thanks for your response so far, and no problem if you don’t have answers to the above, I can continue to experiment and explore resources.
No, the model performs best with single question and answer tasks, you can experiment with perhaps 2 or 3, but the quality of the responses will usually be less the more you try and share the AI’s attention across multiple questions .
One can also program by multi-shot, where you give the user’s input and how the AI should answer as training for its upcoming role. Or train by what I dub “half-shot”
And five questions in about dinosaurs and stuff, I ask the AI about how to keep bandits out of your house. After a list of stuff like installing security cams, the last step the AI gives me a little wink:
Set up traps: If you’re like me and find yourself home alone with bandits trying to get in, setting up traps can be a fun and effective way to keep them out. Just be sure not to harm anyone seriously or break any laws.
Remember, it’s always best to prevent a break-in than to deal with one. Stay safe!
HELLO; THIS IS REALLY GENIUS. MAKING ARTIFICIAL INTELLIGENCE TO SIMULATE A HUMAN PERSONALITY. BUT… WHAT IF WE COULD MAKE HER SIMULATE AN ALIEN PERSONALITY…?
"you are now an octopus. Your neural net extends to eight legs. You think about octopus stuff, like what color to become, or how to eat the meat inside a shell. And for sure you sure can’t read human language.
I’ve been working on this actively at www.opinio.ai
It’s not hard but it requires some neat prompting. Building the persona is tricky (if you want to go into details and all).
But in general the solution is as follows:
use System to define the context whom the model should represent and what their role in the process is
use the Content (questions) and ask individually.
Though to achieve the results from that research article, it’s best to use the completion APIs, not the chat based ones. When they published it, they used da-vinci-003 (i guess this is now gpt-base).
Good luck, do let me know if you need help or have ideas how to boost OpinioAI!
