• Previous part (3)

Yesterday I came across the most exciting AI news in a while, as far as my interests in the subject go. Matthew Berman, a YouTuber who’s always on top of this stuff, posted the following video:

A scientific paper claimed a method to achieve about 85% accuracy in modeling human behavior using large language models. I have read the original paper, which I found fascinating. In summary: AI agents, such as the ones I include in my Python project neural narrative, usually use either a demographic-based or a persona-based approach to mimic specific human behavior. In practice, that means that the agents are either provided demographic attributes or a paragraph summarizing the target person’s profile in order to perform tasks. My app, as many other modern ones, use a combined approach: demographic attributes and a summarized profile. The paper convincingly demonstrates that an interview-based approach is far more accurate at predicting human behavior. That paper even provides the base interview questions, although the interview process is actually dynamic, carried by an AI agent.

I intend to implement the interview-based system in my app: a whole module with an AI agent that asks a series of pre-set interview questions and possible follow up ones to an AI agent, which will answer the questions based on demographic and persona info that are already generated in the app. Once the interview is finished, I intend to use solely the interview transcript to produce any character’s actions and speech. I’m excited to find out how realistic the resulting behaviors will be.