This is awesome! Game idea to build on top of this: Table top deception type game where each agent has the goal of convincing the real users that they are in fact also real users.(So each agent is trying to pass a turing test). Every AI agent uses RL to optimally prompt their personal LLM for how they should chat with the human players. eg should they try to frame a certain person, should they play it dumb, should th…
It would be pretty easy to beat. One sure way to tell an LLM apart from a human is to type something nonsensical and optionally repeat it multiple times. A human would inevitably answer with something along the lines of "what the hell are you on about dude???", which is something you'd never see from an LLM.
So you make it learn. It collects everything humans say and matches it to the situation, replaying human conversation as needed.