> I think it is just pattern-matching. > For role playing, the pattern-matching approach will probably work well. The potential for using LLMs to depict charac

> I think it is just pattern-matching. > For role playing, the pattern-matching approach will probably work well. The potential for using LLMs to depict characters for movies and games appears staggering. I deeply agree with this. Role playing and simulating agents (ultimately via clever prompting) seem to be the tools we need to get to the next level. I've had an unbelievable amount of success by simply iterating a dynamic prompt and keeping track of the last things the LLM did and any relevant error messages. The LLM part is mostly solved. The new challenge is role playing to such an extent that you can engineer agents (via prompts) that are difficult to distinguish from the real deal. The other hard part is figuring out how to expose your domain as functions to the prompt. What seems obvious usually doesn't work. You oftentimes need to decompose the activity to accommodate context limitations - Multiple modes of operation become necessary for more complex agents who can't reliably consider all of their capabilities at once. To put this back into the language of the article, I would say the real magic trick is figuring out how to project a rule-based problem domain into a pattern-based one (i.e. prompt) so that the LLM can make an effective decision.