A few months back, I started trying to tackle this problem, initially with prompts and eventually with models fine-tuned on character interactions.

I created this website as a testbed for the concept, and I open-sourced a fine-tuned model for those interested in experimenting with this themselves [1].

During the instruct tuning phase of training, LLMs learn to act as assistants. And they need to stop acting as assistants to produce engaging interactions. In Spellbound, telling NPCs they are AI will generally cause them to be confused or protest that claim. Most NPCs are "convinced" they are real people who exist separately from the LLM.

One downside of this approach is that the guardrails formed during training generally stop functioning [2], so it's a technique that needs refinement before being applied to more mainstream applications.

We're really interested in applications of LLMs for simulating characters and worlds without the typical blandness of an instruct-tuned LLM. Solving this problem will unlock a ton of use cases in media where AI currently doesn't cut it.

[1] https://huggingface.co/hf-100/mistral-spellbound-research

[2] Regardless of how you feel about guardrails, some ability to guide content is useful for making children's stories, for example.