Hi,
I think the discussion is really interesting - since LLMs basically contain all books ever written, mirroring fictional settings against that is pretty cool (imo).
I came up with an open system based on the world wide web; like an interactive fiction web, where users can create documents that link to one another, and users can browse this interactive fiction web using an LLM attached to a normal web browser.
Since I’m new to this forum, I cannot post links, but the main architecture components are the following:
- World Text Markdown Language (similar to HTML) – Specifies World. Includes link to other WTML-documents with proximity-criteria. The LLM dynmically calculates whether these criteria are met and the new WTML-document is loaded.
- World Text Transfer Protocol (similar to HTTP) – Prompt-Pattern, which specifies communication with LLMt. It returns a XML response, based on which the HTML/JS-site for the user with choices is generated by the browser. The choices already include consequences for the world state. It could include a field for the user to enter custom prompt what to do next, from which the consequence is derived on the go by the LLM
- World Web Browser (to be run inside a Web Browser with connection to local LLM) – Let’s the user browse the world web
I already tried this out, however, since I’m running it locally, it works pretty slow and the LLM still has the same problems discussed in other threats of this forum (trouble keeping track of current state; generated text is boring - it would be cool to have a CERN for AI as Tim Berners Lee proposes, which could result in more creative LLMs.).
But it’s overall great fun. I hope this isn’t too off topic. Would love to hear you thoughts on this 