| HN Mirror

Interesting. I've been exploring this space too, but for a LLM-driven MUD (multiplayer text-based game). The most challenging aspect is coming up with an architecture that can easily lend itself to handling "semantic collisions" in a multiplayer context, pretty much like non-neural game engines handle spaceship collisions, or prevent a player from entering a box if another player is in it, but going beyond and handling more abstract, more "semantic" aspects, such as narrative elements.

By the way you should get interested in "narratology" at large. There are benefits to driving the LLM generation with an explicit representation of the narrative structure. For instance if the bad guy turns out to be the good guy, then, narrative logic demands that the good guy was the bad guy all along. As a consequence, appearance and deceit should be keystones values of the hero's arch. There are many interdisciplinary approaches to this domain, some more naturalist/empirical such as trope databses, some theoretical like 70s semiotics while some other framework approach the problem from the perspective of cognition and entropy.

Even though I haven't really started coding anything, this is a very interesting topic to think about because of the technical challenges it raises as well as the scope of the project. Any papers exposing new LLM use and prompting techniques is of potential interest to me now.

It is a wide space