I think procedural story generation is actually best when there’s a simulated autonomous system and you just watch it play out according to its own rules. LLMs can’t touch this, though i think they can emulate something that looks like this at first glance, poorly
the critical difference is that an LLM understands nothing
it's a slot machine that pieces words together into sentences; it can play the odds but it doesn't have a "mental model" of whatever it's describing
what Dwarf Fortress is doing is describing its internal world which abides by its internal rules; the behavior is programmed but the specific interactions are emergent
