i think good dynamic quests will eventually be a thing but not coming from gen AI or old school complicated rule books. if you generate the whole thing with a LLM you get a lot of results that make no sense. if you go with a deterministic rule based generator or a constraint solver you get low diversity because the rules are fixed.

my solution is a hybrid approach using a machine learning model trained to estimate player ratings for a quest script. you have a simple RNG that generates a ton of different script, the model picks out the ones that are likely to be fun and you give one of them to the player.

the meta might be using all the methods (maybe not LLMs) together to get a result thats fun, logically consistent and even fits into a larger narrative. like tracking NPC relationships to assign a target for combat missions, coming up with extra plot twists along the way then using ML to rate all the generated options.