[mathieudombrock]

This is a neat idea and I wish it worked. I've spend hours and hours trying to get LLMs to be a "dungeon master" for text adventures. I've written a good amount custom code trying to facilitate this. Trying to force the LLM to keep it's story straight.

I'm pretty convinced that the current generation of LLMs is nowhere close to being capable of this. No matter how many context hacks you throw at it.

It inevitably derails and ruins the immersion.

Best of luck on this. If you can pull it off it would be really cool I think.

[benbreen]

Currently working on an idea like this, but its a history simulator for educational use - I find that LLMs respond rather well to being grounded in a specific time/setting in real world history, as opposed to being told to roleplay a fictional setting. The latent space of any fictional world is close enough to other fictional worlds that they will rapidly slide off into other similar-sounding settings. Whereas if you return them to a super-specific historical context each go-around ("The time is now 3:13 pm. It is August 3, 1348. You are currently simulating the functioning of a small vineyard in Normandy. The farmer, [NPC name], is looking for helpers in the fields") they will be able to pull from a pretty solid baseline of background knowledge and do a decent job with it.

Some fun things I've been experimenting with is 1) injecting primary sources from a given time and place into the LLMs contex to further ground it in "reality" and 2) asking the LLM to try to simulate the actual historical language of the era - i.e. a toggle button to switch to medieval French. Gemini flash lite, the only economical model for this sort of thing, is not great at this yet but in a year or so I think it will be a fascinating history and language learning tool.

Have been meaning to write this project up for HN but if anyone wants to try a very early version of it, it's here - you can modify the url to pick a specific year and region or just do the base url for a fully random spawn, i.e. here is Europe in 1348: https://historysimulator.vercel.app/1348/europe

[ianbicking]

I think a big part of it is not so much that they aren't capable of being a dungeon master, but they are constitutionally unfit due to their agreeability.

The biggest improvement there is to treat the game engine as the "user", and the player (and their input) is merely one among many things the game engine is managing. But then you also need a game engine that manages lots of the state programmatically, with formal transitions of that state. The LLM can manage the state transitions, but the actual state needs to be more formal.