It seems to work in the short term, for generic adventures, but it has no sense of place. What I mean is it doesn’t understand geography, or the layout of a city, or factions, or metaplot. It’s not creative in that way. If you ask it to create factions for a city it will do it, or a city with particular history and culture it will do that, but if you ask it to run a city adventure, it doesn’t know it needs to do those things first. So you’d need to hand hold it through generating a specific setting, and then you’d hit the token window limit.

1. I openly question gpt's ability to generate compelling narratives. Perhaps modules could be feed into it to provide enough background for it go off of, but I would consider this a weakness to GPT compared to humans.

2. Current GPT based DMs allow the player to do basically anything they want. Human DMs offer more control over the game, and will follow closer to tradition and unwritten roles than GPT allows. There's also a fun part of the game for the player in which they convince the DM to allow an action they wouldn't otherwise.

3. D&D's low tech nature is a draw for a lot of players. Some books, some paper, pencils, dice, a table. Simple stuff. I think many players would push back on screens and other high tech.

A domain specific LLM trained on fantasy, science fiction, role playing etc, embedded into a hardware speaker / mask. Something you can prop up at the table, and have disconnected from the internets.