Game idea for someone with more time than me: an LLM powered "roguelike".
You fine-tune the LLM to spit out RPG skills based on the player's chosen class. (Dwarven princess = summon dwarf minions, has a rare bloodline magic, whatever). The classes could get _real_ weird, like a school bus character could summon school children minions and skill up to flying a la The Magic School Bus. (Or that could be a prestige class haha.)
A second model would match the RPG skills to a known/finite set of game mechanics supported by your game. This is ultimately a simple classifier - not really an LLM (though the skills on the input side are language, okay pedants). Determining the set of game mechanics to support shouldn't be too hard - how many different ways of doing damage and (de)buffs could there be? (Okay there can be a shit-ton.) Balancing the game could be done live via reinforcement learning. To constrain the search space, it'd follow the old Zelda format of limiting you to one room/screen until the enemy is defeated (or like Binding of Isaac for those too young.)
It'd have to use 2d/sprite based/placeholder graphics because I don't think the rigging/animation AI is currently good enough... but bad graphics hasn't stopped fun gameplay before. Imagine Charizard vs Death Star. Good luck with the DMCAs.
Another problem is that with simple graphics and literally infinitely variable enemies, there's no learning on the player's side so there's no real "skill" accumulation to improve on for the next run. There would be a lot of reading on the player's part if you decide to just show a stats-list for an enemy type. I suppose you could turn it into a zombie game where new enemies are slowly introduced in a horde of other enemies the player's familiar with.
