There is probably some limited utility for using ML in games, lifelike multiplayer bots trained on behaviour of different types of players.
I think in most scenarios good gameplay comes from iteration which comes from tweak-ability, you see players get stuck in area x so you do a minor or major adjustment to area layout, enemy composition, or give the players more tools. If something is controlled by an LLM it becomes unpredictable and hard to reason about for both player and designer, and consequently also hard to adjust.
I think it is possible there is some role for LLM's in games, like how Left 4 Dead had an AI (this is classic AI not ML) director who would decide how to structure enemy waves and for example send a boss after a player who strays too far from the rest. But even there rules based AI is probably easier to adjust for the designer.
Using machine learning to make enemies smarter has been thought about and talked about for decades, but I don't see it being used either for similar reasons. Training a ML algorithm takes resources which are in short supply and time which could take longer than for the player to finish the game (if you want smarter enemies play against real human players). There's also the fact that the game has to run on the minimum specification computer, which means different players will have differently skilled enemies and their experience of the game will suffer.
There have been some games in the past which used some form of machine learning as a core gameplay feature, but they have been pet breeding simulators.
The purpose of AI in games isn't making something thats impossible to beat, its making something that is ALWAYS possible to beat while remaining fun and engaging.
You could tune this to handicap the LLM such that the player wins 80% of the time or so and it should be more fun (except for the competitive aspect).
I haven't been able to use it for more than events since the tokens per second aren't high enough - I would like to use the models to e.g. generate actions for characters based on their history of interactions with other characters, but that ends up taking so much time that the player never gets any events.
I also did a demo of an LLM-powered NPC, I think the issue here is that you need more than just the LLM in e.g. a 3D RPG where you can actually go up and talk to the character. You can generate the text, but from there you also need to generate speech, from that you need to dynamically generate realistic facial and body animations that match up with it - this means you need 3 or 4 models chained together in a pipeline.
[1] https://store.steampowered.com/app/2240920/Vaudeville/ [2] https://art-from-the-machine.github.io/Mantella/
There have been a number of games that have tried AI conversations, but you're basically trying to sell a Chat GPT prompt, and many players are aware they don't have to pay you to get a similar result on their own.