I’ve been thinking about how traditional game AI can be improved by generative models. One of the biggest problems with games like Civ is that the AI strategy is predictable - especially if you’ve played a few dozen hours.

LLMs with some decent harnesses could build up unpredictable - but internally consistent - strategies per each new game you play.

This is close to a proof of concept for those improvements.

I wonder how could you keep the LLM from going bonkers as the game progresses? I have a feeling it's possibly better to re-create the prompts after some time, and have the LLM work more like one of those "reasoning models" with the game as something it can interact with.

Otherwise you run into the risk of "TOTAL NUCLEAR FINANCIAL LEGAL DESTRUCTION" ;)

Absolutely! Max and I were huge Civ fans and always tried to make the game AI deviate from its programmed strategy. We also believe you can get some really interesting story arcs by adjusting parameters like temperature and how context is presented. Some of the things you'll notice in the game is we have a no-holds barred approach -- you can fully modify system prompts and adjust how the LLM interprets the state of the world.

As someone that plays those games pretty heavily: I’d rather not have LLMs take over game AI like that. If I want different gameplay I’d play online. We don’t need to bog down already heavy games with LLMs.

I can't help it, the first thought that came to mind was "Huh...talk about sheer senseless brute force." Why use a Large Language Model on something as clearly defined in scope as a game instead of a model designed and trained for the task/ruleset? Sure, there's the argument of not having to train that model, but OTOH, "decent harnesses" does some very heavy lifting there...

Definitely the case. That being said, I think it would be hard, at least in the immediate future, to translate the concept of difficulty to a universal LLM for a bespoke/specific game. I assume most game AIs are tuned by hand to feel fair for a given difficulty level...but if you just give an LLM some new game, explain the rules and what resources/abilities it has available to it, you're stuck with adding some addendum to the tune of "and you're meant to represent an entity of 'medium' difficulty." For very well established games, it might have a sense of how given actions might fall into a skill-level hierarchy, but not for anything new.

Fine tuned LLMs though with actual experience with the game, maybe?

From a player perspective, oftentimes the best AI systems are the most trivial ones. You can get really far with an agent that is allowed to cheat. It's a hell of a lot easier to build and troubleshoot a model that manipulates the amount of in-game resources received per unit time than it is to implement actual strategic intelligence.

even being able to scheme with / backstab leaders, and they would "understand" all that's happened (and acts accordingly) would be so fun

I'd love to see LLM based versions procedural tasks like "radiant quests" which are generally disappointing, though I've heard it discussed before and the real challenge is keeping it from going way off the rails.

The other challenge I think you'll run into in general is that there's a huge knee jerk reaction against any use of LLMs or other popular types of gen AI in games in places like Reddit or Bluesky.