Posted by vuciv 3 days ago
It would make single player game so much more alive, even railroaded one like red dead redemption because the npc could adapt more easily to what you just did, how you behave in game and so on.
Games are already demanding on gpu, running a very tiny local LLM and only when people interact with npc wouldn’t require so much power.
I’m sure there are issues which explain why it’s not already being used but the first one to do it will be remembered.
The biggest problem I faced at the time (during ChatGPT 3 era) was that, without a good context, LLMs are the most vanilla roleplayers you’ve ever seen. By themselves, LLMs are just not interesting enough for a player to choose to talk to in-game.
If you want them to be “interesting” to talk to, you must provide (or generate and keep track of): a backstory, chat history, the scene, NPC inventory, the NPC’s current emotional state, the weather, literally everything needs to be given to the model before it generates messages for the player.
At which point you’ve got a big task. You need a way to automatically get the relevant data to the model for the specific conversation you’re having. There might be tools to pick appropriate text documents from a db given a conversation topic, but I didn’t/don’t know how to make that work for games.
I’m sure there’s a way to accomplish this with more modern tools and models. (Maybe instead of providing all that data up front, you would now give the model tools to call to retrieve that data on-demand?) But that’s what made me give up in 2022.
I mean it looks like to me the next big step in gaming after 3D and yet this is being ignored.
But also, why couldn't you look at the code to find the addresses used for dialogue? If it's already disassembled I would think you could just look at the addresses used by the relevant functions/set a breakpoint etc.,?
While it's impossible for game developers to write code to cover every situation, AI could make general reactions possible.
It's surprising that really simple things like this haven't been tried yet (AFAIK!). Like, even if it's just the dialogue, it can still go a long way.
Old text adventures honestly did this heaps better than modern games do, but the reality is there was a more finite action space for them and it wasn't surprising when something wasn't catered for.
I’m only aware of experimentation in making more “difficult” NPC AI which was found less enjoyable for obvious reasons, so would be interested to see why similar but different attempts down another path also failed.
I would love to see a Zelda game implement LLM dialogue for random inconsequential dialogue or relating dialogue to the game context.
https://github.com/jmarshall23/Quake3LLM
jmarshall23 is a beast, with tons of interesting id tech-based projects.