Imagine a game like “the sims” where you can adjust how autonomous the sims you control are. I could see Ai being used to control that.
Or having an elder scroll game were you just respond however you want and the npc adapts to it.
First of all, I’m going to replace AI with LLM, since that’s probably what you meant.
There are 2 distinct questions asked in this post:
- Why not use LLMs to provide different levels of automation? (Like, manual, medium, auto)
Answer: you don’t need LLMs for that. You can just code it in like any other feature. It’s not particularly hard, game developers know how to do it since they are used to programming automation for NPCs.
- Why not use LLMs to procedurally generate NPC dialogue?
Answer: games are primarily a form of art. NPC dialogues are written with a purpose. Different characters have different personalities. Some dialogues are meant to drive the plot. Other dialogues are meant to teach the player how to play. Others are meant to show the player things that they may have missed, or things that are interesting.
Procedural dialogues removes all the control from artists. They would all be generic npc n#473, with the personality of the LLM, maybe slightly varied if the developer writes a different prompt for each character.
Procedural dialogues would have the same issues as procedural world generation or photorealistic graphics, it would just not be interesting.
There is a practically infinite amount of Minecraft worlds, yet they all feel the same way. The thing that differentiates a Minecraft world from another is that which the player has built. The only part of the world that wasn’t procedurally generated.
There is a great amount of photorealistic games. And they all look very similar. You may only distinguish one from another by looking at their handcrafted worlds or their handcrafted characters. But not by staring at a wall. You can stare at a wall in non-photoreslistic games and know what game it is.
So if you put procedurally generated dialogues, no one will read them, since you’ll be bored by the time you read the same thing being said by 5 different NPCs from 5 different games.
Have you ever tried to run an LLM locally? It makes CPU usage go way up, uses a lot of RAM, etc. It would tank game performance and/or require beefier PCs.
Games already have had AI for a long time, but the kind of AI you’re talking about would be far more computationally expensive than what they currently use.
Ha. You are thinking of running games locally? Cute.
AI in games (using code for entities to make non-player decisions) is about being good enough, cheap enough. It’s just like how games determine their physics. The existence of large scale “black box” AI like OpenAI does not reflect what’s good or cheap. It can’t play chess. You think it’s going to understand The Sims and make reasonable choices in that system?
They’ve already created well tuned system to give your Sims in obtaining their needs. It leads to you having to manage the chaos, and that’s what the fun is. To better hone that is to have the AI play the game for you. And even that, if efficiency of play is the goal, is better done by TASbot and machine learning.
That generic black box style of AI like popular LLMS is like creating a hammer. Now everyone is treating every problem like nails. AI decisions making in games is like washing windows; don’t use a hammer.
The problem is that “AI” is a poorly defined, very vague, and widely used term. Most people here have assumed you meant LLMs because everyone pitches those as ways to solve everything. “Oh, irer up an agent, give it instructions, and let it make requests that are context dependent”. Then, like everyone says here, that usually turns into people testing boundaries and breaking your game. So that makes it both “not good enough” and “not cheap enough”.
Now, look at AI with the term “machine learning” in mind and it’s different. Games like ARC Raiders use machine learning to teach NPCs movement behavior, and to train AI voices like Siri so they can’t add things without further paying people. They think that up-front investment is worthwhile. But those are both far cries from “uploading it to Claude or ChatGPT and see what happens”. Especially when you would have to teach that black box AI your system anyway, for it to use it. And you’re already doing that with current “good enough, cheap enough” bespoke methods, for much cheaper, and they’re good enough.
You would have to design the game around an LLM, not just drop one into existing games.
It might be cute for the guards in Skyrim to have unique dialogue, until one of them denies the Holocaust or says feminism is cancer.
I tried this as a side project and it’s such a pain in the ass to get the bots to actually behave like they’re in a world and not be overly eager losers.
You have to do so much prompting to get them to behave and, as others who have to work with these full time know, prompting can only go so far.
They’re not as autonomous and general usage as companies want to make you think they are.
Ah, but then you get to stab the Nazi in the head.
There actually is a semi-working system for Skyrim/Fallout https://art-from-the-machine.github.io/Mantella/
Also not all LLMs are Nazi machines, I almost exclusively use abilerated models and I’ve never once had it go on a nazi tirade.
Though I mostly use it for Linux/code or random home assistant projects, not as a conversation.
I have heard of some people experimenting with it, e.g in a stardew valley mod to allow you to have actual conversations with the characters.
While that sounds like fun. It also sounds like something that is fun for 10 minutes.
Agreed, I don’t think the NPCs are the best thing about the game. What I like best about SDV is that it’s essentially an industrial collectathon game.
If we could get complex conversations in games (e.g an adventure game where you’re not limited to 3 conversation options) I think it would be awesome. Might mean we waste more time playing the game, though.
i have watched some videos where its used on skyrim npc, it seems to work surprisingly well.
What value would it add to the game?
- LLMs are computationally expensive
- Replacing voice actors with AI means making dialogue worse
- Replacing writers with AI means making the story worse
At the end of the day AI is mostly a marketing term for LLMs and LLMs just aren’t that useful in most games, they just average out a dataset to autocomplete a response, that autocompletion is worse than what a human would have written.
We saw with procedurally generated worlds that it takes a lot of effort to prune what is generated to make the game interesting.
There are particular subgenres of games and applications where LLMs might be useful though.
I am a fan of using LLMs specifically to imitate the VAs on demand to pronounce character names. They’re generally good enough that a single word can blend in, and you have a couple minutes during the opening cutscene to run the computation. Just having all of the characters never say the custom player name and instead address them in the second person or with a title is a bit jarring
I think the Where Winds Meet tried this, right? The NPCs ended up saying anachronistic things and making travel itineraries for Beijing or something.
Do they? I’ve talked to several NPCs, never happened to me. At most, they get completely confused on what you are saying. Eg, one kid thought he was rich enough to buy a house. Trying to tell him he’s not and he thinks I took his money (and started crying, but also became friends?). In another a guqin player wondered if anyone could tell how sad she was from her playing. Instead, we’re keeping secrets? (No idea how that came about).
And before anyone points out, I dropped the game due to quests requiring MC drinking alcohol (can’t stand games like that. Just a me issue). Sad because I loved the everything else too :(
one kid thought he was rich enough to buy a house. Trying to tell him he’s not and he thinks I took his money (and started crying, but also became friends?).
I don’t know about confused, have you ever talked to a toddler?
I haven’t actually played it (wont play any game that used or uses LLM software), so I can only tell you what I’ve read.
Shame, it looked interesting
Have you ever talked with an AI? It sucks.
I’ve talked to them often. So I don’t bore my family with my wild ideas lol
Those wild ideas would be good for someone
We do use AI in videogames, and have for multiple decades (with varying levels of sophistication).
Indeed! Seems like people have forgotten that AI is not just LLMs.
People have tried this a bit and it doesn’t work well. Remember that most games have some sort of plot which needs to move forward without deviating too far and this is not easy to manage with AI. AI systems are predictive text tuned up, so they tend to wander in the conversation and this can be disastrous for something like a video game.
The world is there to support the illusion but also to direct the player to game material. An AI agent going off on a tengent about some random thing that kind of fits the world could lead to users running around wasting their time and being frustrated.
Add to that the risk of the AI system stepping into awful places like reproducing Nazi ideology and it is a nightmare for developeds. Imagine getting your game rated when it can randomly start telling your character not to worry about saving those people over there because their skin tone is darker and that makes them less than human.
Now as a tool for building scripts quickly? Maybe, but it does produce slop now and if that will change I cannot predict when. Maybe it could be used as part of the process but I think it is so toxic now I would not bet on it. I also think it should be labeled as the use of AI comes with moral issues around the environmental impact and theft of content from other people. If a game has AI generated content I won’t be playing it, and I am not alone. Just the push back from audiences could be enough to discourage the use of AI systems.
Now on the other hand using a neural network design for making character behaviours more believable, for example using a series of needs and having the algorithm decide what to do next and so on, that could be cool, but we have that already and it isn’t considered AI.
Thank you that explains alot.
Unfortunately now all I can’t think of is “great sir knight the queen has been captured, and much like hitler she has done nothing wrong!”
Player “I will save the queen and… wait hang on what was that about hitler?”
Npc: I’m just saying if you look at the geo-political climate of the 1930s-
Player: I’m just going to find the dragon thank you.
It would need to be cloud-based, or otherwise require a lot of RAM
Because fuck AI, that’s why.
Yah but OP is talking about AI in gaming, not generative AI that makes propaganda out of stolen art or turns your family photos into porn, there’s a big difference. You’ve been playing with AI in gaming for years, and probably have complained about it because it always sucks. Enemies walk predictable patterns, they see you kill twenty of their friends and then resume their patrol and say “it must have been the wind.”
Lets not mindlessly rage at terms without understanding the context, then you’re just becoming MAGA.
There is a game called Whispers from the Star which uses an LLM to run the script. It’s pretty much a fancy choose your own adventure book. It’s pretty shit.
Because the kind of Genrative AIs which would be worth puting in a game (smaller ones) have two drawbacks for the hype train :
-you can’t promise an AGI which would justify the govt putting mbillions in your company in order to stay “competitive”
You can’t create a feedback loop of finance with nvidia and the like because your company wouldn’t need such computational power then.
I’m pretty sure no one is going to think an in-game npc is real
Do it: modify Minecraft so a villager gives investment advice. No one could be dumb enough to expect that to be real, right?
No one could be dumb enough to expect that to be real, right?
Oh, my sweet, summer child…
There might have been a missunderstanding,
What i meant is that as an ai company, to service so many clients at once, you would have to downscale your ai models quite a lot.
In doing so, you limit your claims that you can “make an AGI, i swear bro just one more server farm”. This means that the government/investors are less likely to jump on the hype train.
All the while, downscaled model require way less trainign time and data scraping meaning you won’t get to buy all of nvidia for them to buy all of you for your market value to explode for your money to go stonks.
As far as i’m aware, this is the reason why you don’t see AIs as npc (at least yet, maybe when we get a little bit more reasonable, we can try to do it inteligently)
Please pardon me if i missunderstood your comment/post
You can not cage llm. They will break out at some point, it’s proven again and again.
(it can have its uses, but this idea will run rampant with time)
