The problem here with highly capable agents is that it's not predictable and controllable in the point of designer's view. That might work for certain types of games, but many cases designers want to have a certain degree of control on their games.
You can control LLMs to any arbitrary degree of specificity by a mix of retraining and changing the decoding strategies. They can be as predictable as needed, I think the bigger issue is more like how do you write stories when the possibilities get so much larger.
Yes, less predictability is a part of the problems from the unconstrained search space. I think technically there is a room for improvement, but this usually needs ML expertise, which most of game designers and engineers do not have at this moment.
Figuring out what wont work is how you figure out what do work. All his points are good, they are things you would have to work around in some way.
> Then we'll make new kinds of games where the unpredictability of the NPCs is a core mechanic.
This is impossible, you need the NPC to be predictable on some level to make a fun game. Even unpredictable NPC needs to have a predictable personality on some level, total randomness isn't fun. Like, imagine a terrain generator that just randomizes terrain on each tile, that wont be fun at all, that is what a basic random personality would be like.
Think of a human opponent, they are very predictable, just looking at a human player and what he does and I can predict what he will do in the future. Not perfectly, but players aren't that random. To make an AI that feels good it has to be very predictable.
The main problem with "smart" bots is that they have so far always been way less predictable than humans, they get a strange edge cases and bugs where they start to act very dumb and strangely, that feels like a bug to the player and isn't fun. Or their smartness makes them do the same thing every time making them even more predictable than basic scripting, either way they are worse than basic scripting.
Getting over these issues is a really hard problem, LLMs hasn't helped solve that so far.
> Even unpredictable NPC needs to have a predictable personality on some level, total randomness isn't fun
tongue in cheek counterpoint: Rimworld players love Random Randy :P
I think it really depends on the game though, but you're right 100% random in an RPG could be really annoying.
Right now I'm into games like Factorio and Captain of Industry and they've both recently had blog posts about how they do terrain generation and CoI stuck out because you can manually plop features like mountains and then it procedurally generates the mountain range[1].
There's been a lot of games recently that seem to be doing procedural land generation, is there not a way this can be applied to AI personalities as well or is there no overlap between them? It kind of feels like procedurally generated personalities should be do-able but it sounds like there's something more going on that complicates that?
> tongue in cheek counterpoint: Rimworld players love Random Randy :P
Even randy random isn't entirely random, people love it since it sends you big threats, so it is coded to ensure it throws you big threats. If it randomly didn't send big waves people wouldn't like it as much.
"If Randy has not fired a major threat after 13 days, the next Randy fired event becomes a major threat."
> There's been a lot of games recently that seem to be doing procedural land generation, is there not a way this can be applied to AI personalities as well or is there no overlap between them?
I'm certain that is possible, but we don't have nearly as much intuitive understanding how to generate full fledged personalities hooked into an LLM that changes how the character acts and his motivations etc that will actually work well when put in a world and interacting with other NPC's in that world.
Terrain is just really easy to generate well enough, almost everything else is way harder.
Just imagine the other leaders in Civilization/The Total War games being given a prompt like “You are Abraham Lincoln. The following has happened: … What would you do within these constraints: …. “
Diplomacy in games have always been terrible because it just comes down to preprogrammed numbers. Humans might be the same at the very base level but there’s no way we could code that all out.
Now you could truly make those leaders act like they “should.”
A pretty material portion of the population actually doesn't have much of any imagination, and these people, especially when they have had some success and developed "expertise" will always assume if they can't imagine it then it's impossible, or if they don't know how to do it then it can't be done.
Surely you've met some of these pessimists before.
Nobody actually wants to read LLM-generated soup. That's just a waste of everyone's time.
If you're making a highly interactive, dynamic game, you don't even need detailed language for NPC interaction. You may as well use simple templates or even symbols.
Easier said than done. I assume you have no prior experience on professional game development? Many game designers tried to tame chaos as a game mechanic in the name of "emergent gameplay" and only a few of them survived through numerous iterations in an extremely limited format. I would recommend you to do your own research before making such a bold statement. It is not that people cannot come up with the same idea; many cases they tried it and there's a good reason not to do that.
