I saw a clip somewhere saying Neuro SAMA was the top female streamer for the month or something, and while I don’t particularly care about stuff like that, I find AI super interesting.
So what would it take for AI to play this game solo, in a dungeon group, or in a raid group? If you just turned AI controlled characters loose with the prompt to complete X dungeon or raid on Y difficulty, would they eventually sort out what talents, enchants, gems, etc to use to optimize their situation and create a better chance of success?
I’m relatively certain we cannot perform the experiment due to the ToS, and this wouldn’t necessarily be human inputs beyond the initial prompts and providing stat weights and damage formulas, but I’m really curious what everyone’s take on the idea is.
Blizz worked with Google DeepMind before for AlphaStar. I’m sure Microsoft could provide a bespoke World of Warcraft server instance to train a community of AIs, if the fancy ever struck OpenAI or Microsoft’s own research team. Currently the hot sandbox for such projects tends to be Minecraft, actually.
Here are some cool papers around that idea, since you mentioned the topic of AI particularly interested you:
Voyager: An Open-Ended Embodied Agent with Large Language Models: https://arxiv.org/abs/2305.16291
Project Sid: Many-agent simulations toward AI civilization: https://arxiv.org/abs/2411.00114
I always had a deep fondness for NPC companions, too. I really dream someday I can actually experience Skyrim with a fully real-time open-ended interactive Lydia.
Is it actual AI that learns and adapts or just a series of scripts tailored to each dungeon? There’s a difference, and that’s probably one reason why they haven’t added follower options to all past dungeons yet.
Monsters in turn based RPGs don’t have “AI”, they just have a set of conditions. It doesn’t sample data and adapt to it the way true AI does.
We already have bots that can play the game better than a human so to answer your comment, yes an AI could do it.
Why aren’t these bots a bigger deal? Because they’re more expensive (both literally and from a processing perspective) to run for the job that bots are made to fill.
It also isn’t as deep as people have made it out to be.
The model needs parameters for success and failure, and likely for specific events given the complex state of mechanics in the game, at least from a viewpoint of what failure and success mean in the short term.
The biggest hurdle is giving them an environment to train in, since you’d ideally want the timescale changed.
Yes and it would easily rank as number 1 player in the world on pretty much all specs.
I mean Honorbuddy already had some of the top ranked players in the world and that wasn’t even able to account for things like phases, getting PI etc where an AI could learn to account and adapt to a fights phases and group comp.