Hey folks, so I was talking with a friend of mine over Discord about what kinda games could be made with a program like ChatGPT, and it got us into brainstorming a game idea.

The game idea had a bunch of inspirations like Kenshi, Cyberpunk 2077, the 2012 Dredd movie and more. It was very basically an open world city set in a dirty, dusty, poverty-stricken (very non-glamorous) cyberpunk dystopia ruled by megacorporations where you start the game as a homeless vagrant and the gimmick/core mechanic of the game was that all NPCs were generated with a set of perameters like a job title, personality traits, rank in a faction etc… that will dictate their interractions with the player through AI (though higher ranked ones won’t be found outside of an armoured limo when out on the streets) and here’s the kicker, every mission, every quest, every interaction with an NPC is with ChatGPT or an AI like it, and, much like in some chatGPT mods in games like Bannerlord and Fallout 4, world events, other NPCs etc… can be spawned by what the NPC is talking about and you can work or even scheme your way up the ranks in world factions. If you ask an NPC to do something, and you outrank them or they owe you, or you’re threatening/blackmailing them, they actually do it like telling a faction underling you’re working with to lure a rival out so you can snipe them, thus opening up a spot to move up in the hierarchy and you might have some favors saved up with other higher ups to make sure you fill that new job opening etc… the thing is, everything the player said to every NPC in the game, was all typed out themselves, there is no charisma stat to put points into, charisma is how persuasive the player can be.

We kinda went wild with it. But the thing is, from what we’ve seen with AI mods in existing games, it no longer felt far fetched at all. It all felt like it was entirely possible in the here & now, all it would take is an engine designed to react to and work with the AI, being able to generate new events on the fly like spawning a group of gangsters at a certain location because an NPC said they were there and the AI itself being heavily trained in the context of the game world, what perameters to play NPCs with (like having 5000 personality traits and having the AI pick 5 to build a character around) etc… allowing players to have full conversations with random NPCs, get quests, build relationships and more.

So, as you can see, the capabilities in the use of AI, especially in Role-playing games, seems to be near infinite, but as game devs, what do you think? Do you think AI will never find a place within gaming? Do you think it’s inevitable? Do you think it’s impact will be cataclysmic or usher in a new age of gaming? Or do you view it as just another tool? One thing seems certain though, there’s no way the industry will just refuse to use it at all.

  • lurkingllama@discuss.tchncs.de
    link
    fedilink
    English
    arrow-up
    4
    ·
    1 year ago

    Some thoughts:

    Use of AI techniques in game development

    • Some AI techniques are already being used in the development pipelines of some studios. For example, image upscaling has been used to scale up textures in some remasters of older titles. But these approaches are tools that speed up the development process, not automatic solutions, and they still require being polished by an artists. See also this YouTube video on texture upscaling by Tommy Thompson.
    • Similarly, image upscaling is also already being used as a post processing step in rendering pipelines. nVidia’s DLSS allows rendering a game at a lower resolution (and thus at a higher frame rate) and then using a NN to scale the frame to the target size. YouTube video on DLSS, YouTube video on the newer DLSS 3, both by Tommy Thompson.
    • As far as I know, neither LLMs like ChatGPT nor text-to-image models like stable diffusion have been widely adopted so far. Personal opinion: I don’t think we will see an adoption of these techniques in the near future, for a variety of reasons. Without a lot of tinkering with their prompts, they tend to produce output that, while certainly not bad, is fairly generic and recognizable in style. They are also trained on copyrighted material and sometimes include substantial portions of a training sample in their output, which is a legal liability. Finally, you give up a fair bit of artistic control over the end result.

    Your game idea

    I don’t think this is viable at this point. The crucial thing to understand about LLMs like ChatGPT is that they don’t have any actual understanding of the world. The only thing they learn is to predict the most likely next word of a sentence, given the context of the conversation so far. So you could probably use them for some background character that only ever talks but never takes any actions (say, a drunk patron at a bar). But there is no way for an LLM to actually affect your simulation, so you’d need to somehow parse the conversation into a set of events in the game world, which does not seem feasible.