cross-posted from: https://piefed.ca/c/[email protected]/p/1004331/ai-driving-game-players-are-burning-through-so-many-tokens-its-developer-had-to-take-out-a
The developer behind Teach My Little Sister How to Drive, a driving instruction sim where you instruct a generative-AI-powered, dynamic NPC, has been spending $1,000 dollars a day to keep the demo working.
NPCs in games may be one of the very few actual good uses for AI. But not using the cloud obviously.
Theoretically.
In practice, 95% of consumer GPUs are too busy and too small to run an LLM good enough for dialogue, or a decision model smart enough to be warranted over hardcoded logic.
CPUs are too slow. NPUs aren’t universal and are, unfortunately, all but useless with modern software frameworks.
I am very much for LLM NPCs; there are some fantastic mods for such things.
But in shipping games? That can’t be jank?
It is not time. And I would never buy a game with a hard cloud dependency for SP. We will have to get past the RAMpocalypse, at the very least.
The graphics processor is busy doing graphics processing while playing a game. All supporting features of game development: logic, physics, npcs, only run on cpus.
No. No, they’re not.
I thought we all learned this with the Batman Arkham games. “Real” AI would have made those games unplayable Souls-like nightmares as the enemies learn every single trick or something. We only got the dark knight power fantasy because of the very basic state machines they used. Like I teach them being praised at the time.
I think that it would be highly dependant on how the AI model is trained and what genre of game it is implemented into. For example, games that more heavily rely on dialogue or social interaction.
Take Danganronpa, which is a series if linear narrative games about students trapped in a school forced to kill each other (and get way with it) to escape. You as the player have to investigate each murder and arrive at who the culprit was to win the game. They are fun on your first playthrough, but on any subsequent playthrough you already know the solutions. Like playing Clue but the game only ever has one outcome: Mr. Green, in the Library, with the Lead Pipe. You wouldn’t even have to open the game box and you already know. Absolutely zero replayability.
With an AI in such a game, you could have a new gamemode, like a New Game+, which players can play after the hand-crafted story. This is where I think something like a very carefully trained LLM could be useful, because it could create these kinds of situations at random, making them unpredictable and therefore adding replay value. Now, I don’t think AI is currently sufficiently advanced for this use, but if it continues to develop I could see that niche making good use of it.
Nah
First of all, video games, including detective adventure games with randomised riddles to solve and different endings, have been done for quite some time, long before llms. See e.g. https://en.wikipedia.org/wiki/Blade_Runner_(1997_video_game) . You don’t need a language model for that. Just be passionate about the project.
Second, the whole idea of a video game like that is to tell a story. The last thing I would want is randomised bullshit to click through, wondering all the time if what I’m reading is even still within the boundaries of what the developers wanted or if the llm was already going off the rails. And that’s not even considering the shale feeling of consuming not some created, curated story but merely the condensed diarrhea of stolen work.
No thanks.
A locally run model with quest components and npc dialog as its training data could be tuned in such a way to make radiant quests more unique and engaging. The scenario you’re putting forward assumes a frame around a general purpose LLM, which would be a very bad idea (as the OP article shows.)