Somewhat is key. You can try to guide it in a direction, but that’s it. Also, as a player, you can never be sure if the dialogue is meaningful or not. Does it reveal something about the plot? Is it a key information about the character? Is it just hallucinated gibberish to fill the space?
Besides, LLMs struggle with retaining contextual information for long and they’re pretty dang resource hungry. Expect a game with LLM-driven dialogue to reserve several gigs of VRAM and a fair chunk of GPU processing power solely for that.
And then you still get characters who hallucinate plot points or suddenly speak gibberish.
You can run checks and fence it in with traditional software, you can train it more narrowly…
I haven’t seen anything that suggests AI hallucinations are actually a solvable problem, because they stem from the fact that these models don’t actually think, or know anything.
They’re only useful when their output is vetted before use, because training a model that gets things 100% right 100% of the time, is like capturing lightning in a bottle.
It’s the 90/90 problem. Except with AI it’s looking more and more like a 90/99.99999999 problem.
You can control what it spits out, though. They already do somewhat.
Edit: Gonna go out on a limb and assume most of you haven’t actually played any of the projects currently doing this. Or mess with chatbots at all.
Somewhat is key. You can try to guide it in a direction, but that’s it. Also, as a player, you can never be sure if the dialogue is meaningful or not. Does it reveal something about the plot? Is it a key information about the character? Is it just hallucinated gibberish to fill the space?
Besides, LLMs struggle with retaining contextual information for long and they’re pretty dang resource hungry. Expect a game with LLM-driven dialogue to reserve several gigs of VRAM and a fair chunk of GPU processing power solely for that.
And then you still get characters who hallucinate plot points or suddenly speak gibberish.
You really can’t.
You can run checks and fence it in with traditional software, you can train it more narrowly…
I haven’t seen anything that suggests AI hallucinations are actually a solvable problem, because they stem from the fact that these models don’t actually think, or know anything.
They’re only useful when their output is vetted before use, because training a model that gets things 100% right 100% of the time, is like capturing lightning in a bottle.
It’s the 90/90 problem. Except with AI it’s looking more and more like a 90/99.99999999 problem.