Having used Claude to help build Sharpee, I can see it had already added some NLP type logic to my parser. I questioned and researched what it added and determined, “this is all normal NLP logic that’s been around for decades” and left it in the code.
You don’t need an LLM to help with text input. There are plenty of other historically proven techniques to do so.
In my opinion, using an LLM enabled parser is a potential enhancement only in sections of the game involving NPC dialogue. What I noticed when introducing my girlfriend to the Witness for the first time is, she gets to the back room, the part where you interview Mr. Linder. What do you think she’s going to type in? Naturally, she does what a detective would do, she asks him questions. Of course I had to explain to her it doesn’t work like that.
But now, with this technology, if I’m making a game, I don’t see why the player shouldn’t be able to ask that character questions.
Same thing when I’m playing Planetfall. I meet that funny robot for the first time. I want to talk to that robot damn it!
Now I certainly do feel if I’m talking the time to make something and someone plays it. I don’t want them to see AI generated slop. I want them to see my human written slop! So in my view, the goal should be to restrict the AI to outputting specific lines of human-written dialogue. The language-understanding capability is only there to determine if your input “probably” leads to it. If the player types in something completely wild, it should lead to an improvised response, but with a shortened maximum length, like one sentence, to mitigate the obnoxiousness. Also, the AI should be instructed to not answer any questions about things it doesn’t know about based on the context you provided. When there’s hallucinations, because there inevitably will be, maybe there’s a creative way to contain them. You’re a schizophrenic, you’re seeing things.
Number 2. The AI should be used for persuade checks. Think of dialogue the trees in the old Fallout games. You have to bluff your way past the guard. In Fallout, you just selected a dialogue option from a menu like in a CYOA, and then it rolled your charisma stat + speech skill to determine if your bluff was successful. Well, I think, F— that. Write your own damn dialogue. You see the thing about interactive fiction is, it’s not that I want AI to do more of the work, I want you the reader to do more of the work. Why should the author have to write everything? So, if you wrote something clever enough. The AI will evaluate you as successful and you get to pass, if not, get ready for fisticuffs. I think this is an entirely appropriate use of LLMs in a role playing game that uses dice rolls during game play, especially when you consider the probabilistic nature of the way they generate words.
You are not wrong, but i think there are simpler ways to do this. If you’re going to emit pre-written responses, the old way would be to have some kind of “scanning” or “keyword” parser in addition to the usual command parser. Think “Eliza”.
Of course, this would also not be perfect in detecting meaning, but also an LLM can often get the wrong idea.
That would get old fast.