I have an idea that we could use AI to host the IF. Recently, I have been studying the LangChain, which is a framework for using AI, and it provides a method for AI to understand documents. So I think maybe can write some opera. as an upload to AI, and let AI host the game?
This has been hashed out in several threads (including this one). I think the general consensus is that LLMs can’t really maintain the story or world consistency to really run a game. At least that’s my interpretation.
6 Likes
Agreed. It’s been tried countless times before, and it’s not worked. Nothing adds up. Plus that takes out the point of making it yourself. People like me, I don’t care too much for playing (that’s a lie, but the main point is I much prefer making my own).
It’s also just unethical. We’ve had too many conversations about this in the chats. And it takes up so many carbon emissions we’re heading for oblivion much faster than scientists predicted. So, yeah-
5 Likes
hmmm Actually, try recording the chat and go with a more detailed script, I’m thinking with the help of RAG, which is still essentially getting information from text. It’s similar to ai summarisation or something like that. and recently google claimed to have context free models, I wonder if that would be used to help?
I mean, sure, go ahead, but this has been tried so many times before. Don’t worry, I’m not stopping you, I’m just telling the truth.
7 Likes
I think we could use more intelligent parsers (especially for new IF players), I mean the algorithm that tries to find out what the player wants to do. [You could call it input AI.] But I don’t think this can be done with ChatGPT.
I don’t think it would work to let AI decide what the game world does in reaction to the input. [This could be called game logic AI.]
Edit: Sorry I wasn’t aware that this has been discussed excessively already.
For example read this post:
1 Like
I’ll just say some of my thoughts here.
I’m going to start by writing a script which should contain the plot of the game, the map of the game, and every item in the map, the more detailed it is the better, and then write a program through the framework of langchain to allow ai’s to interact with the player on this script, and the program should save model and user’s record, use to next play.
I’m going to try out this idea, but it’s going to be next month because I have an important exam this month.
AH that topic is too old, I don’t recover it.
From reading people’s experiences, my impression is many AI routines seem to include an element of random improvisation to make it seem more ‘alive’ in conversational gambits when it’s a bit confused to prevent it from always saying the same thing over and over - which does help in casual stream-of-consciousness conversation. However, this coupled with a short-term memory about what has been established previously in a longer interaction crops up inconsistencies over time, which usually comes across as the AI lying to the human interlocutor as the AI argues that it remembers what it said three nodes back as the truth, but doesn’t remember the other contradictory thing it said ten nodes ago.
AI doesn’t have a concept of “yes, and…” for improvisation and truly believes the most recent established facts in its short-term memory, so the behavior can be interpreted as if you’re speaking with a pathological liar or someone who is trying to gaslight you!
3 Likes
You make it sound almost human!
1 Like
I think that’s the goal with LLM AI!
Part of the “sounding human” part almost requires the AI to be given leeway to establish ‘opinions’ and possibly get stuff wrong sometimes, which might make it more realistic to converse with, but not as useful as a source of information since it can also develop bias.