Questions about LLMs in Group Chats(vineeth.io)
vineeth.io
Questions about LLMs in Group Chats
https://vineeth.io/posts/llm-groupchats
20 comments
I've been contemplating these exact same thoughts and ideas, and indeed have been very surprised how little exploration there seems to be around these nuances!
given the mechanics of language models, i think it's really interesting to consider them in group settings. how do you create an environment where they can "decide" to respond? how do they make that decision?
this is something that's talked a lot about in education -- how to foster a productive group discussion: https://www.teacher.org/blog/what-is-the-harkness-discussion...
this is something that's talked a lot about in education -- how to foster a productive group discussion: https://www.teacher.org/blog/what-is-the-harkness-discussion...
This is very interesting and thought provoking. On its face it sounds so simple - just put the bots in a room together and let them go at it. But of course it's much more complicated than that, at least if you want good and/or interesting results.
I did a small test a year and a half ago (very basic) and it was already funny xD
https://github.com/realrasengan/chatgpt-groupchat-test
https://github.com/realrasengan/chatgpt-groupchat-test
I'm interested. Could you use a LLM function call to decide whether or not to respond instead of randomness so it feels more intelligent?
Also, I have no idea what the use case for this would be but making it work sounds cool and kinda fruitful.
Also, I have no idea what the use case for this would be but making it work sounds cool and kinda fruitful.
> Could you use a LLM function call to decide whether or not to respond
I've implemented this for a toy project before and it worked surprisingly well, yes! It can take some creative prompting to have the model understand which messages are actually directed at it, though, which I guess makes sense since every User message is _supposed_ to be?
I've implemented this for a toy project before and it worked surprisingly well, yes! It can take some creative prompting to have the model understand which messages are actually directed at it, though, which I guess makes sense since every User message is _supposed_ to be?
Curious what kind of things you had the model look for? Did they have any kind of goal or was it more a "vibe" thing.
Also would it only respond if a message was directed at it or could they proactively decide to join a conversation?
Also would it only respond if a message was directed at it or could they proactively decide to join a conversation?
What model did you use for this?
> Could you use a LLM function call to decide whether or not to respond instead of randomness so it feels more intelligent?
Yes, and once you collect enough samples from the LLM you could train a SLM to do it faster and cheaper.
Yes, and once you collect enough samples from the LLM you could train a SLM to do it faster and cheaper.
What does this achieve?
Main thought is just that there are tons of situations that involve more than 2 people being in a discussion. some the of links I mentioned in the post get into this with the "facilitator" model.
Moderation/mediation is probably the most obvious thing "usecase." I'm more interested in just pushing the limits of what the language models can do and see what kind of things they'll do.
There's a corner of the internet interested in these models as a form of "alien intelligence." Don't immediately know what the application would be, but the results are super interesting
Moderation/mediation is probably the most obvious thing "usecase." I'm more interested in just pushing the limits of what the language models can do and see what kind of things they'll do.
There's a corner of the internet interested in these models as a form of "alien intelligence." Don't immediately know what the application would be, but the results are super interesting
Fun and satisfaction of curiosity for me.
I think it's inevitable that group chats will involve LLM agents in the future.
For?
Ideally? Automatically answering support questions in open source discord channels, automatically taking minutes in business slack conversations.
Cynically? Your group chat is hosted by a tech company. That tech company wants to tell investors they've got an LLM, and show user counts going up. And the fastest way to boost user counts on a failing product is to make it an always-on feature of a successful product.
Cynically? Your group chat is hosted by a tech company. That tech company wants to tell investors they've got an LLM, and show user counts going up. And the fastest way to boost user counts on a failing product is to make it an always-on feature of a successful product.
I don't think your first usecases are applicable to what people normally call "group chats"
For all the same reasons you talk to humans...
To find out what’s going on in their lives and to arrange to meet up socially?
Is that the only thing you talk to your friends about?
I mean yeah literally you could talk about your plans [i.e. by yourself or with the other real people] or what's going on in your life and have the bot chime in with thoughts too. You could talk about food, philosophy, politics, travel, complain about stuff, get advice for relationships or work, gossip. All in a group chat setting.
It's not hard to imagine a bot that's able to contribute here and there in a meaningful way in any of these subjects, and other unlimited subjects. But perhaps you weren't open to it in the first place so I'm not going to try to convince you further if you don't buy it; If you don't want to really give it consideration then nobody is going to convince you otherwise.
I mean yeah literally you could talk about your plans [i.e. by yourself or with the other real people] or what's going on in your life and have the bot chime in with thoughts too. You could talk about food, philosophy, politics, travel, complain about stuff, get advice for relationships or work, gossip. All in a group chat setting.
It's not hard to imagine a bot that's able to contribute here and there in a meaningful way in any of these subjects, and other unlimited subjects. But perhaps you weren't open to it in the first place so I'm not going to try to convince you further if you don't buy it; If you don't want to really give it consideration then nobody is going to convince you otherwise.