some people like to interact with another human for technical questions or significant needs. simplifying things is very good but ordinary people dont always accept it and cant communicate well with such a ai platform.
2 Likes
The weird feedback loop is when the AI is answering something based on other previous answers, and this becomes a new “answer”, that is then propagated further for the AI to use next time.
So it starts to parrot itself, and can get stuck in a loop if there isn’t enough fresh/good data for it to parrot and propagate.
1 Like
I have not! I was really curious about it though. I want to try it out for a few different reasons. The data from airchat is either a goldmine for multi-turn speaking data, or is utterly useless for that kind of information.
One of the interesting topics in discourse analysis (a subfield of linguistics) is about turn-taking in groups. How do people know when to speak or interject? How long do people wait until someone speaks? These are questions that are pretty tough to answer, because there’s seldom a definitive answer. This makes it hard(er) for AIs to handle multi-speaker discourse. Airchat’s unique setup should, in theory, make this kind of data easier to gather, and easier for an AI to make sense of. Or, it might turn into a new mode of interaction altogether, and this new style of communication would be the perfect bridge between regular spoken discourse and internet discourse.
Yeah, it’s kinda difficult to ask about something technical like error handling when right under your post is your aunt Karen sharing a Q-Anon photo talking about how Obama is going to come vaccinate your kids, isn’t it? 
Just remember that everyone has different abilities to communicate, and there’s going to be a lot of that here as more folks come in. Forums are just as good an avenue for ordinary people as experts. Everyone starts somewhere!
If it wasn’t for this old forum for this game editor called RPGtoolkit, I wouldn’t be here. I was either 12 or 15 when I first joined that forum. That place fostered my desire to develop things quite well. There’s going to be plenty of situations like this too I bet!
2 Likes
Hey! I’m new to this community and hope I’m gonna find friendly people here!
I believe a full-AI forum would be a bad idea, BUT an in-between would be excellent
=> AI bot adding informations (like the community notes on Twitter), indications, review, or globally a synthesis of all answer for a topic under the main post would be great!
There’s 1000 things to do, but adding complementary sources or counter-sources in an automatic way would be great for everyone, it could also help with code and suggestions of answer. Maybe even for multilingual people as me, allowing user’s to write in their native language, and translating every post in every main language to definitively break language barriers!
Maybe with or w/o a bot help to navigate the forum, and providing help with a RAG on data available there!
3 Likes
Welcome to the community!
Nice ideas. It would be great if there was an option to include sources from the Documentation and Forum at the bottom of your post by just clicking a button. It would make everyone’s posts more trustworthy.
1 Like
While this feature is not something we can build, once Discourse’s AI integration starts supporting Google’s API, I think that’s when we will start seeing a bot that can answer questions using Discourse AI(advanced).
1 Like
Welcome in!
I was thinking about this myself actually!
Google really shines here when it comes to translation. It’s one of the few things I trust Gemini to do better than other language models, and they have shown evidence for even understanding rare languages they have not been trained on. Google translate is practically in everybody’s toolbox already.
I would love to build something like this!
Also,
I think I mentioned something similar in another topic here, where Gemini could find quotes or posts on this forum for citations or something similar. It helps compliment what humans do, not replace the human action itself.
Speaking of, in said post I mentioned, I asked “How good is Gemini Pro at ‘needle in a haystack’ type testing?”
Well, I’ve found the answer to that question in their technical paper, which can be found here:
According to their technical paper, the model achieves a 99.7% accuracy rate amongst needle in a haystack questions across its million-token context length across all modalities. It can even extend up to 10M tokens, according to the paper.
This means that these suggestions and requests aren’t just on the table, it’s realistically achievable.
If anyone is curious where I found the paper, I got it from Google Deepmind’s Gemini page here:
1 Like
Sorry for being a little late on the response here, but as @Curt_Kennedy also mentioned this is one:
But I was thinking more about human interactions like this:
This traffic can be create both positive and negative feedback loops. An example of a positive one could be people finding a solution to a problem they’re having, and decides to post about another issue that is subsequently solved.
But imagine if someone with billing or account related issues decides to post their complaints here, which brings in more people having the same issue that we cannot help with, because it requires access to a lot of personal information that shouldn’t be posted on a public forum.
1 Like