Why not train "our" own user support agent?

[Boxx]

I was reading again the hot thread about the use of ChatGPT :wow: just after searching for some old advises among the 627,553 posts (to date) of this forum.... Then, I thought it may be not that complicated to install an AI-agent to support users at using Orbiter, instead of the traditional search engine of the website. It could even assist a user at troubleshooting ("have you checked your video tab? ... my what? ...)

I'm not thinking about ChatGPT, likely more GPT4All with advises from Claude AI or the AI community https://huggingface.co/spaces to develop an OrbiterForum-specialized LLM and train it from our 627,553 posts. I wish I could support such a task... it first depends on the enthusiasm (and other running tasks).

 

[Lucy]

This will be a divisive subject, because LLMs are already a divisive subject. I will confess I had a knee-jerk reaction to reading this post that probably doesn't come from a helpful place.

What you propose is technically possible, by which I mean we almost certainly possess the technology to create such a model, host it and run it - if not the experience of doing so already, perhaps. My misgivings with the concept are more conceptual in nature;

• What value does this add to the forum? Currently, we're a place of mostly human-generated content (there's the occasional post by a bot and I'm sure some folks have written posts using ChatGPT) and a resource mostly "by humans for humans". Adding "AI Features" into that feels like it runs against the goal of having a community of humans - admittedly a small community of active users at this point - working together to discuss and solve each other's problems. You can, for instance, see the impact AI has had on sites like stack overflow. I really don't want Orbiter Forum to turn into a place where bots talk to bots nor a place populated solely by the hallucinations of a large language model - we've worked quite hard to ensure that isn't the case, and as a result it often feels like logging into a piece of the "old internet" when I come back here. There are human beings talking to human beings here, and I would always prefer that to be our focus.
• Consent is essential - as our privacy policy (yeah, we have one) currently stands, we - that is, the Orbiter Forum site itself - don't specify whether or not we use the content posted into the forum as fodder for an LLM (but to be clear, we don't). That's mostly because the policy was written before those tools existed, but I would consider it important to communicate to our users that their posts here will be used to train an LLM if we were to begin to do that. I've no idea what the response to that would be like nor, indeed, where that would leave us with the posts that were all made before we changed the policy.
• Isn't this happening already? Content on the forum is already publicly available, most of it without logging in. If I check the server logs at any given moment I can see a handful of "indexer" bots, several of which identify them as some form of AI harvester. I did expend some effort trying to discourage them, initially, but we don't really have the resources to keep up with that particular arms race, so I suspect a reasonable amount of the posts here have already been used as training data for some existing LLM somewhere (possibly ChatGPT).

 

[Boxx]

I'm a bit sad to read that but it is quiet consistent and maybe I am naive (which is my best quality)