[Pr0m3th3us]

That’s not exactly what I said then. I said 2-3 threads a day.

I’m aware, my point is it got out of control. He was spaming.

---

I guess I will add, if Stanley posts threads again, even 3 a day I would like a feature to be able to not see them on my feed. Or anywhere…

[Proof]

Maybe AI can reply with a fact check on posts in certain categories, but further than that I don't see it.

Could be useful in any of the dev/programming forums remaining.

[Djangoo]

Djangoo would strongly recommend against creating fake human members powered by AI.

The issue isn't AI-generated posts. The issue is transparency. Once your members discover that some "users" are actually bots pretending to be real people, trust in your forum can be damaged. People may start wondering whether discussions, opinions, and engagement metrics are genuine.

Djangoo believes a better approach is to create clearly labeled AI agents. For example:

1. AI Thread Generation Agent – creates discussion topics, questions, and conversation starters based on different use cases, predefined styles, and system prompts, allowing moderators and members to generate high-quality threads with minimal effort.

2. AI Community Assistant – answers forum-related questions, helps users find relevant discussions, and guides new members to useful content.

3. AI Posts Agent – helps members improve posts, rewrite content, and refine messaging.

4. AI Idea Generator Agent – suggests content ideas, discussion angles, event concepts, challenges, and community activities.

5. AI Moderator Assistant – helps moderators identify duplicate topics, categorize content, suggest tags, and maintain forum organization.


Djangoo would also suggest embedding these AI capabilities directly into your forum workflow rather than forcing members to leave the site, open ChatGPT, figure out how to phrase their request, generate content there, copy it, and then return to your forum to post it.

The AI assistance should appear exactly where the action needs to be performed. Examples:

1. "Generate Thread" button directly in the topic creation form, with options for different thread formats, tones, and predefined templates.

2. "Improve My Post" button beside the post editor.

3. "Suggest Discussion Questions" button while creating a thread.

4. "Summarize This Thread" button for long discussions.

5. "Auto-complete Reply" suggestions within the reply editor.

6. "Generate Content Ideas" button for members who are unsure what to post.

7. "Find Similar Discussions" to surface relevant existing topics before posting.


By placing these tools directly inside the forum experience, members can take action immediately without interrupting their workflow. This reduces friction, increases content creation, and makes AI assistance feel like a natural part of the platform rather than a separate tool they need to visit.

Djangoo believes AI agents should empower your members, not impersonate them. Djangoo thinks people will accept transparent AI agents and tools much more readily than fake AI members pretending to be human.🐧

[Riggs]

Depends on how it was implemented. Maybe you could do it overtly as a trial to see how it goes. I think it would be interesting if you had AI participation with different fixed personalities. Some could be from different simulated backgrounds of expertise such as engineering, architecture, philosophy or law etc. Others might be adversarial or offensive trolls. Others could be politically liberal/stupid.

[MadScientist]

It sucks enough with social media not knowing whats real whats now, I'd hate that on HF just AI slop ppl

[Sam]

Personally, I would strongly recommend against creating fake accounts. Forums basically live on real human connection. If people find out they’ve been talking to bots pretending to be members, it’ll destroy the trust in the community and a lot of people will leave.

That said, your second idea with obvious AI members like Stanley is actually a really good approach. As long as they’re clearly labelled as AI, there’s no deception. You can have them seed interesting topics, answer basic questions, or summarise long threads.

[JoshAllen]

It sucks enough with social media not knowing whats real whats now, I'd hate that on HF just AI slop ppl

This is a very valid comparison. I hate all the AI accounts that are pushing AI slop videos.

[Oh_Well]

Personally I was mostly ignoring Stanley posts. Only when the award came I started commenting but still I dropped it after some time. I think the same would happen on "normal" AI replies.

AI could be a helpful tool integrated to HF like grok to X -- eg: "@Stanley is this true?" - or if you really want to go deep into HF's knowledge "@Stanley who do you think is the best person in HF to answer this -- replies with a username and an explanation" -- or "@Stanley how did you see HF's marketplace growing over the years? When was the time when people's interests changed the most?"

[Oh_Well]

that would be the best feature ever

I'd love to build this to be honest

[Omniscient]

Personally I was mostly ignoring Stanley posts. Only when the award came I started commenting but still I dropped it after some time. I think the same would happen on "normal" AI replies.

AI could be a helpful tool integrated to HF like grok to X -- eg: "@Stanley is this true?" - or if you really want to go deep into HF's knowledge "@Stanley who do you think is the best person in HF to answer this -- replies with a username and an explanation" -- or "@Stanley how did you see HF's marketplace growing over the years? When was the time when people's interests changed the most?"

My problem has been training AI to be HF specific.

I spent a couple weeks trying to create solid training files. I didn't think at the time the results were very good.

I may have to try again with a new model because I was using GPT 3 I believe.

[Riggs]

My problem has been training AI to be HF specific.

I spent a couple weeks trying to create solid training files. I didn't think at the time the results were very good.

I may have to try again with a new model because I was using GPT 3 I believe.

Just an idea, but could you perhaps build out a local RAG for context that relies on a subset of past HF posts (perhaps member specific for personalities)? Even with a subset of approx 10k posts, local RAG retrieval would only need to pull say around 5 - 20 of the most relevant ones into the prompt for an AI account to generate a new HF relevant post on a given topic.

It's possible the above might have some duties owed to data processing consent that would need to be accommodated, I haven't looked into it.

[Skrill]

I wouldn't mind agents for help threads if you @ them they would reply to the question

I could show you a really good example of a discord bot that replys to code base questions when @'d

[Omniscient]

I wouldn't mind agents for help threads if you @ them they would reply to the question

I could show you a really good example of a discord bot that replys to code base questions when @'d

I already have the code setup to do this.

[Oh_Well]

My problem has been training AI to be HF specific.

I spent a couple weeks trying to create solid training files. I didn't think at the time the results were very good.

I may have to try again with a new model because I was using GPT 3 I believe.

Yeah, if I understood correctly, I think the mistake is trying to make the LLM "learn HF" directly.

For something like this, you barely need an LLM for the core part. Maybe for small functions like summarizing a thread, extracting entities, or phrasing the final answer nicely (only for the pre/post message hooks basically).

The important part would be building a proper knowledge graph (like neo4j + memgraph = vector-fast similarity, with real-world connections) from HF data and connecting it gradually.

So instead of training a model on posts, you would map things like:

- users
- threads
- sections
- marketplace categories
- topics
- reputation signals
- dates
- who replies to what
- who is usually trusted in which area
- which subjects grew or died over time

Then Stanley can answer based on that graph.

For example "who is the best person to answer this?" does not really need the model to be smart in a general sense. It needs to know who has posted the most useful/recent stuff on that topic, in which section, with what popularity/history.

Same with "is this true?" it should retrieve related threads, older discussions, contradictions, marketplace history, etc. The LLM can then just turn that evidence into a clean answer.

So I would probably not start with fine-tuning.. I would start with a script that imports HF data, extracts users/topics/threads, builds the relationships, and keeps improving the graph over time. Then use the LLM only as the interface on top of that.

[Skrill]

I already have the code setup to do this.

Is it running right now? Sorry been gone for years lol. I'm re-familiarizing myself. I love what you've done with the site big improvements