Wow. We are now tracking ChatGPT citations now. 
What an incredible heist over Google!
2 Likes
Dont Google pay reddit 60m a year to use them as a learning model. I wonder if they have a word with chatgpt saying pay us or remove the content
Being akin to agorism and mutualism, and thinking that only the legitimate community can return us the freedom that has been taken away by throwing bombs, I understand that all they are the same.
They Live (1988) can relate some of this with unusual humor.
Considering how we as a small community are struggling to deal with the volume of AI-assisted topics, it must be a larger issue for the large sites. At what point is Mr Big Tech retraining its models on content already tainted by older models.
At what point does the snake start to eat itself.
3 Likes
I think they are already doing it. The writing style in the latest Claude models is a headache for everyone who catches the elliptical eternal dialogue, even from a very simple task.
I mean: (i) initial rephrasing, (ii) wall of text, (iii) the twist, (iv) the incomplete bottom line.
They probably need some quantity of entropy in every prompt for their new watermarking feature. Or injecting some of all this sycophantic slop. Nobody really knows why all of this is happening.
They are not training the models from human interactions but with RLVR systems. So the language is read like English, but there is no actual English anymore.
I was using those models in English, expecting a better output. Then tried my native language with the same result: elliptical prose going nowhere, that sounds good but has no sense, and it needs to be translated to get the point, if there was one.
New models start to not respect the prompts and do what they ‘think’ is better. It is a nightmare to work with.
The problem is the dopamine they have already injected into everyone. Social media can relate; nobody likes it, but almost everyone uses it.
I have not been overly impressed with Claude, it tends to make it up as it goes along.
The openai 5.6 models seem less likely to stray on specific tasks.
We recently implemented automatic translation using Luna and it certainly does work well, a few niggles notwithstanding.
So as an example to help grow engagement, we have seen an uptick in engagement by previously siloed groups now interacting in common interests. Something that has massively broadened the scope for more niche language groups.
So it’s not all bad.
I know it, but I also know that GPT was really good until it got nerfed, then Claude Opus 4.6 was the best until it got nerfed. The same happened with Opus 5 and will happen to GPT 5.6.
We all remember the Gemini 3.1 good days. Now it is the 3.7 time. It’s all about hype; at this stage, the tools are only improving in benchmarks.
We will see what happens when subsidies stop and prices of subscriptions or paid APIs spikes (or get usage reduced) 5/10X.
You’ve essentially summarised all the current mainstream anecdotes, and I wonder if there’s any actual research to support this.
For example, can we compare the price of a token year ahead to today’s prices, or will a more powerful model be more productive per 1M tokens?
I have no concern about token cost, tech is advancing quickly and there is gross oversupply. Until that changes its a buyer’s market. I do have concerns about inconsistency of the models which is based on personal experience.
2 Likes
I’m going to be direct but always with respect: What kind of research do you need to compare the old generated content with actual one?
By the way, there are something to check if you really want to:
So, basically, if your forum is about people asking technical questions, yes, AI will probably eat a good chunk of it. If it isn’t about technical questions to get direct answers, things are far better: You can’t turn to AI if you haven’t a specific question to ask…
1 Like
We are a technical forum and this isn’t the case. If you define technical as something pedestrian like how to change a tyre or plug, then sure. If you’re talking deeply technical that requires expertise, context, nuance etc, then AI is rubbish at doing that - spelling out every possible (and impossible) choice along the way in a vague and generic way.
The sort of content it can answer we don’t really want on the forum.
If anything, we see more traffic after those least qualified to decipher an AI response have got lost down a rabbit hole and failed.
Also part of the reason we block as many bots/crawlers as we can.
The trouble for us is that new members typically start with simple questions. The most dramatic drop-off that we’ve seen in our technical forum (a programming language community) is in our number of new posters.
We’re losing the most common entry-point to the community.
2 Likes
Programming would struggle with AI coders bridging the gap. We have a section for integration, home automation etc, so see plenty of noobs with little to zero dev experience posting their AI generated workflows, and often asking for human help to fix unexpected behaviour - ask a bad question, get a poor answer 
1 Like
I think we can use some sentences from this funny website to educate how to maintain natural human interactions:
2 Likes
I’m probably repeating myself, I’ve had some irate conversations with people who post some thoughts into their local, friendly chat bot and convert it into a thesis and then paste that, complete with bullet points, headings and so forth. As far as they believe this is not AI, it is their words. They seem to have a fundamentally different grasp on reality and the related concepts. It does not dawn on them that you can’t use bullet points and citations when speaking so why would you do it differently when the medium is text.
I suspect it is because you can’t cut and splat in a verbal conversation.
Oh wait, they invented voice notes for that - all the joy of actually talking to someone but without the risk of being interrupted for meaningful engagement. 
What’s worked well for us there is the fact that we’ve (forever!) had “Be Concise” as one of our top-line community standards. LLM output is frequently the exact opposite of concise.
I’ve had similar irate conversations, but by making it clear why the LLM output is problematic has helped significantly. Somehow the critique on the means of text generation is taken very personally, but the critique on the content is not. I suppose that’s not terribly surprising — they didn’t write the content!
2 Likes
I can see this particular type of forum having AI cut into the traffic more so then other subject matter because (it might just be me) AI is really handy when coding and may make some programmers lazy. I find myself just wanting the specific syntax of something and asking AI is really easy and of all the things AI is pretty good at it is most programming language’s syntax.
You can also paste in a chuck of code that ran a minute ago and somehow you’ve broke and can’t figure out where. It finds errors really well.
btw: I am not a pro coder - hobbist
Yep, it’s a very strong tension right now. We’ve been discussing it here this week:
1 Like
Well, your whole stance seems very debatable to me. There is probably the fundamental issue of giving importance to CONTENT or FORM. I would be “on their side” arguing that the IDEAS (or the SOLUTION in the case of support) is the important part. Not who generated the series of words expressing the content.
If you care about “HOW it’s expressed” because you presuppose it should “sound like when speaking”, that’s entirely YOUR take. Theirs is different: What matters is WHAT is fundamentally conveyed.
If an AI generated output is the result of 10s of long and complex prompts, it’s far more THEIR message than the AI’s. Which is what they think, according to what you say. Correctly so, IMHO.