[quote=“kofa, post:16, topic:43615”]
I agree with you that an LLM summary can be useful, …
[quote]
István, I’m referring to more that just “an LLM summary”. It is not unusual to have a conversation with AI, even correcting wrong information - or having to re-phrase the query - or having to tell it my understanding in order to test it’s previous responses.
I understand that English is not your native language but, taken literally, “clean” is implying that any LLM AI-based information posted here somehow “taints” the Forum whether it is correct or not.
Again, terminology, why will information provided by “regular users” somehow be always “clean”?
It won’t, and we have a long history here in this forum of drilling down to the most technically correct things, even when its pedantic and unwarranted
I guess I’d prefer LLM content to be curated/provided by a human, preferrably someone who has the critical thinking skills to make sure the LLM input contributes materially to the conversation. Sorta like:
“I provided this query to (pick your LLM poison), and got this response: …”
What is the difference between querying with a search engine then reading the ‘hits’, picking one, then copying and quoting the relevant part here - apart from speed?
A big difference. AI is likely to mash together text from different texts about the topic. For example, an ImageMagick user had problems with an IM command suggested by AI. Parts of the command were suitable for the Windows BAT shell, and parts were for bash. This was obvious to anyone who knew both Windows BAT and bash. Neither the IM user, nor AI, had that knowledge.
… then, by posting such information on the site, it will be picked up by search engines (AI-powered or traditional), and presented to others.
Yes. I think that is likely to become a major problem. AI can already mash-up good text to produce garbage. When it starts using AI garbage as input, the output really will be useless.
Perhaps the AI trainer can watch out for words like “generated by AI”, and discount that text for training.
I’m not anti-AI, or more specifically anti-LLM. There are many potential benefits. But also many possible pitfalls.
Good point … I read the one about ImageMagick, and was thoroughly confused because I too know little to nothing about shell stuff, BAT stuff, or most other kinds of scripting stuff.
It has never appeared to me that discuss.pixls.us has had hard prohibitions on anything other than abusive posts. In that spirit, I tend to think that there should not be a prohibition on posting AI-generated material. But I think that we should at least encourage people to label anything AI-generated as such.
Generated content is fine as long as it is curated, considerate and not dis/misinformation[1]. Spam on the other hand is unhealthy for the body, soul, vegans or certain religious folks.
[1] The thing about dis/misinformation is that it is often insidious and easy to mistaken as factual.
Which leads to a slight tangent regarding the discussion of reactions other than “like” - I think the negative reactions would be detrimental to the forums, but “haha” would be useful sometimes.
Along the lines of “LOL look what ChatGPT came up with this time! HAHA!”
So far my favorite uses of AI have been ones that “leaned in” to their tendency to hallucinate.
On the other side of things, people using Github Copilot and similar tools scare the sh*t out of me, who knows where that garbage might end up.
It is interesting that the majority of AI-related posts in this forum are like “look how wrong it was!!” - whereas my personal experience shows ChatGPT to be right most of the time, with it’s misunderstandings quickly cleared up by further questions or even corrections by myself.
In a playful mood, I just asked ChatGPT 4o “what is the Exposure Triangle?” Amongst all the standard stuff, it told me that the ISO setting adjusts the sensitivity of the sensor (as does a great many if not most tutorials on the subject). That simply ain’t so. So I told it that, with reasons why.
It agreed and corrected itself, mostly repeating my wording.
Whether it learned from that exchange, I have no idea.
They are not retained after each correction, however, as long as the new information is within the ‘context window’, it will shape subsequent output.
Try asking the same question in a ‘temporary conversation’ (I don’t use ChatGPT often, but I think that’s what it’s called).
They are good at generating text, without understanding. For example, ChatGPT can tell you what tic-tac-toe is, or write a Python program to play the game using the optimal strategy, but ChatGPT 3.5 not only makes random moves, it does not even detect when you won. GPT4 also makes random moves, but detects when it loses.
Thing is, it literally isn’t even wrong. It literally has no idea what it’s saying. I’m not even sure that “saying” is the right word for whatever it does.
It “says” things that are right or wrong in the same way that the photocopier I used to photocopy my graduate thesis did.
The university lecturer is coming out in me but I see this as a form of plagiarism or the photographer comes out in me and I see it as a breach of copyright since they are stealing someone else’s words and using them without expressed permission. That is besides the fact that huge errors are made. The only AI I like is the voice from google maps that guides me back to the safety of my hotel in Morocco so I can read this post.