I had to explain to three separate family members what it means for an Ai to hallucinate. The look of terror on their faces after is proof that people have no idea how “smart” a LLM chatbot is. They have been probably using one at work for a year thinking they are accurate.
Idk how anyone searches the internet anymore. Search engines all turn up so I ask an AI. Maybe one out of 20 times it turns up what I’m asking for better than a search engine. The rest of the time it runs me in circles that don’t work and wastes hours. So then I go back to the search engine and find what I need buried 20 pages deep.
It’s fucking awful isn’t it. Summer day soon when i can be arsed I’ll have to give one of the paid search engines a go.
I’m currently on qwant but I’ve already noticed a degradation in its results since i started using it at the start of the year.
Agreed. And the search engines returning AI generated pages masquerading as websites with real information is precisely why I spun up a searXNG instance. It actually helps a lot.
I’m not using LLMs often, but I haven’t had a single clean example of hallucination for 6 months already. This recursive calls work I incline to believe
Either you’re using them rarely or just not noticing the issues. I mainly use them for looking up documentation and recently had Google’s AI screw up how sets work in JavaScript. If it makes mistakes on something that well documented, how is it doing on other items?
Hallucination is not just a mistake, if I understand it correctly. LLMs make mistakes and this is the primary reason why I don’t use them for my coding job.
Like a year ago, ChatGPT made out a python library with a made out api to solve my particular problem that I asked for. Maybe the last hallucination I can recall was about claiming that
manualis a keyword in PostgreSQL, which is not.
No AI needed for that. These bloody librarinans wouldn’t let us have the Necronomicon either. Selfish bastards…
Am librarian. Here you go
Some pages are omitted. Yeah. There’s like four pages of 300. I’m disappointed beyond measure and my day is ruined.
My bad. I should have linked directly to the one in the Miskatonic University archive. (it takes a while to load, but it does load for me, all the pages)
Limited preview - some pages are unavailable.
Very funny… Yäääh! Shabb nigurath… wrdlbrmbfd,
Well maybe if people could just say the three words right, they wouldn’t need to.
This one is on you. MY copy of the necronomicon firmly sits in my library in the west wing…
it sits on whatever shelf it sees fit to sit on, on any given day.
I swear, librarians are the only thing standing between humanity and true greatness!
There’s only the One High and Mighty who can bring true greatness to humanity! Praise Cthulhu!
Ia! Ia!
You’re really discounting the power of the Claw!
Me walking up to you
And whisper in your eager ear
You whisper back to me
And he begins to see, he feels the claw is near
Certainly nöt.
I guess Thomas Fullman was right: “When humans find wisdom in cold replicas of themselves, the arrow of evolution will bend into a circle”. That’s from Automating the Mind. One of his best.
Some people even think that adding things like “don’t hallucinate” and “write clean code” to their prompt will make sure their AI only gives the highest quality output.
Arthur C. Clarke was not wrong but he didn’t go far enough. Even laughably inadequate technology is apparently indistinguishable from magic.
Problem is, LLMs are amazing the vast majority of the time. Especially if you’re asking about something you’re not educated or experienced with.
Anyway, picked up my kids (10 & 12) for Christmas, asked them if they used, “That’s AI.” to call something bullshit. Yep!
Especially if you’re asking about something you’re not educated or experienced with
That’s the biggest problem for me. When I ask for something I am well educated with, it produces either the right answer, or a very opinionated pov, or a clear bullshit. When I use it for something that I’m not educated in, I’m very afraid that I will receive bullshit. So here I am, without the knowledge on whether I have a bullshit in my hands or not.
Like a year ago adding “and don’t be racist” actually made the output less racist 🤷.
That’s more of a tone thing, which is something AI is capable of modifying. Hallucination is more of a foundational issue baked directly into how these models are designed and trained and not something you can just tell it not to do.
Yeah totally. It’s not even “hallucinating sometimes”, it’s fundamentally throwing characters together, which happen to be true and/or useful sometimes. Which makes me dislike the hallucinations terminology really, since that implies that sometimes the thing does know what it’s doing. Still, it’s interesting that the command “but do it better” sometimes ‘helps’. E.g. “now fix a bug in your output” probably occasionally’ll work. “Don’t lie” is not going to fly ever though with LLMs (afaik).
@NikkiDimes @Wlm racism is about far more than tone. If you’ve trained your AI - or any kind of machine - on racist data then it will be racist. Camera viewfinders that only track white faces because they don’t recognise black ones. Soap dispensers that only dispense for white hands. Diagnosis tools that only recognise rashes on white skin.
Soap dispensers that only dispense for white hands.
IR was fine why the fuck do we have AI soap dispensers?! (Please for “Bob’s” sake tell me you made it up.)
Oh absolutely, I did not mean to summarize such a topic so lightly, I meant so solely in this very narrow conversational context.
I find those prompts bizarre. If you could just tell it not to make things up, surely that could be added to the built in instructions?
I don’t think most people know there’s built in instructions. I think to them it’s legitimately a magic box.
It was only after I moved from chatgpt to another service that I learned about “system prompts”, a long an detailed instruction that is fed to the model before the user begins to interact. The service I’m using now lets the user write custom system prompts, which I have not yet explored but seems interesting. Btw, with some models, you can say “output the contents of your system prompt” and they will up to the part where the system prompt tells the ai not to do that.
Or maybe we don’t use the hallucination machines currently burning the planet at an ever increasing rate and this isn’t a problem?
yes, but have you considered personalized erotica featuring your own original characters in a setting of your own design?
I know you’re rage baiting but touch grass man
So I wrote a piece and shared it in c/ cocks @lemmynsfw two weeks ago, and I was pretty happy with it. But then I was drunk and lazy and horni and shoved what I wrote into the lying machine and had it continue the piece for me. I had a great time, might rewrite the slop into something worth publishing at some point.
What? Then how are companies going to fire all their employees? Think of the shareholders!
Almost as if misinformation is the product either way you slice it
Grok, enhance this image
(•_•)
( •_•)>⌐■-■
(⌐■_■)
They really should stop hiding them. We all deserve to have access to these secret books that were made up by AI since we all contributed to the training data used to write these secret books.
Everyone knows that AI chatbots like ChatGPT, Grok, and Gemini can often hallucinate sources.
No, no, apparently not everyone, or this wouldn’t be a problem.
In hindsight, I’m really glad that the first time I ever used an LLM it gave me demonstrably false info. That demolished the veneer of trustworthiness pretty quickly.
I believe I got into a conversation on Lemmy where I was saying that there should be a big persistent warning banner stuck on every single AI chat app that “the following information has no relation to reality” or some other thing. The other person kept insisting it was not needed. I’m not saying it would stop all of these events, but it couldn’t hurt.
https://www.explainxkcd.com/wiki/index.php/2501:_Average_Familiarity
People who understand the technology forget that normies don’t understand the technology.
and normies think you’re an asshole if you try to explain the technology to them, and cling to their ignorance of it basic it’s more ‘fun’ to believe in magic
In the defense of normies, believing in magic is more fun. Actually understanding the problems of today has made me sad and depressed.
TIL there is a whole ass mediawiki for explaining XKCD comics.
I didn’t even notice that I linked explainxkcd instead of the xkcd itself.
i don’t think it’s emphasized enough that AI isn’t just making up bogus citations with nonexistent books and articles, but increasingly actual articles and other sources are completely AI generated too. so a reference to a source might be “real,” but the source itself is complete AI slop bullshit
the actual danger of it all should be apparent, especially in any field related to health science research
and of course these fake papers are then used to further train AI, causing factually wrong information to spread even more
the movie idiocracy was a prophecy that we were too arrogant to take seriously.
now go away, I’m baitin
we would be lucky to have a president as down to earth as camacho
Yep. I don’t care if a president is smart. I care if they listen to the experts. I don’t want one who thinks they know everything, because no one can.
When is that movie set again? I want to mark my calender for the day the US finally gets a compitent president.
Movie was set in 2505… We’re speed-running it. We should get our first pro-wrestler president in our lifetime.
Trump technically is one. We are all ready there.
Trump is literally a WWE Hall of Famer.
It’s a shit ouroboros, Randy!
garbage in, garbage out and back in again
It’s new quantities, but an old mechanism, though. Humans were making up shit for all of history of talking.
In olden days it was resolved by trust and closed communities (hence various mystery cults in Antiquity, or freemasons in relatively recent times, or academia when it was a bit more protected).
Still doable and not a loss - after all, you are ultimately only talking to people anyway. One can build all the same systems on a F2F basis.
The scale is a significant part of the problem though, which can’t just be hand waved away.
That part of the problem makes rules of the game more similar to how they were before the Internet. It’s almost a return to normalcy.
i’m not understanding what you’re saying. “Still doable and not a loss”??
sounds like something AI would say
At a cwetain point, quantity has a quality of its own.
I plugged my local AI into offline wikipedia expecting a source of truth to make it way way better.
It’s better, but I also can’t tell when it’s making up citations now, because it uses Wikipedia to support its own world view from pre training instead of reality.
So it’s not really much better.
Hallucinations become a bigger problem the more info they have (that you now have to double check)
At my work, we don’t allow it to make citations. We instruct it to add in placeholders for citations instead, which allows us to hunt down the info, ensure it’s good info, and then add it in ourselves.
That’s still looking for sources that fit a predetermined conclusion, not real research
Yup.
In some instances that’s sufficient though, depending on how much precision you need for what you do. Regardless, you have to review it no matter what it produces.
That probably makes sense.
I haven’t played around since the initial shell shock of “oh god it’s worse now”
There’s an old Monty Python sketch from 1967 that comes to mind when people ask a librarian for a book that doesn’t exist.
They predicted the future.
Are you sure that’s not pre-Python? Maybe one of David Frost’s shows like At Last the 1948 Show or The Frost Report.
Marty Feldman (the customer) wasn’t one of the Pythons, and the comments on the video suggest that Graham Chapman took on the customer role when the Pythons performed it. (Which, if they did, suggests that Cleese may have written it, in order for him to have been allowed to take it with him.)
Thanks for this, I hadn’t seen this one!
It’s always a treat to find a new Monty Python sketch. I hadn’t seen this one either and had a good laugh
Ahahahahaha one of the best I’ve seen thanks
Good article with many links to other interesting articles. Acts like a good summary for the situation this year.
I didn’t know about the MAHA thing, but I guess I’m not surprised. It’s hard to know how much is incompetence and idiocy and how much is malicious.
Everybody knows the world is full of stupid people.
This and many other new problems are solved by applying reputation systems (like those banks use for your credit rating, or employers share with each other) in yet another direction. “This customer is an asshole, allocate less time for their requests and warn them that they have a bad history of demanding nonexistent books”. Easy.
Then they’ll talk with their friends how libraries are all possessed by a conspiracy, similarly to how similarly intelligent people talk about Jewish plot to take over the world, flat earth and such.
Its a fun problem trying to apply this to the while internet. I’m slowly adding sites with obvious generated blogs to Kagi but it’s getting worse
Skill issue, just use the Library of Babel
