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.
I legitimately don’t understand how someone can interact with an LLM for more than 30 minutes and come away from it thinking that it’s some kind of super intelligence or that it can be trusted as a means of gaining knowledge without external verification. Do they just not even consider the possibility that it might not be fully accurate and don’t bother to test it out? I asked it all kinds of tough and ambiguous questions the day I got access to ChatGPT and very quickly found inaccuracies, common misconceptions, and popular but ideologically motivated answers. For example, I don’t know if this is still like this but if you ask ChatGPT questions about who wrote various books of the Bible, it will give not only the traditional view, but specifically the evangelical Christian view on most versions of these questions. This makes sense because they’re extremely prolific writers, but it’s simply wrong to reply “Scholars generally believe that the Gospel of Mark was written by a companion of Peter named John Mark” because this view hasn’t been favored in academic biblical studies for over 100 years, even though it is traditional. Similarly, asking it questions about early Islamic history gets you the religious views of Ash’ari Sunni Muslims and not the general scholarly consensus.
I mean. I’ve used AI to write my job mandated end of year self assessment report. I don’t care about this, it’s not like they’ll give me a pay rise so I’m not putting effort into it.
The AI says I’ve lead a project related to windows 11 updates. I haven’t but it looks accurate and no one else will be able to dell it’s fake.
So I guess the reason is they are using the AI to talk about subjects they can’t fact check. So it looks accurate.
I have a friend who constantly sends me videos that get her all riled up. Half the time I patiently explain to her why a video is likely AI or faked some other way. “Notice how it never says where it is taking place? Notice how they never give any specific names?” Fortunately she eventually agrees with me but I feel like I’m teaching critical thinking 101. I then think of the really stupid people out there who refuse to listen to reason.
The results I get from chatgpt half the time are pretty bad. If I ask for simple code it is pretty good but ask it about how something works? Nope. All I need to do is slightly rephrase the question and I can get a totally different answer.
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.
I usually skip the AI blurb because they are so inaccurate, and dig through the listings for the info I’m researching. If I go back and look at the AI blurb after that, I can tell where they took various little factoids, and occasionally they’ll repeat some opinion or speculation as fact.
At least fuck duck go is useful for video games specifically, but that one more or less just copy pasted from the wiki, reddit, or a forum shits the bed with EUV specifically though.
The paid options arnt any better. When the well is poisoned it doesn’t matter if your bucket is made of shitty rotting wood, or the nicest golden vessel to have graced the hands of a mankind.
Your getting lead poisoning either way. You just get to give away money for the privilege with one and the other forces the poisoned water down your throat faster.
I’ve asked it for a solution to something and it gives me A. I tell it A doesn’t work so it says “Of course!” and gives me B. Then I tell it B doesn’t work and it gives me A…
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
I got hallucination from trying to find a book I read but didn’t know the title of. And hallucinated NBA play off results of the wrong team winning. And gotten basic math calculations wrong.
Its a language model so its purpose is to string together words that sound like sentences, but it can’t be fully trusted to be accurate. Best it can do is give you source so you can got straight to the resource to read that instead.
It’s decent at generating basic code, and testing yourself to see if it outputs what you want. But I don’t trust it as a resource when it comes to information when even wrong sports facts have been provided.
The three things I’ve found search engine LLMs to be useful for. Searching for laptop since it’s absurdly good at finding weird fucking regional models or odd configurations that arnt on the main pages of most shops.
Like my current laptop wasnt on newegg Amazon or even msi’s own shop. It was on a fucking random ass page on their website that nothing linked to and was some weird ass model that wasn’t searchable even.
The second most useful one was generating a metric crapload of boiler plate json files for a mod.
The third thing is bad dnd roleplaying while I’m bored at work. The hallucinations are a upside lol
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?
Just a few days ago I tried to feed my home automation logs to copilot in hopes that it might find a reason why my controller jams randomly multiple times per hour. It confidently claimed that as my noise level reported by controller is -100dB (so basically there’s absolutely nothing else on that frequency around, pretty much as good as it can get) it’s the problem and I should physically move the controller to less noisy area. A decent advice in itself, it might actually help on a lot of cases, but in my scenario it’s a completely wrong rabbit hole to dig in. I might still move the thing around to get better reception on some devices but it doesn’t explain why the whole controller freezes for several minutes on random intervals.
I use them at work to get instructions on running processes and no matter how detailed I am “It is version X, the OS is Y” it still gives me commands that don’t work on my version, bad error code analysis, etc.
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 manual is a keyword in PostgreSQL, which is not.
It’s more the hallucinations are due to the fact we have trained them to be unable to admit to failure or incompetence.
Humans have the exact same “hallucinations” if you give them a job then tell them they aren’t allowed to admit to not knowing something ever for any reason.
You end up only with people willing to lie, bullshit and sound incredibly confident.
We literally reinvented the politician with LLMs.
None of the big models are trained to be actually accurate, only to give results no matter what.
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.
I legitimately don’t understand how someone can interact with an LLM for more than 30 minutes and come away from it thinking that it’s some kind of super intelligence or that it can be trusted as a means of gaining knowledge without external verification. Do they just not even consider the possibility that it might not be fully accurate and don’t bother to test it out? I asked it all kinds of tough and ambiguous questions the day I got access to ChatGPT and very quickly found inaccuracies, common misconceptions, and popular but ideologically motivated answers. For example, I don’t know if this is still like this but if you ask ChatGPT questions about who wrote various books of the Bible, it will give not only the traditional view, but specifically the evangelical Christian view on most versions of these questions. This makes sense because they’re extremely prolific writers, but it’s simply wrong to reply “Scholars generally believe that the Gospel of Mark was written by a companion of Peter named John Mark” because this view hasn’t been favored in academic biblical studies for over 100 years, even though it is traditional. Similarly, asking it questions about early Islamic history gets you the religious views of Ash’ari Sunni Muslims and not the general scholarly consensus.
I mean. I’ve used AI to write my job mandated end of year self assessment report. I don’t care about this, it’s not like they’ll give me a pay rise so I’m not putting effort into it.
The AI says I’ve lead a project related to windows 11 updates. I haven’t but it looks accurate and no one else will be able to dell it’s fake.
So I guess the reason is they are using the AI to talk about subjects they can’t fact check. So it looks accurate.
I have a friend who constantly sends me videos that get her all riled up. Half the time I patiently explain to her why a video is likely AI or faked some other way. “Notice how it never says where it is taking place? Notice how they never give any specific names?” Fortunately she eventually agrees with me but I feel like I’m teaching critical thinking 101. I then think of the really stupid people out there who refuse to listen to reason.
The results I get from chatgpt half the time are pretty bad. If I ask for simple code it is pretty good but ask it about how something works? Nope. All I need to do is slightly rephrase the question and I can get a totally different answer.
I mainly use it as a search engine, like: “Find me an article that explains how to change a light bulb” kinda shit.
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.
I pay for Kagi search. It’s amazing
I usually skip the AI blurb because they are so inaccurate, and dig through the listings for the info I’m researching. If I go back and look at the AI blurb after that, I can tell where they took various little factoids, and occasionally they’ll repeat some opinion or speculation as fact.
Usually the blurb is pure opinion.
At least fuck duck go is useful for video games specifically, but that one more or less just copy pasted from the wiki, reddit, or a forum shits the bed with EUV specifically though.
This is the one time in all of human history where autocorrecting “fuck” to “duck” would’ve been correct.
Worst part is I’m pretty sure it autocorrected duck to fuck cause I’ve poisoned my phones autocorrect with many a profanities.
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.
The paid options arnt any better. When the well is poisoned it doesn’t matter if your bucket is made of shitty rotting wood, or the nicest golden vessel to have graced the hands of a mankind.
Your getting lead poisoning either way. You just get to give away money for the privilege with one and the other forces the poisoned water down your throat faster.
I’ve asked it for a solution to something and it gives me A. I tell it A doesn’t work so it says “Of course!” and gives me B. Then I tell it B doesn’t work and it gives me A…
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
I got hallucination from trying to find a book I read but didn’t know the title of. And hallucinated NBA play off results of the wrong team winning. And gotten basic math calculations wrong.
Its a language model so its purpose is to string together words that sound like sentences, but it can’t be fully trusted to be accurate. Best it can do is give you source so you can got straight to the resource to read that instead.
It’s decent at generating basic code, and testing yourself to see if it outputs what you want. But I don’t trust it as a resource when it comes to information when even wrong sports facts have been provided.
The three things I’ve found search engine LLMs to be useful for. Searching for laptop since it’s absurdly good at finding weird fucking regional models or odd configurations that arnt on the main pages of most shops.
Like my current laptop wasnt on newegg Amazon or even msi’s own shop. It was on a fucking random ass page on their website that nothing linked to and was some weird ass model that wasn’t searchable even.
The second most useful one was generating a metric crapload of boiler plate json files for a mod.
The third thing is bad dnd roleplaying while I’m bored at work. The hallucinations are a upside lol
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?
Just a few days ago I tried to feed my home automation logs to copilot in hopes that it might find a reason why my controller jams randomly multiple times per hour. It confidently claimed that as my noise level reported by controller is -100dB (so basically there’s absolutely nothing else on that frequency around, pretty much as good as it can get) it’s the problem and I should physically move the controller to less noisy area. A decent advice in itself, it might actually help on a lot of cases, but in my scenario it’s a completely wrong rabbit hole to dig in. I might still move the thing around to get better reception on some devices but it doesn’t explain why the whole controller freezes for several minutes on random intervals.
I use them at work to get instructions on running processes and no matter how detailed I am “It is version X, the OS is Y” it still gives me commands that don’t work on my version, bad error code analysis, etc.
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.It’s more the hallucinations are due to the fact we have trained them to be unable to admit to failure or incompetence.
Humans have the exact same “hallucinations” if you give them a job then tell them they aren’t allowed to admit to not knowing something ever for any reason.
You end up only with people willing to lie, bullshit and sound incredibly confident.
We literally reinvented the politician with LLMs.
None of the big models are trained to be actually accurate, only to give results no matter what.
What is a hallucination if not AI being confidently mistaken by making up something that is not true?