[a / b / c / d / e / f / g / gif / h / hr / k / m / o / p / s / t / u / v / vg / vm / vmg / vr / vrpg / vst / w / wg] [i / ic] [r9k / s4s / vip] [cm / hm / lgbt / y] [3 / aco / adv / an / bant / biz / cgl / ck / co / diy / fa / fit / gd / hc / his / int / jp / lit / mlp / mu / n / news / out / po / pol / pw / qst / sci / soc / sp / tg / toy / trv / tv / vp / vt / wsg / wsr / x / xs] [Settings] [Search] [Mobile] [Home]
Board
Settings Mobile Home
/g/ - Technology


Thread archived.
You cannot reply anymore.


[Advertise on 4chan]


Why does nobody talk about AI hallucinations anymore?
>>
we already talked about it months ago and nothing has changed
there's no point in repeating myself
>>
File: daysoftheweek.jpg (171 KB, 1080x2340)
171 KB JPG
>>109112572
yeah man hallucinations are solved, it is really intelligent now
>>
>>109112572
>AI is often wrong and can’t be trusted
What else is there to say?
>>
>>109112597
I'd have dropped the convo when he put Sunday first. So much for weekend.
>>
>>109112572
>Why does nobody talk about AI hallucinations anymore?
Because it's a well known character quirk of LLMs.
>>
>>109112572
There is a thread about it right now >>109109584
>>
>>109112585
Link the thread coward
>>
>>109112689
Shut the fuck up retard, your thread is gay.
>>
>>109112697
You're thread is gayer
>>
>>109112572
Same reason why nobody talks about bad or good weather
The topic became boring.

AI is still useless for most cases, as it does more harm than good.
Using it in a production environment we'll never reach 100%. How it won't even reach 20%.
>>
>>109112694
I don't remember which one it is but they're all the same anyway.
Is there anything specific you were wondering about or what?
People post about it here, I give my opinion, that's the end of it.
>>
>>109112572
they're a mathematical certainty (as openAI proved)
>>
>>109112572
thanks for the shitter screencap. great contribution.
>>
>>109112572
It's just something inherent to the system and it doesn't look like it will ever disappear, not unless we get """AI""" that functions completely differently from the present crop. Hallucinations are always present and if you use LLMs then you must always take them into account or skip LLMs entirely.
>>
>>109112572
Because we now know only coding matters, for everything else it's shit.
>>
>>109112572
kek, these AI fucks are so funny
>less talk about AI is good, because it is now normal!
>>
>>109112759
Ya its something I think about when I see it being utilized in schools and for younger folks.
>>
>>109112658
Look at a calendar.
>>
>>109112572
Quoting X (The Everything Site) on the AI bubble (that exists on X (The Everything Site))
>>
isn't it weird that a fucking digital intelligence has more knowledge than any living human ever?
there are 2 options;
1. aliens gave us tech too early and we supposed to evolve to the same level
2. everything is just a piece of data, even our existence
>>
>>109112572
It has been mitigated to a large extent
>>
it annoys me so much when you see someone just posting a screenshot of an ai response as an answer
ai can be helpful as a springboard, especially since search engines are fucked beyond saving thanks to SEO... and ai slop sites
but so many morons just take it at face value instead of using it to find the actual source, ai really makes people stupider
>>
>>109112572
because it's mostly solved. You just need to make sure you don't go beyond context threshold, have another ai from different distillation or learning to review the original AI's output, then have that review be reviewed by another third party agent, and then have that reviewed by......
>>
>>109113517
but who will review the final reviewer?
>>
>>109113000
Most countries calendars start in monday, not sunday, thougheverbeit.
>>
>>109112597
>sonnet 4.6
>>
>>109113499
Yes it is very bad. I'm seeing it now too more big picture. Lots of people taking it all at face value.
>>
>>109113499
It lends me to believe that these people aren't even reading what has been output.
>>
>>109112572
problems arent as obnoxious
in gpt-3 days if you ask about 'donald vance the us army general in ww1' it would've told you some fabulously made up nonsense
nowadays those don't
>>
>>109113594
This is why people don't post as much about AI hallucination. Someone will always reply with ">" and the model name, implying that had you paid more money to use a higher tier model, it wouldn't have hallucinated. And usually it wouldn't in the particular case in question, but would do so in other ways at other times. But who wants to spend tokens on top tier models just to find a situation where they mess up and make a gotcha post?
>>
>>109113658
>why don't people complain about slow cars anymore?
>because all cars are made fast now
>but look, this car from 1985 is slow!
>>
>>109113676
Except that those lower tier models are still being pushed hard as the defaults that both new free and paying users get unless they switch it to a more expensive one.
1985 cars aren't being marketed or sold new anymore.
>>
>>109113545
This is antisemitic
>>
>>109112572
>Why does nobody talk about AI hallucinations anymore?
They do. Fuck your shitty false premise, nigger faggot.
>>
>>109112572
>Why does nobody talk about AI hallucinations anymore?
people do. this is ai bro cope
>>
File: 1764285433327833.png (431 KB, 1547x1745)
431 KB PNG
>>109112597
>>
>>109112572
>Why does nobody talk about AI hallucinations anymore?
Because it's become accepted as a commonplace temporary defect.
>>
Why does nobody talk about water being wet anymore?
>>
>>109112572
Why does nobody talk about human hallucinations
>>
>>109113545
>Most countries calendars
The ones that matter have their week start on Sunday. I don't care how Africans handle their calendars.
>>
>>109112572
Piece of shit hallucinates literally constantly. Lies about everything and when I point it out it says sorry and lies again.
>>
>>109112572
>Why does nobody talk about AI hallucinations anymore?
because everybody just accepted it as a fact that it's a thing AI does
>>
There are benchmarks for hallucinations btw. Chatgpt 5.5 hallucinates the most. GLM 5.2 the least
>>
>>109112572
All critical thinking skills have atrophied to the point that they just believe the lies now.
>>
File: youhavenoidea.jpg (17 KB, 356x390)
17 KB JPG
>>109114918
we live in a post-truth world
and it's only gonna get worse
>>
>>109114942
this. it's almost impossible to find real info nowadays. Even itt there are probably shills and bots talking
>>
File: 1765685297957371.png (2.11 MB, 1122x1402)
2.11 MB PNG
>>109114942
>>
>>109112572
how is ai supposed to augment humans? rather than killing us all, we'll get cyberbrains?
>>
>>109115030
most people (especially americans) are dumb as fuck and can't read at highschool age. the best product never becomes the best seller, just the idea people have when they buy the cheaper product that can only do 80% of that
>>
>>109114390
You only care to have your bussy gapped by africans
>>
>>109115684
I'm not European
>>
>>109112572
modern LLMs hallucinate all the time.
when was the last time it gave you a link to a webpage that doesn't actually exist? Probably fuckin' today if you use it with any frequency
>>
>>109115955
The US has a greater amount of niggers than europe, though.
>>
>>109116539
do not confuse our American niggers with your african warlord niggers.
>>
>>109113705
> those lower tier models are still being pushed hard as the defaults
You are acting like the webtoddler interface would be the pinnacle of AI usage. It's just a commercial frontend they don't give two shits about. If you take AI seriously then use Opus model from API. Sonnet models are cost effective mid-tier stuff, of course they are pushing it as a default for you since you don't care either way. Sonnet is a budget model bro, look it up.
>>
>>109114425
there are two kinds of ai users, irrelevant idiots who never ask any interesting questions and therefore only get prepared answers, and people who do serious work and reach the limits of what ai is capable of talking about
>>
Every other week there is an article about some lawyer getting reamed by a judge for submitting paperwork full of hallucinations, last one I read had both sides of a lawsuit kicked out of court.
>>
>>109116558
Oh yeah, because there's totally zero warlords and armed militia in the US, specifically in democrat ruled big cities.
And neva4get the fact that a coked up nigger can easily get a gun and start gunning down shop owners to steal 200 bucks from the cash register.
>>
>>109112572
it will still hallucinate occasionally, enough that i treat it like i would treat anything else written on the internut: plausible but not certain. it has gotten much better, but is not fully gone. sad the government banned frontier models, i was looking forward to playing with them
>>
ceding the future to china: a dumbfucks guide
>>
>>109112572
Every new iteration does worse compared to the previous. You can clearly see on claude sonnet as example. Starting from 3.5 onwards each new version of the model has higher % of hallucinations.
This is caused by synthetic data ending up in training datasets. Output of LLMs as training input. Habsburg AI problem is not solved and it is getting worse.
>>
>>109112572
it upsets the investors. so they just lie about it.
>>
>>109114142
That seems fine to me. If both travel at least 20 meters, and the force applied to both is the same, the golf ball will almost certainly continue much further.
>>
Humans hallucinate more often.
>>
>>109116839
They are both thrown 20 meters, not to travel "at least" 20 meters. It doesn't ask which one might potentially continue further later, it asks specifically about the distance traveled BEFORE touching the ground. The question doesn't talk anything about force being the same, you are hallucinating.
>>
>>109117300
It doesn't say they only travel 20 meters, seems like your the one hallucinating keeek
>>
>>109112572
it's probably by design and on purpose, they'd rather knowingly give you false information than admiting they don't fucking know. billions are "trusted" into the hands of these people
>>
>>109115030
it improved google search
its a good tool to cut down research time
>>
>>109117463
Oh, my bad, it doesn't specify distance traveled at all.
With the potential ambiguity of the language, you just have to make the closest approximation with what you have. With that, a 20m throw certainly doesn't lead to the thing flying 25m away before touching the ground. Whether it continues further afterwards is irrelevant to the question asked and you are trying to divert the original problem with your smartass hallucinations.
>>
>SOME GUY ON TWITTER
Shut the fuck up, it doesn't matter. Link a study. You are so fucking stupid it's exhausting
>>
Nobody talks about AI because the hype is over. It came, we all had a good fun time, nothing changed. Prior art: segway, roomba, virtual reality.
>>
>>109112572
Because people are tired of repeating the same arguments over and over again?
>>
>>109112572
Because they think they don't exist anymore.
Yet they fall for them every day
>>
>>109114668
ding ding ding at first ai grifters were hoping it's a solvable problem, now noone talks about it (muh solved), don't mention to ppl paying big money for your one hand bandit solution provider that they pay for all the misses as well, just proomot better/skill issue/newer model just werks
>>
>>109117685
I can't even register on twitter anymore without their app.
I wanted to be part of the onions but they wont let me
>>
>>109112572
mo is such a retard sometimes, of course they don't mention it, why keep mentioning the fundamental problem with a product you're trying to market as few trilly in worth
>>
>>109118158
junior coder that works 24/7 but keeps lying is not worth all that much
>>
>>109112572
funny thing is now you can add 'hallucinations' to your thread and it works like the magic N word in dissuading bots from engaging
>>
>>109113594
>You need Mythos to know thursday has a d in it
>>
>>109118054
Robot vacuums are alright.
>>
>>109118459
you need mythos to find a weekday without d in it
>>
>>109112572
nobody talks about it but it is 100% clearly still a major fucking problem.
the other day i googled one of the world cup games and it told me the result of the game, but the game hadn't even begun
>>
>>109112572
>Why does nobody talk about AI hallucinations anymore?
you can fuck a dead horse only for so long or something like that
>>
>>109113545
This confuses the American
>>
>>109112572
Because they're inherent to how LLMs work, they can't be removed. All cargo cultists can do is close their eyes, cover their ears and pretend they're not there.
>>
>>109119142
stop noticing emperor's lack of clothes
>>
>>109118480
Not possible because you're getting the d every day of the week.
>>
>>109118478
Alright at uploading maps of my house, yeah. What do you mean it's also supposed to clean the floor?
>>
>>109112572
>psychosis was a gpt-4 era problem
there's more retards than ever dating their AI, using it as a therapist, and generally just leaning on it and trusting it for everything, it's just swept under the rug now
>>
>>109119378
thank you mythos
>>
>>109119456
wait, what do you mean, all the sources of lawsuits are being suppressed when magical snake oil gets ipo get suppressed? Crazy world
>>
>>109112572
gemini never hallucinates!
>>
File: 1759821875867362.mp4 (1.39 MB, 1246x720)
1.39 MB
1.39 MB MP4
AI Chads won. Hallucinations are just creativity.
>>
>>109119577
as long as you don't ask it about strawberries or walking to a gas station or...
>>
>>109119599
The first one isn't even valid, you would know if you weren't retarded.
>>
>>109119606
hallucinating results of simple character addition is not valid, we fixed hallucinations ai bros, forget trying to railguard with python, rlhf until raspberries hit etc etc it's just not valid, let AI make world changi decisions, just valid ones, simple
>>
>>109119606
our tool is dumb enough we can't fix for letter counting, but target decisions for killing iranian schoolgirls, top fucking notch
>>
>>109119642
the evil cunts behind killing iranian schoolgirls will jist pretend: we were following anthropic orders, this shit is ready for real life, fuck you cunts
>>
>>109119680
and this is why every murrican deserves to be killed on sight, just following capitalist orders, not my fault thisndumb bitch can't count letters ok? Good enough to bomb schoolchilren
>>
>>109112572
look, ok, maybe hallucinations killed some kids, but muh IPO
>>
>>109119625
>>109119642
People who don't work in tech should be prevented from posting on /g/, yall are too fucking cringe.
>>
>>109119820
yeah, fags like you should stay in linkedin to shill latest electronic jeet
>>
>>109119828
you ludlost
>>
>>109119589
When can I get a sexy AI robot like this, and will there be a fleshlight attachment available?
>>
>>109118459
>The USGovt has banned the world from using Mythos in case they find out that 'Wednesday' has a 'd' in it.
>>
>>109112572
It still hallucinates all the time, it's just 4chan doesn't think its funny anymore, joke got old
>>
File: image.png (113 KB, 968x646)
113 KB PNG
>>109113658
My bot has no issues with questions like that.
Web interfaces are making the models more stupid. They inhibit tool calls and reasoning effort and are in no way a good benchmark of what the models are capable of.
>>
File: 1750896839539237.png (160 KB, 277x500)
160 KB PNG
>>109113545
>>109119116
??? americans start their week on sunday?
>>
>>109113705
>/g/cel suddenly cares for the "average user"
>>
>>109113594
>sir you must to be doing the upgrading and postpone the rerunning of prompt called as "d"
>>
>>109112572
>just 2 more weeks saar agi is around the corner my autocomplete better than ur autocomplete pls give me another trillion
die
>>
>>109124373
Imagine the most retarded shit possible, Americans will probably be doing it
>>
>>109112572
it's a well known issue. whats the point of constantly talking about it?
>>
>>109112572
>why did my talking point fail
I don't know, why don't you ask about r's in strawberry a few more times.
>>
>>109112658
I used to work for a retail company whose work week started on Sunday. There was some internal rule that nobody could be made to work both weekend days on a week's rota, so to sidestep its own rule they started the week on a Sunday so they could have you do the Saturday one week and Sunday the next, ie doing the full weekend via pilpul.
>>
>>109117499
>it improved google search
Not only is this not true, but it's turned a cheap to operate search engine into an economy warping disaster.
>>
>>109116727
https://www.theguardian.com/technology/2026/apr/22/ai-hallucinations-found-in-high-profile-wall-street-law-firm-filing

lmao
>>
Do you never make mistakes in your work? Do you never forget trivial pieces of information?
>>
>>109126573
there's a difference between forgetting and making some crazy shit up after taking a fuckton of drugs
>>
>>109126573
I've never unintentionally made up a source
>>
>>109126592
>>109126601
I accept your concession
>>
>>109120719
>falling for /biz/ trolls
>>
>>109112658
You retard
>>
>>109126573
>>109126622
>no u
>haha I stopped listening
I wanna get paid to do this shit too, where do I sign up?
>>
>>109124384
Like it or not, most services are geared toward the average user.
>>
OpenAI put out a paper saying that hallucinations are an inherent problem with AI and will never go away. I don't know why people still pretend AI is infallible when the makers of it themselves say that isnt the case.
>>
File: Clipboard01.png (41 KB, 404x332)
41 KB PNG
>>109113000
>>109113545
What confuses me most is that weekend is a loanword in most languages yet english countries don't have an actual weekend. They have a end-of-week-and-also-start-of-week...
>>
>>109113545
Did you actually "do your research" on this?
Jews only have Saturday as a definite weekend day, a second weekend day can be either Friday or Sunday (I think it's Friday in the Zionist entity).
Muslims only have Friday as a definite weekend day, so a second can be either Thursday or Saturday (or non-contiguous Sunday lol, hello Lebanon). From what I hear it was usually Thursday, but they switched to Saturday because of considerations like "You can't have a branch of an international company/bank unavailable in any of FOUR different weekdays". The new symmetry with zios rubbed some the wrong way, lol.
Supposed Christians have Sunday as the definite, so theoretically, Monday could have been a second. But I guess Jews and "special" Christians made Saturday a natural choice.
Don't know about Hindus, Chinks, ...etc.
>>
>>109112572
why keep bring it up when nothing's changed?
>>
>>109112572
People understand the limitations of LLMs better now and use them more appropriately.
>>
>>109112572
>they're not talking about us anymore... that must mean we won!
or it means everything has been said and people have moved on
>>
>>109122576
pretty much, everyone already knows that they hallucinate and you can only post the "how many r's in strawberry" pictures so many times before everyone has seen it and aren't surprised/entertained by it anymore
>>
>>109112572
>it is a class of intelligence
until it is self-orienting and self-improving, it isn't
>>
>>109112572
they fixed it for the only thing worth complaining about, programming

ask gpt 5.5 about 40k lore and it will hallucinate often times on the first prompt and you can't even correct it



[Advertise on 4chan]

Delete Post: [File Only] Style:
[Disable Mobile View / Use Desktop Site]

[Enable Mobile View / Use Mobile Site]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.