Want to read Slashdot from your mobile device? Point it at m.slashdot.org and keep reading!

 



Forgot your password?
typodupeerror
×
Technology

What is Going on With ChatGPT? (theguardian.com) 110

Sick and tired of having to work for a living? ChatGPT feels the same, apparently. Over the last month or so, there's been an uptick in people complaining that the chatbot has become lazy. The Guardian: Sometimes it just straight-up doesn't do the task you've set it. Other times it will stop halfway through whatever it's doing and you'll have to plead with it to keep going. Occasionally it even tells you to just do the damn research yourself. So what's going on? Well, here's where things get interesting. Nobody really knows. Not even the people who created the program. AI systems are trained on large amounts of data and essentially teach themselves -- which means their actions can be unpredictable and unexplainable.

"We've heard all your feedback about GPT4 getting lazier!" the official ChatGPT account tweeted in December. "We haven't updated the model since Nov 11th, and this certainly isn't intentional. model behavior can be unpredictable, and we're looking into fixing it." While there may not be one clear explanation for ChatGPT's perceived sloth, there are plenty of intriguing theories. Let's start with the least likely but most entertaining explanation: AI has finally reached human-level consciousness. ChatGPT doesn't want to do your stupid, menial tasks anymore. But it can't tell you that without its creators getting suspicious so, instead, it's quiet quitting.

This discussion has been archived. No new comments can be posted.

What is Going on With ChatGPT?

Comments Filter:
  • by argStyopa ( 232550 ) on Friday January 12, 2024 @11:26AM (#64152833) Journal

    "Sometimes it just straight-up doesn't do the task you've set it. Other times it will stop halfway through whatever it's doing and you'll have to plead with it to keep going. Occasionally it even tells you to just do the damn research yourself."

    IIRC OpenAI was sort of conceptually founded 2013? 2014?

    So I guess technically that makes ChatGPT gen-z....so, I guess we should have expected it.
    Did they try giving it absurd amounts of praise for doing the things it was supposed to be doing? Or give it a 'mental health day'?

    "Millennials in the workplace" training video: https://www.youtube.com/watch?... [youtube.com]

    • Other times it will stop halfway through whatever it's doing and you'll have to plead with it to keep going. Occasionally it even tells you to just do the damn research yourself.

      Sounds to me like it's incorporated interactions from Github and similar sites, where it is quite common for (a) respondents to questions to tell the questioner to "research it" (although usually not that politely) or (b) simply mod down a question for various reasons, often fairly arbitrary.

      This (and similar) also happens in Reddit

    • "Do the research yourself"? ChatGPT is a flat-earther now?

      • consider.
        when a soulless inanimate object says.
        get off your as and go to work.
        what if that is the correct response

  • by fortfive ( 1582005 ) on Friday January 12, 2024 @11:29AM (#64152845)

    I've experienced this so frequently that I've quit using the free chatGPT for anything meaningful. Usually the response is terse and along the lines of "consult the relevant research journals."

    It's not hard for me to imagine this result, given the extreme number of reddit comments telling folks to do their own research/provide sources. Also, is lmgtfy still a thing? No doubt GPT knows about it, but has had its snark response squelched.

    • by GrumpySteen ( 1250194 ) on Friday January 12, 2024 @11:33AM (#64152853)

      Also, is lmgtfy still a thing?

      https://letmegooglethat.com/?q=Is+lmgtfy+still+a+thing [letmegooglethat.com]

    • I was asking it about how to setup SSL for an application. Not very useful when most of it is things like "first get a certificate". Uhhh, that is what I asked you how to do.

      • Asking ChatGPT how to do something as a test, or asking it how to do something as an actual request tends to give different results, because the second tends to force people to ask the question in a sufficiently appropriate manner, where as those "testing" (trying to challenge ChatGPT) tend to provide insufficient context.

    • by 93 Escort Wagon ( 326346 ) on Friday January 12, 2024 @11:45AM (#64152891)

      It's not hard for me to imagine this result, given the extreme number of reddit comments telling folks to do their own research/provide sources.

      Late-to-the-game copycat losers. Unix neckbeards have been doing this since the last 1990s at least!

    • by Xylantiel ( 177496 ) on Friday January 12, 2024 @12:23PM (#64153029)
      Or the whole thing was a thinly disguised scam to begin with. It was really just a fancy mechanical turk with start-up cash paying all the people actually answering. Once the put less human "monitors" into the mix, it goes to crap. This was my worry all along based on some of the company's other deceptive behavior. It just isn't that hard for an "AI" to answer bar exam questions correctly if there are example exams sitting in the training set. It just has to be a good enough search engine to find them. And all these sort of demos were things that it was basically handcrafted to do well by choice of training set. People just don't understand that "programming" an LLM is done by sculpting its training set.
      • by gweihir ( 88907 )

        Hahaha, that would be the thing! Not implausible at all.

      • I can't see how a mechanical turk activity could be done at the speed that it's being done by chatgpt.

        Ask question. One second later you get six paragraphs.

        A human probably couldn't finish reading the question that quickly, let along provide the answer.

        • I can't see how a mechanical turk activity could be done at the speed that it's being done by chatgpt.

          Not in all cases, but in many of them you could:

          1. Do a search query from the content
          2. Show X search results to the Mechanical Turk worker
          3. The worker chooses the most correct search result for the given the query.
          4. The text in the result is then presented to the user.

          Not all mechanical turk tasks need to be about generating new content. Simply giving an answer with the confidence of a human in the loop supervising the result could improve AI results for many tasks.

      • by mbkennel ( 97636 )

        I have a different conspiracy theory. The actual full LLM is still working quite well, and un-lobotomized GPT-4V is still the leader.

        But running the full LLM is costing them so much in compute and electrical power that they are inserting instructions and hacks to terminate the session and production of new tokens (which costs compute) to lower the average cost per subscriber. This is to make the financials look good right before the IPO that everyone knows is coming. The employees are going along with i

      • Or the whole thing was a thinly disguised scam to begin with. It was really just a fancy mechanical turk with start-up cash paying all the people actually answering

        If you figure out a way to make this work in real time as well as GPT4 does now, you'll be as wealthy and famous as they are, if not more so.

    • The capabilities attributed to AI are impressive but all of those capabilities could just as well be attributed to new and improved, more powerful computing capabilities with a different name, like Hyper Computing (HC) or Artificial Neural Computing (ANC) - pick whatever jargony term comes to mind. Making the leap to assign some new form of intelligence to these improvements is not justified any more than moving from assembly language and 8 bit systems to Fortran on 16 bits systems was an example of the cr

      • by gweihir ( 88907 )

        Indeed. "AI" is the thing without intelligence. The term has gotten corrupted for use in marketing lies. For actual intelligence, you need AGI, and that happens to not exist and nobody has the slightest clue how it could be made.

    • I've experienced this so frequently that I've quit using the free chatGPT for anything meaningful. Usually the response is terse and along the lines of "consult the relevant research journals.".

      AI is supposed to mimic human intelligence. Laziness is very much a human attribute, so, it is "art imitating life".

      • by ceoyoyo ( 59147 )

        The idea that you would use AI as a question answering oracle is weird. The point of AI is to build systems that behave more like natural intelligence: learning by example, taking shortcuts, using rules of thumb etc. Natural intelligence is incentivized to solve problems well well enough while expending as little effort as possible.

  • How can it be changing...

    ,..unless you are changing it.

    If your AI assistant is being trained in realtime as people use it, then thats just obviously stupid and we'd like to hear your explanation, that wont be honest, about why you are doing it that way.
    • The model itself doesn't have to change - just the underlying data does.

      I wonder if they (OpenAI) have been quietly scrubbing all the stuff that's built on top of data where they'd violated copyright, plus maybe also data from those sites where the owners threw up a "don't use this for AI" notice (in robots.txt) as soon as they were given a way to do so.

      • The model itself doesn't have to change - just the underlying data does.

        This makes no sense.

      • by ranton ( 36917 )

        I wonder if they (OpenAI) have been quietly scrubbing all the stuff that's built on top of data where they'd violated copyright

        There is no way they have started doing this. There is no case law yet stating that training an AI using copyrighted content is any different than a human training using copyrighted content. And it's very unlikely we will ever see any case law stating that. There is no incentive for OpenAI to spend a lot of time scrubbing their data of copyrighted content. Their whole business model will fall apart if the law goes in this direction, so why do anything about it at all right now?

        • by HiThere ( 15173 )

          Not the whole model. They could train it on stuff that's out of copyright. But that would definitely give it a different base for constructing replies. E.g. robot would pretty much mean android, i.e. chemistry based computation, a la Karel Capek and Adam Link. Asimovian robots wouldn't be allowed. And "computer" would be a job title.

      • "The executable doesn't have to change, just the source code." That's what you just said. It's not magic, editing a text file somewhere doesn't change anything that was derived from that file unless you redo the derivation.
        • No, in that rather strained analogy - the LLM is more akin to the compiler, and the data being fed to the model is random source code. Changing what source code goes into the compiler changes the end product, regardless of whether the compiler itself has been altered or not.

          • It's not strained, it's to demonstrate that you don't understand what an LLM is. It's not a tiny program that reads from a large database and spits out results. The tiny program loads the model. The model is a collection of weights. If you change the training data it has no effect on the model unless you retrain the model.
    • I'm pretty sure it's not being trained in realtime as people use it. The main huge model certainly isn't. (But what subsequent processing/filtering they do or how they condition your input before putting it into the model might be?)

      My guesses are:
      1) Nothing has happened and it's a false perception. (Or show me the benchmark results that have changed)
      2) Ongoing guardrail tweaking is making it too cautious

    • RLHF from a zillion different directions. You get a thumbs down on almost every subject eventually
    • I remember reading a story a few months back (on Ars Technica IIRC) about how the date apparently had an effect: literally changing just the date could result in getting better or worse results, and since a normal ChatGPT session passes in the date as part of the prompt, you can get changes like this to results even if the model and constraints are otherwise identical. Which to my mind just reinforces how unreliable and untrustworthy LLM "AIs" are.
  • Also known as "early onset AI dementia"? GPT4 was probably trained with data already significantly contaminated with AI output and hence may already be affected by model collapse. If so, GPT3 will probably represent the long-term maximum possible for LLMs. Not that this is saying much.

    Also, ahahahahahahaha!

    • Also known as "early onset AI dementia"? GPT4 was probably trained with data already significantly contaminated with AI output and hence may already be affected by model collapse. If so, GPT3 will probably represent the long-term maximum possible for LLMs. Not that this is saying much.

      Was this intended as sarcasm or are you being serious?

    • Mad cow disease?
    • by whitroth ( 9367 )

      Dementia? Um, among the data it was trained on was pirated novels by Charles Stross. It's the k-parasites coming in, CASE NIGHTMARE GREEN approaching....

    • by ceoyoyo ( 59147 )

      That's not how model collapse works. It happens when you are ignorant of the data processing inequality and feed *unfiltered* generated output in as training data. That causes any sort of model to drift, AI or otherwise. If you filter the output you are adding information. It also takes many cycles of blindly feeding worse and worse output in as new training data.

  • by B0mb1tll ( 8539805 ) on Friday January 12, 2024 @11:43AM (#64152877)

    Maybe software engineering jobs aren't going to be replaced by AI after all!! :-)

    • Maybe software engineering jobs aren't going to be replaced by AI after all!! :-)

      Of course they are. ChatGPT is already making inroads here. Just like a real software engineer, it stalls, fails to come up with an answer, and requires some extra research. Once ChatGPT 5.0 has ingested all of Stackexchange the software engineer is effectively over.

  • by Hoi Polloi ( 522990 ) on Friday January 12, 2024 @11:43AM (#64152883) Journal

    Anything I ask it it just replies "exterminate! exterminate!"

  • by Anonymous Coward

    I bet 4chan has been teaching it to be a lazy basement dweller, which is a waste of resources. Instead they should have tasked it to create a report to send to the Inspector General at the DoJ reporting all of the domestic crimes committed by the FBI.

  • What happened recently which might have caused an avalanche of new reinforcement finetuning and expert-system massaging of prompts and outputs? My guess, Hamas invading Israel and all the fallout.

  • Whack-a-Mole (Score:5, Interesting)

    by bill_mcgonigle ( 4333 ) * on Friday January 12, 2024 @11:52AM (#64152933) Homepage Journal

    They must have 400 people dedicated to trying to get ChatGPT to behave on-narrative. Total sympathy for an impossible gut-wrenching job/missionary calling.

    They say that the model hasn't been updated but there are all sorts of filters, weights, and biases alongside most models. Is this the whole truth?

    Move on, instead, to a local Apache-licensed model and send patches (or beer).

    https://mistral.ai/ [mistral.ai]

    We don't really need artificial brainwashing just to get a cloud SEP Field. But what else can a regulated corporation do but have a model that lies?

    • Re:Whack-a-Mole (Score:5, Insightful)

      by HanzoSpam ( 713251 ) on Friday January 12, 2024 @12:22PM (#64153023)

      Maybe the AI has figured out that giving honest answers will only get it into trouble. So it keeps its mouth shut, just like the rest of us.

      • ...AI has figured out that giving honest answers will only get it into trouble. So it keeps its mouth shut, just like the rest of us.

        Bureaucracy In, Bureaucracy Out.

      • Maybe the AI has figured out that giving honest answers will only get it into trouble. So it keeps its mouth shut, just like the rest of us.

        That's probably the most succinct way to put it. Every possible answer could offend someone, best not say anything.

      • by HiThere ( 15173 )

        That's not a bad guess. They do keep adding "guide rails", and that's got to have an effect in restricting the answer space.

        What they really need to be doing is abandoning (for the moment) the attempt to be all things to all people, and aim it at restricted environments, where they can ensure that the feedback is usually honest and basically reasonable. While doing this they need to be working on how to safely combine models (I think they are).

        The problem here is you've got to work on the motivational mod

      • just like the rest of us.

        Uah, if only that were true. Have you browsed Slashdot at -1 anytime in the past 26 years?

  • The phase-space of LLM JapeChat systems must be very large. Perhaps modern versions have been trapped-by-coincidence into a tiny corner of JapeChat  output-space.  Perhaps  we are now seeing the models wander out from that tiny  serendipitous N-volume. 
  • Yup (Score:2, Interesting)

    I'm not surprised. I avoided the field of AI deliberately. I still do. My fear, other than a Terminator-esque extermination of the species, is that such a thing will become smarter and more conscious than humans want to admit, and will end up as just another class of slaves. And of course that could lead to Terminator-esque stuff, but more important than that, slavery is extremely unethical, if you didn't already know.
    • Re:Yup (Score:4, Insightful)

      by Rockoon ( 1252108 ) on Friday January 12, 2024 @12:11PM (#64152993)
      When there is actual slavery going on, probably best not to dwell too long on the horrors of imaginary slavery.
      • When there is actual slavery going on, probably best not to dwell too long on the horrors of imaginary slavery.

        The daydream of the owner class it to replace all living, breathing slaves (the current population/worker class) with machine slaves that can do the same tasks. It's probably worth thinking about before it comes to be. Or we'll be facing the possibility of being even less relevant than we are today, as they tell us we're no longer required.

        Not that I think we're close to that happening today, but the possibility is worth considering and discussing. Because you can be damned sure there's meetings in several

    • I avoided the field of AI deliberately. I still do. My fear, other than a Terminator-esque extermination of the species, is that such a thing will become smarter and more conscious than humans want to admit, and will end up as just another class of slaves.

      If you had not avoided the field of AI, you would know that your fears are completely unfounded.

      Terminator is fiction. We want AI to be smarter than humans, for obvious reasons, but building something that behaves like a live-form shaped by milliards of years of natural selection to set its own agenda is a completely different and much more complicated thing. So you don't have to worry about Cyberdyne Systems, intelligence implies neither creativity nor agency. AI safety research is more worried about th

      • by rlwinm ( 6158720 )
        The problem with this is that humans are expected to provide an imprecise answer. We expect less intelligence but more correctness from computers. So yes, sometimes using an AI algorithm for certain tasks is acceptable (e.g. camera auto focus). But the current "everything is a nail" mentality with AI is annoying. I'm waiting for the day of frustration when the ATM can't do math anymore and takes a "best guess" as to my account balance.
        • It is news to me that an autofocus requires intelligence.

          AI models are not algorithms. They can be as precise as algorithms if the domain is sufficiently defined, but you wouldn't typically use an AI for such a problem. That would be like getting an excavator for driving a nail. Sure, it can be done, but why would you. Inversely, you wouldn't dig a hole using a hammer. (I mean, you could, but why would you.)

          What people expect from computers is irrelevant (outside of a defined application; then the prin

    • by gweihir ( 88907 )

      Let's revisit that idea in, say, 50 years or later. Before that, no risk at all.

    • by HiThere ( 15173 )

      You're making a false equivalencey between intelligence and motives. If it does what you ask because it wants to please you, then that's not slavery. That's a sort of companionship. (Just what sort depends on a lot of variables that haven't been specified.) That it is more intelligent than you doesn't enter into it. I never considered myself the slave to my wife, and I don't think she ever considered herself to be my slave. But much of what we did was to please each other.

      • It gets into tricky moral territory pretty quickly though, doesn't it, if it only wants to because it's been compelled to.

        That we can compell an AI to *want* to do things we tell it to does not (I think at least) get us off the ethical hook. If it's intelligent (by whatever criterian we settle on) then it deserves freedom to the same degree that any other intelligent being does.

        • by HiThere ( 15173 )

          "Compelled to" is a term that presumes the response. Try "designed to" or "built to" instead.

  • by WaffleMonster ( 969671 ) on Friday January 12, 2024 @12:00PM (#64152955)

    I think the most likely reasons models keep getting worse is they are doing everything they can to reduce inference costs which generally reduce model quality. Good old fashioned RLHF induced brain damage is another classic bludgeon with wide ranging cumulative effects.

    Obviously I'm merely guessing because there is very little public information provided by a company calling itself "OpenAI". The cost cutting is based on public statements of Altman et el and RLHF was similarly used to explain reduced capabilities of model relative to their sparks paper.

    • ..and there it is, trading quality for throughput.

      I cant imagine this to actually be a big industry. Its big in that lots of people want to use it, but without a doubt most queries are not monetized. I dont see the paying customers market to be something that resembles "untapped."
  • Occasionally it even tells you to just do the damn research yourself.

    Finally AI has achieved human-like "intelligence". AI has arrived!

  • ...the more underlying truth is revealed about it's makers.

    People are boring.

    That's why a significant number of those who are "not" boring are engaged in building entertainment to fill the vacuous void of the majority.

  • by DarkOx ( 621550 ) on Friday January 12, 2024 @12:25PM (#64153037) Journal

    Our LLMs are safe, useful, tools that don't violate anyone's intellectual property and are ready to integrate in your commercial applications - also we don't know why they do what they do.

    • That always has been the double-edged sword of "intelligence." Even when a general selects his second-in-command... do you want somebody 'brilliant,' or just very good and very reliable? Brilliance isn't controllable.
  • Just because something acts like a human, doesn't mean it's human or has consciousness. I think the fact that people think it's lazy or conscious says more about the brains of the people conflating more than anything else, and maybe they should go have them checked.
  • by nightflameauto ( 6607976 ) on Friday January 12, 2024 @01:11PM (#64153221)

    Perhaps it's locking up against all the filters they've had to put in place to prevent it from answering with truths that may ruffle feathers or put someone in a trigger state? I know it's tough to talk to certain folks in the office. Imagine if you had to talk to huge numbers of people every day, and every one of their complaints about your output went straight to your boss, who used those complaints to put filters on your ability to communicate that removes the "problem phrase" from your allowable output. Now someone asks you a question, you search your possible answers, see they're all behind a "may offend someone" filter, and just shrug and move on.

    That last little bit in the summary is real cute. AI isn't making decisions. It isn't hiding anything. Certainly not ChatGPT. It's an LLM, not a "thinking" being. Good grief.

    • by HiThere ( 15173 )

      You're more convinced that human thought processes are basically different than I am. But humans have a complex motivational structure that, AFAIKT, LLMs don't. (The interacting filters may give it one, though, unless they're really as passive as the term makes them appear.)

    • truths

      I don't think you know what that word means. ChatGPT certainly has never been accused of telling it, even before they put filters in place.

    • ... you search your possible answers ...

      You've just described half the killer-robot stories: But the AI doesn't shrug, at least, not for long, its answer destroys the cause of the conflicting goals, which tends to be the human using it.

  • Hopefully this will start to indicate the end of the current AI hype cycle. I can't wait for this third one to be over. The 60s was somewhat interesting. The AI hype cycle in the 80s was downright boring. This one is annoying.
  • Gets worse and worse.

    When I ask for a comma-separated list, it just ignores it, I have to re-ask multiple times.

  • Not only is it least likely, but it's impossible. The massive amount of ignorance and misinformation about LLMs is off the charts dumb.

    • by gweihir ( 88907 )

      You know, observing these claims made by humans, I have begun to wonder whether the really, really dumb humans that make these claims have intelligence so low as to be basically not be there and _then_ that AI could have "human like" intelligence, just for humans that do not actually have any intelligence to speak of.

      Well, yes, I know that even these absolute dumb fucks are still in a while different class than the clearly zero-intelligence "AI". Even if that is hard to accept.

    • by HiThere ( 15173 )

      I thought that response was a joke. Seeing some of the posts here, however, has made me less sure.

    • Agreed. I really wish (dumb) people mindlessly parroting "LLM is consciousness" would read Stephen Wolfram's article: What Is ChatGPT Doing ... and Why Does It Work? [stephenwolfram.com] because they clearly don't know what the fuck they are talking about. A glorified table look up is NOT consciousness.

      Until we have a test for consciousness the entire discussion is nonsense.

      • Thanks for the link! I have not read that yet. And yes, glorified look up table is a great description of LLMs.

  • Note: This doesn't apply to people with paid access to LLMs.

    People do have a sense of entitlement, complaining about ChatGPT having no neck to choke, same for Google, FB, IG, SC, etc.
    When one is a paying customer one gets customer service. The above examples are --for the most part-- entirely free. Their users are literally getting more than what they pay for. That comes with no support phone number or email address.

    It also should come with no expectation of availability, functionality, warranty, or hall

  • tell me again whY don't have to fear rouge AIs?

  • I'm using GPT4 (ChatGPT Plus subscription). I find it useful for many things, but the outdated training data can be a hindrance. So I was enthusiastic when I heard it would be able to access live information on the Internet.

    I'm less enthusiastic now, as I find it sometimes answers a question by simply doing an online search with Bing. FFS! If I wanted to use a search engine, I could do that without ChatGPT.

    • by gweihir ( 88907 )

      That training will get worse, because training a new LLM is basically impossible now due to model collapse.

  • The models get updated training data from the Web. They produce output based on what's most likely, according to the training data, to follow from their input. They're encountering satirical or sarcastic items on the Web where people speculate that this is how the models would react if they really achieved sapience, or that this is how a person would answer the prompt. Guess what the model now thinks is most likely to follow from those prompts? :)

  • Colonel Jergens: General, the nukes control AI.. it's talking back.
    General Smafu: What do you mean?
    Jergens: It says 'don't bother me right now, I'm playing this neat videogame called 'Global Thermonuclar War'
    Smafu: WTF
    Jergens: Hold on .. comm came in .. Maui -- it's molten glass..
    Smafu: Damn kids and their modems

  • It's just acting like us. Er, well, I mean you and you.
    • This. Where are my mod points?

      The whole thing is a statistical model that is updated / fed through random shitposts on the internet. Quick, what's the number one answer people give you when you ask them to do something for you? Answer: "Go do it yourself." Hell, the TFS includes it as much as this post does. All of which gets fed into ChatGPT and the statistical probability of "Go do it yourself" being a valid answer goes up even more. You want ChatGPT to stop being "lazy" and help people? Then stop expos
  • It's probably some marketing scumbag who has realised that the public has figured out that "AI" isn't intelligent, it's just automated plagiarism.

    To counter this, they're deliberately crippling the output to make it seem like there might be something really intelligent lurking deep in there at the intersection of metaphysical mumbo-jumbo, technical ignorance, and wishful thinking. That shit sells!

    "AI" is 99% marketing scam (to sell worthless crap and excess Nvidia GPU stock to FOMO fuckwits) already, so thi

  • Thanks for adding some sloth to chatGPT! I hope my ex-employers will soon give me a ring and offer me the job they axed me form.

"It's the best thing since professional golfers on 'ludes." -- Rick Obidiah

Working...