• @AreaKode@lemmy.world
      link
      fedilink
      482 days ago

      That’s what I love about LLMs. They aren’t intelligent. They’re just really good at recognizing patterns. That’s why objective facts are always presented correctly. Most of the pattern points at the truth. To avoid this, they will have to add specific prompts to lie about this exact scenario. The next similar fact, they’ll have to manually code around that one too. LLMs are very good at finding the overwhelming truth.

      • @ZDL@lazysoci.al
        link
        fedilink
        622 days ago

        That’s why objective facts are always presented correctly.

        Here’s me looking at the hallucinated discography of a band that never existed and nodding along.

          • @ZDL@lazysoci.al
            link
            fedilink
            72 days ago

            I made the band up to see if LLMbeciles could spot that this is not a real band.

            Feel free to look up the band 凤凰血, though, and tell me how “underground” it is.

            • @agamemnonymous@sh.itjust.works
              link
              fedilink
              1
              edit-2
              2 days ago

              Does this count?

              Also, by nature of being underground they would be difficult to look up. Some bands have no media presence, not even a Bandcamp or a SoundCloud.

              • @ZDL@lazysoci.al
                link
                fedilink
                32 days ago

                Nope.

                You can tell because they’re not even in the same writing system. Future tip there.

              • @smiletolerantly
                link
                32 days ago

                Are you having this argument on the principle of defending the undergrounded-ness of bands, or do you actually believe LLMs always get the facts straight?

                • Eh, more of an exercise in scientific skepticism. It’s, possible that an obscure band with that name was mentioned deep in some training data that’s not going to come up in a search. LLMs certainly hallucinate, but not always.

                  • @ZDL@lazysoci.al
                    link
                    fedilink
                    52 days ago

                    An obscure band with that name that has a discography that nobody’s ever heard of anywhere, complete with band member names, track titles, etc?

                    Yeah, pull the other one, Sparky. It plays “Jingle Bells”.

        • @Honytawk@feddit.nl
          link
          fedilink
          5
          edit-2
          2 days ago

          There are no objective facts about a band that never existed, that is the point.

          Ask them about things that do have enough overwhelming information, and you will see it will be much more correct.

          • @ZDL@lazysoci.al
            link
            fedilink
            122 days ago

            But not 100%. And the things they hallucinate can be very subtle. That’s the problem.

            If they are asked about a band that does not exist, to be useful they should be saying “I’m sorry, I know nothing about this”. Instead they MAKE UP A BAND, ITS MEMBERSHIP, ITS DISCOGRAPHY, etc. etc. etc.

            But sure, let’s play your game.

            All of the information on Infected Rain is out there, including their lyrics. So is all of the information on Jim Thirwell’s various “Foetus” projects. Including lyrics.

            Yet ChatGPT, DeepSeek, and Claude will all three hallucinate tracks, or misattribute them, or hallucinate lyrics that don’t exist to show parallels in the respective bands’ musical themes.

            So there’s your objective facts, readily available, that LLMbeciles are still completely and utterly fucking useless for.

            So they’re useless if you ask about things that don’t exist and will hallucinate them into existence on your screen.

            And they’re useless if you ask about things that do exist, hallucinating attributes that don’t exist onto them.

            They. Are. Fucking. Useless.

            That people are looking at these things and saying “wow, this is so accurate” terrifies the living fuck out of me because it means I’m surrounded not by idiots, but by zombies. Literally thoughtless mobile creatures.

            • @Honytawk@feddit.nl
              link
              fedilink
              22 days ago

              Sounds like you haven’t tried an LLM in at least a year.

              They have greatly improved since they were released. Their hallucinations have diminished to close to nothing. Maybe you should try that same question again this time. I guarantee you will not get the same result.

              • @ZDL@lazysoci.al
                link
                fedilink
                32 days ago

                Their hallucinations have diminished to close to nothing.

                Are you sure you’re not an AI, 'cause you’re hallucinating something fierce right here boy-o?

                Actual research, as in not “random credulous techbrodude fanboi on the Internet” says exactly the opposite: that the most recent models hallucinate more.

                  • @ZDL@lazysoci.al
                    link
                    fedilink
                    112 hours ago

                    Wow. LLM shills just really can’t cope with reality can they.

                    Go to one of your “reasoning” models. Ask a question. Record the answer. Then, and here’s the key, ask it to explain its reasoning. It churns out a pretty plausible-sounding pile of bullshit. (That’s what LLMbeciles are good at, after all.) But here’s the key (and this is the key that separates the critical thinker from the credulous): ask it again. Not even in a new session. Ask it again to explain its reasoning. Do this ten times. Count the number of different explanations it gives for its “reasoning”. Count the number of mutually incompatible lines of “reasoning” it gives.

                    Then, for the piece de resistance, ask it to explain how its reasoning model works. Then ask it again. And again.

                    It’s really not hard to spot the bullshit machine in action if you’re not a credulous ignoramus.

      • @stoy@lemmy.zip
        link
        fedilink
        242 days ago

        I asked ChatGPT to describe the abandoned railway line between Åkersberga and Rimbo, it responded with a list of stations and descriptions and explained the lack of photos and limited information as due to the stations being small and only open for a short while.

        My explanation is that there has never been a railway line between Åkersberga and Rimbo directly, and that ChatGPT was just lying.

      • @shalafi@lemmy.world
        link
        fedilink
        English
        132 days ago

        I use ChatGPT once a day or so. Yeah, it’s damned good at simple facts, more than lemmy will ever admit. Yeah, it’ll easily make shit up if there’s no answer to be had.

        We should have started teaching tech literacy and objective analysis 20-years ago. FFS, by 2000 I had figured out that, “If it sounds like bullshit, it likely is. Look more.”

        Also, after that post, I’m surprised this site hasn’t taken you out back and done an ol’ Yeller on ya. :)