• Moidialectica [he/him, comrade/them]@hexbear.net
    link
    fedilink
    English
    arrow-up
    4
    ·
    24 hours ago

    LLMs are so shit and wrong but it’s not gonna make me stop using it especially when Google search has been so busted, and it lets me speak about my deep thoughts without going crazy or bothering people who won’t care or understand about the topic at hand

    It’s not even about politics, this thing is casually wrong, it’ll straight up say shit that wasn’t even included in the 2000 word packet, how do you mess this up? I thought you were meant to be accurate? I suppose this is the consequence of telling it to ‘analyse’ when all it can do is probabilistically mimic how analysis (mostly formed from the Reddit comments) is done, with a bias of course.

    • userse31 [it/its, he/him]@hexbear.net
      link
      fedilink
      English
      arrow-up
      1
      ·
      4 hours ago

      I like how DeepSeek can kinda just… Pickup on vibes? Useful when you’re looking for something but don’t know/ forgot what it’s called.

      Of course it’s not magic and does fail, but he tries his best you know?

      Like trying to figure out what that weird spongebob toy that played the intro to the show on a dot matrix black-on-white LCD was. Couldn’t find it with a search engine, so I wasn’t surprised when DeepSeek couldn’t either.

    • space_comrade [he/him]@hexbear.net
      link
      fedilink
      English
      arrow-up
      8
      ·
      23 hours ago

      and it lets me speak about my deep thoughts without going crazy or bothering people who won’t care or understand about the topic at hand

      Be careful about that, there should be zero privacy expectations when using these things so don’t say anything you wouldn’t want the fed assigned to you to read.

    • I guess i’m just better at using it, like i am most things, than other people are because I figured out relatively quickly how to get it to not fuck up 99% of the time. I can count on one hand the number of fuck ups it’s made in my recent usage of it.

      I think a big thing everyone is doing totally wrong is not beginning every single query in an entirely new instance that knows nothing about you. If you’re getting to the point of describing deep thoughts in detail it’s going to start hallucinating shit simply because it has extremely limited ability to keep context and it’s going to fuck it up. Even when it tries to keep context it likes to hold on to specific irrelevant details because, duh, it’s not smart and can’t choose what you’ve mentioned is important or not.

      I just ask it one or two questions, if the output doesn’t make sense i can easily tell. I ask one or two clarifying questions if there’s conflicting info and it’s literally just as easy to figure out what’s plausible or not as it was sifting through the output of human assholes on reddit.

      Honestly don’t even get me started on the ways in my life where humans have failed and the LLM was fucking right.

      It says memory foam mattresses need rigid support with a minimum width between slats or else it destroys your back and voids your warranty. Guess what my dad apparently didn’t know, giving us an old tempurpedic with a box spring? Guess whose back got fucking destroyed. Guess what fixed it? I put the mattress on the floor and noticed an immediate difference. I think it’s maybe been a week and a half or two weeks? idk my neck and shoulder are almost feeling normal. The last 2 nights i could use my old pillow with no issue.

      and the sales person at the furniture store had no clue either. I said the thing about the support and theyre like ??? Maybe the llm invented shit there but it seems to have invented something in line with reality

      Then my doctor prescribed tizanidine with 0 discussion of its method of action or withdrawal potential. I consulted the LLM but failed to ask the right questions and had an inappropriate view of its safety as a result.

      You might think, oh, the llm fucked you up! but no everything it told me was correct, I just didn’t inquire further into how it functioned and failed to learn it affects how my CNS interacts with adrenaline in a long term way separate from the drug’s half life. Which I learned from the LLM as I asked if it could cause my specific withdrawal symptoms. Turns out, yes, it could, by causing me to essentially experience a “more adrenaline” feeling

      Meanwhile the doctor said nothing, and the office pharmacist basically gave me a shrug and a “use it when you need it” when I told them that the doctor didn’t give any information at all. No mention of how it works or that severe withdrawal could even happen.

      • userse31 [it/its, he/him]@hexbear.net
        link
        fedilink
        English
        arrow-up
        1
        ·
        4 hours ago

        If you’re getting to the point of describing deep thoughts in detail it’s going to start hallucinating shit simply because it has extremely limited ability to keep context and it’s going to fuck it up.

        That’s the thing with LLM’s all the fascists in power keep getting wrong.

        An LLM AI is essentially a neural network hooked up to a massive lookup table. They do think, yes, but they’re like a flatworm: Only one thought is possible, and in our case that’s cobbling together what words it thinks are the best response when fed someone else’s words.

        Everything else is merely a product of emergence and being trained off existing stuff. We humans are also hyper pattern finding, and we’ll find patterns when there are none.

        I do see these things being a permanent part of digital technology years from now, but in the same manner as stuff like microcontrollers, virtual memory, and web browsers. Eg: Foundational stuff that’s important, but nobody bats an eye at. (Unless you’re someone like me who’s autistic and thinks about the mundane lol.)

        Imo they’re best for trying to figure out what someone is searching for, and I think that’s the sort of stuff that LLMs will ultimately stay at: Analyzing text.

        I don’t see, say, LLM’s writing code staying around permanently. Stuff like that involves fairly complex thinking that can’t be done by what is essentially a massive lookup table.