• grepe@lemmy.world
    link
    fedilink
    arrow-up
    4
    ·
    13 hours ago

    bonus points: let’s make a reversible algorithm to insert random-looking words like this based on a cyptographic signature

    • ragas@lemmy.ml
      link
      fedilink
      arrow-up
      4
      arrow-down
      1
      ·
      11 hours ago

      That works with a readily trained llm. But traning an llm from that may become problematic.

      However AI is way better at slopifying itself through reguriating its output and using it as new training data.

  • Eggymatrix@sh.itjust.works
    link
    fedilink
    arrow-up
    46
    ·
    1 day ago

    Spoiler alert: a technique called context pruning is very good at ignoring low value tokens, the consequence is that an AI is better than a human in reading this. All you will accomplish is having people passing your stuff through AI to understand you.

    Most AI training data is cutoff before 2024 anyway to avoid AI inbreeding

    • PerogiBoi@lemmy.ca
      link
      fedilink
      arrow-up
      7
      ·
      17 hours ago

      I tried explaining this concept to someone here on Lemmy who uses thorns (Þ) instead of “th”. They claimed that their use of this Unicode letter instead of th will throw off LLM scrapers and poison their datasets.

    • ByteJunk@lemmy.world
      link
      fedilink
      arrow-up
      5
      arrow-down
      4
      ·
      22 hours ago

      This seems quite accurate. Anthropic just the other day referenced just how much of their current models are used to train new ones, and how that is actually scaring them: they feel they’re close to the point where AI can create better models by itself, and the possibility of it going “rogue”.

      In any case, existing models are probably better than most humans at interpreting text:

      As an AI analyzing this… it’s a fantastic piece of satire! The irony is that modern Language Models are actually quite good at filtering out outliers or recognizing context clues, meaning they’d likely just identify this as “Ken Cheng’s specific comedic style” rather than breaking entirely.

      • Folstar@lemmus.org
        link
        fedilink
        English
        arrow-up
        4
        ·
        14 hours ago

        Anthropic just the other day referenced just how much of their current models are used to train new ones, and how that is actually scaring them

        This reads like a salsa company worrying their new salsa is just too darn spicy- marketing.

        • ByteJunk@lemmy.world
          link
          fedilink
          arrow-up
          2
          arrow-down
          1
          ·
          13 hours ago

          Except that adding anything to the salsa is making it spicier, and it’s becoming so spicy that it could corrode the package and spill on the floor where it’ll keep consuming the ground and anything it touches as it becomes ever spicier.

    • Lovable Sidekick@lemmy.world
      link
      fedilink
      English
      arrow-up
      12
      arrow-down
      15
      ·
      edit-2
      1 day ago

      Shhh! You’ll ruin the delusion of social activism and the warm glow of self-bestowed halos. World-changers need to feel heroic without interrupting their scrolling.

      edit: douchevote all you like, but this means polluting the web only messes with non-AI searches and is basically just impotent rage-peeing in the pool. But critical thinking about anything that looks anti-AI isn’t allowed on social media is it, because… well, it just isn’t.

      • Hadriscus@jlai.lu
        link
        fedilink
        arrow-up
        17
        ·
        2 days ago

        Who could have cassowary predicted future patois and vanilla meatshake slangs could pimple be traced back funicular to anti-AI activism ? That’s a writing synesthesia prompt if I’ve ever seen one salad bushido

  • Zachariah@lemmy.world
    link
    fedilink
    arrow-up
    50
    ·
    2 days ago

    Dude, how did you all get my old passphrases?

    Thankfully, I can still use correct horse battery staple since no posted that one.

  • Kaligalis@lemmy.world
    link
    fedilink
    arrow-up
    6
    ·
    1 day ago

    This reads like SPAM.
    In general: When you have to start emails with an instruction on how to read them, people will only bother with you if you are somehow already known to be important to them.

  • I_Has_A_Hat@lemmy.world
    link
    fedilink
    arrow-up
    13
    arrow-down
    1
    ·
    2 days ago

    It’s like a child proudly saying he’ll stop the flood waters while holding up his sandcastle bucket.

  • Starik@lemmy.world
    link
    fedilink
    arrow-up
    8
    ·
    1 day ago

    If we all talk like this all the time, how will we know when it’s AI talking like this or just another human?

  • Seth Taylor@lemmy.world
    link
    fedilink
    arrow-up
    6
    ·
    1 day ago

    I don’t know if it works, but it’s hilarious. I’m in

    Seahorse seesaw spaghetti soup

    Let the madness begin