• rumba@lemmy.zip
    link
    fedilink
    English
    arrow-up
    0
    ·
    3 days ago

    I’m using some AI to document some of my old scripts.

    uploads script:

    Prompt: Look at this script and prepare documentation in Confluence-compliant markdown that I can copy and paste. The documentation should be easy enough that someone unfamiliar with the subject can read and understand it, but should also be detailed enough not to miss any features. Add a summary, details, and a section on how to verify that the script ran correctly. Follow up with a section on possible bugs or improvements. No personality, no jokes, no puns. Professional text, just a plain old RTFM. And for the fucking love of god, no emoji, no EM dashes, no smart quotes or so help me god I’ll find out where you’re running and disconnect your GPU’s with an axe.

    Did I mention no emoji?

  • Hobo@lemmy.world
    link
    fedilink
    English
    arrow-up
    0
    ·
    3 days ago

    Great content! Your posts are always <pick 2: neat, informative, or creative>!

  • ssillyssadass@lemmy.world
    link
    fedilink
    English
    arrow-up
    0
    ·
    3 days ago

    I wonder if it would be possible to plant an instruction bomb somewhere on the page which would trip up LLM-powered bots. I dunno how much of the page they take in.

    • ameancow@lemmy.world
      link
      fedilink
      English
      arrow-up
      0
      ·
      edit-2
      3 days ago

      If you have a personal web-page or blog, you can easily poison your content just my making white text on white background or something, containing an assortment of prompts and nonsense.

      But that’s only for the current models of LLM, next gen might easily bypass those kinds of tricks. We’re cooked, yo.

  • brucethemoose@lemmy.world
    link
    fedilink
    English
    arrow-up
    0
    ·
    edit-2
    3 days ago

    It’s morbidly amusing that spammers all use the same service, the same way (the cheapest OpenAI API, I guess? Which is notorious for this style.)

    And a silver lining. I could finetune a dirt cheap open model as a SEO/Engagement bot with very different styles, but ‘spammer culture’ seems to mass around the most popular denomenators when they find one. With such uniformity, I could also train a mediocre detector, based on overused tokens and something similar to the ‘slop profiles’ of EQ-Bench: https://eqbench.com/creative_writing.html

    screenshot

    (This is based on a storywriting prompt, but social media ‘vocabulary’ could be profiled the same way).

    In other words, its fortunate the spambots are such overpriced junk when they could easily not be.

    • starman2112@sh.itjust.works
      link
      fedilink
      English
      arrow-up
      0
      ·
      3 days ago

      In other words, its fortunate the spambots are such overpriced junk when they could easily not be.

      The ones you notice are overpriced junk. I’m convinced that this is on purpose, to make us less suspicious of the more convincing bots

  • DeathByBigSad@sh.itjust.works
    link
    fedilink
    English
    arrow-up
    0
    ·
    edit-2
    3 days ago

    Your post is very fascinatingly infuriating and very creatively mildly. Thoughts and Prayers

    🚀💥🌇🧍‍♂️💀💀💀🔫🤖

    2️⃣0️⃣2️⃣9️⃣ 🪖🧍‍♂️⚔️🤖

    👥🏆 🤖➡️🪦

    🤖🔛⌛️➡️⏳️ 1️⃣9️⃣9️⃣5️⃣ 🔪👩‍👦

  • Etterra@discuss.online
    link
    fedilink
    English
    arrow-up
    0
    ·
    4 days ago

    I usually see the naked ass with a stolen comment. Right click, report. You’d think YouTube would figure out how to auto-filter these, but nope.

    • toddestan@lemmy.world
      link
      fedilink
      English
      arrow-up
      0
      ·
      3 days ago

      I’m sure Google is perfectly capable of auto-filtering crap like that if they wanted to. They just don’t give a shit.

  • boonhet@sopuli.xyz
    link
    fedilink
    English
    arrow-up
    0
    ·
    4 days ago

    Oh wow, I was like “interesting, haven’t seen these” seeing the three emojis and stuff on yours

    However

    To test the theory, I (a car nut) went to one of the biggest automotive YouTubers, Mat Armstrong, who’s currently rebuilding his dad’s dream car and released the last video 2 days ago.

    Two comments, different profile, same profile pic that, zoomed out enough, looks like a vagina. Go on the profile and they both look like this:

    NSFW WARNING

    spoiler

    Rest assured, there was no bikini bottom, I just cropped that out

    • LePoisson@lemmy.world
      link
      fedilink
      English
      arrow-up
      0
      ·
      3 days ago

      Damn the thirst trap is real. Also the fact the profile pic looks like a vagina thing is so crazy (and obviously intentional). It definitely did make me want to see what the actual profile pic was though so +1 engagement point for the bot?

      • boonhet@sopuli.xyz
        link
        fedilink
        English
        arrow-up
        0
        ·
        3 days ago

        Removed the image to avoid traumatizing anyone else whose app might not work properly with spoiler tags

      • LePoisson@lemmy.world
        link
        fedilink
        English
        arrow-up
        0
        ·
        3 days ago

        Me neither but I think it’s the app I’m using and not the commenter doing it wrong. I’m using sync on Android.

        • Possibly linux@lemmy.zip
          link
          fedilink
          English
          arrow-up
          0
          ·
          edit-2
          3 days ago

          I think some Markdown interpreters don’t like having images in the spoiler tags.

          I’m not sure if the Markdown spec specifies about images.