• simple@lemmy.world
    link
    fedilink
    English
    arrow-up
    1
    arrow-down
    1
    ·
    edit-2
    2 years ago

    I’ve been talking about the potential of the dead internet theory becoming real more than a year ago. With advances in AI it’ll become more and more difficult to tell who’s a real person and who’s just spamming AI stuff. The only giveaway now is that modern text models are pretty bad at talking casually and not deviating from the topic at hand. As soon as these problems get fixed (probably less than a year away)? Boom. The internet will slowly implode.

    Hate to break it to you guys but this isn’t a Reddit problem, this could very much happen in Lemmy too as it gets more popular. Expect difficult captchas every time you post to become the norm these next few years.

    • Rhaedas@kbin.social
      link
      fedilink
      arrow-up
      1
      ·
      2 years ago

      Just wait until the captchas get too hard for the humans, but the AI can figure them out. I’ve seen some real interesting ones lately.

      • OpenStars@kbin.social
        link
        fedilink
        arrow-up
        2
        ·
        2 years ago

        There is considerable overlap between the intelligence of the smartest bears and the dumbest tourists.

      • Boz (he/him)@lemmy.one
        link
        fedilink
        arrow-up
        2
        ·
        2 years ago

        I’ve already had to switch from the visual ones to the audio ones. Like… how much of a car has to be in the little box? Does the pole count as part of the traffic light?? What even is that microscopic gray blur in the corner??? [/cries in reading glasses]

        • BarbecueCowboy@kbin.social
          link
          fedilink
          arrow-up
          1
          ·
          2 years ago

          ChatGPT isn’t really as smart as a lot of us think it is. What it excels at really is just formatting data in a way that is similar to what you’d expect from a human knowledgeable in the subject. That is an amazing step forward in terms of language modeling, but when you get right down to it, it basically grabs the first google search result and wraps it up all fancy. It only seems good at deductive reasoning if the data it happens to fetch is good at deductive reasoning.