• totallynotfbi@lemm.ee
    link
    fedilink
    English
    arrow-up
    0
    ·
    1 year ago

    The images are too compressed, so I can’t really make out what they say. I’m guessing that EA finally updated their outdated Denuvo implementation, making it much tougher to crack now

    • Kissaki@feddit.de
      link
      fedilink
      English
      arrow-up
      0
      ·
      1 year ago

      Current AI is not smarter than humans. It needs supervised training, and then acts according to that. That’s inherently incompatible to novelty and correct exploration.

      • Send_me_nude_girls@feddit.de
        link
        fedilink
        English
        arrow-up
        0
        ·
        edit-2
        1 year ago

        AI is good in doing complex things but bad at doing easy things. Supervision is required at first for learning of course, there’s no AI that works out of the box.

        • Kissaki@feddit.de
          link
          fedilink
          English
          arrow-up
          0
          ·
          1 year ago

          That assessment entirely depends on what you consider “complex” and “easy”.

          What do you mean by it’s bad at doing easy things but good at doing complex things? I don’t see how something complex would work better than something easy.

          • Send_me_nude_girls@feddit.de
            link
            fedilink
            English
            arrow-up
            0
            ·
            edit-2
            1 year ago

            In short.

            Look up what AI does good right now, like finding complex solutions to mathematical issues a human couldn’t. Calculate stuff very fast, replicate natural language etc.

            Look up what AI struggles with at the moment, like drawing hands or recognizing objects or driving a car.

            This statement is only valid in this current state, as AI is advancing faster than most peoples mind by now. Most people have yet to understand LLM or generative AI models.

            That’s what I’m talking about. If you look at the process required to crack Denuvo, then you’ll notice that there’s a lot of guesswork done, something the AI is good at if learned properly. The amount of people who know how to and are willing to spend time cracking Denuvo is shrinking by the day. The amount of software DRM encrypted is rising every day. We need automation soon.

            AI will soon be mandatory for software security as malicious actors will use AI to find zero day exploits and you want an AI to protect you from those real time threats. Anti Virus software already work somewhat into that direction by now but there’s still much room.

      • B3_CHAD@lemmy.dbzer0.comOP
        link
        fedilink
        English
        arrow-up
        0
        ·
        1 year ago

        No, any sufficiently advanced A.I can and will outclass humans. For example: there are chess A.I’s that have beaten GM’s as good as Magnus Carlsen on multiple occasions. The better an A.I gets at something the tougher it becomes to counter it. This is one of the biggest risks of A.I development that one day we might make something that makes us seem obsolete. On the positive side that day is really really far.

        • Kissaki@feddit.de
          link
          fedilink
          English
          arrow-up
          0
          ·
          1 year ago

          Where between being beaten in chess and making us obsolete do you think we are? Where do you think cracking games is between chess and human mind?

          I think your estimation is off by a lot.

        • Auli@lemmy.ca
          link
          fedilink
          English
          arrow-up
          0
          ·
          1 year ago

          I don’t know if AI is technically better it’s just different and doesn’t play like a human. Humans hate lossing pieces but AI doesn’t care as long as the outcome is a win.

          • Send_me_nude_girls@feddit.de
            link
            fedilink
            English
            arrow-up
            0
            ·
            edit-2
            1 year ago

            AI absolutely plays like a human as it’s trained by humans. The only difference is, AI will do the most optimal move, while humans might hesitate. That’s also the reason why it’s bad to put AI into fighter or bomber jets. The AI has a clear goal but a human might struggle to fire at an unknown target. Because the human has to life with the consequences.

        • reddithalation@sopuli.xyz
          link
          fedilink
          English
          arrow-up
          0
          ·
          1 year ago

          how so? ignoring mathematically unbreakable things like encryption, given enough time, i think pretty much anything could be reverse engineered and cracked, its just a matter of how much time it would take

            • reddithalation@sopuli.xyz
              link
              fedilink
              English
              arrow-up
              0
              ·
              1 year ago

              yeah but your cpu has to run the unencrypted game, and so i would think its currently impossible to have a local, 100% uncrackable game

              • Send_me_nude_girls@feddit.de
                link
                fedilink
                English
                arrow-up
                0
                ·
                1 year ago

                DRM already only does check for validity every other frame or even minute. There’s no use in a game that just closes because it recognized a violation. You do know what causes Denuvo fps spikes? It’s whenever it checks. Of course the software got better by now so it’s less of an issue but it’s still there.

          • Send_me_nude_girls@feddit.de
            link
            fedilink
            English
            arrow-up
            0
            ·
            1 year ago

            Look up RSA algorithms and project that to other mathematically complex DRM protections. You’re wrong because you don’t understand the tech.