• Echo Dot@feddit.uk
    link
    fedilink
    English
    arrow-up
    10
    arrow-down
    1
    ·
    edit-2
    4 months ago

    I managed to get an AI to build pong in assembly. Are are pretty cool things, but not sci-fi level just yet, but I didn’t just say “build pong in assembly”, I have to hand hold it a little bit. You need to be a programmer to understand how to guide the AI to do the task.

    That was something very simple, I doubt that you can get it to do more complex tasks without a more lot of back and forth.

    To give you an example I had a hard time getting it to understand that the ball needed to bounce off at an angle if intercepted at an angle, it just kept snapping it to 90° increments. I couldn’t fix it myself because I don’t really know assembly well enough to really get into the weeds with it so I was sort of stuck until I was finally able to get the AI to do what I wanted it to. I sort of understood what the problem was, there was a number somewhere in the system and it needed to make the number negative, but it just kept setting the number to a value. A non-programmer wouldn’t really understand that’s what the problem was and so they wouldn’t be able to explain to the AI how to fix it.

    I believe AI is going to become an unimaginably useful tool in the future and we probably don’t really yet understand how useful it’s going to be. But unless they actually make AGI it isn’t going to replace programmers.

    If they do make AGI all bets are off it will probably go build a Dyson Sphere or something at that point and we will have no way of understanding what it’s doing.

    • VantaBrandon@lemmy.world
      link
      fedilink
      English
      arrow-up
      6
      arrow-down
      1
      ·
      4 months ago

      I tried to get it to build a game of checkers, spent an entire day on it, in the end I could have built the thing myself. Each iteration got slightly worse, and each fix broke more than it corrected.

      AI can generate an “almost-checkers” game nearly perfectly every time, but once you start getting into more complex rules like double jumping it just shits the bed.

      What these headlines fail to capture is that AI is exceptionally good at bite sized pre-defined tasks at scale, and that is the game changer. Its still very far from being capable of building an entire app on its own. That feels more like 5-10 years out.

      • William@lemmy.world
        link
        fedilink
        English
        arrow-up
        1
        ·
        4 months ago

        I haven’t tried to scaffold whole projects, but otherwise that lines up with my usage of AI copilots so far.

        At this point, they’re good at helping you interface with the built in commands and maybe some specific APIs, but it won’t do your thinking for you. It just removes the need for some specific base-level knowledge.

    • Hroderic@lemmy.world
      link
      fedilink
      English
      arrow-up
      5
      ·
      4 months ago

      Yeah, I don’t see AI replacing any developers working on an existing, moderately complex codebase. It can help speed up some tasks, but it’s far from being able to take a requirement and turn it into code that edits the right places and doesn’t break everything.