• Int_not_found@feddit.de
    link
    fedilink
    arrow-up
    29
    ·
    edit-2
    10 months ago

    As a Data Scientist I can say: The present danger from AI isn’t the the singularity. That’s science fiction. It’s the lack of comprehension what an AI is & the push to involve it more and more into certain decision making processes.

    Current AIs are at there cores just statistical models, that assign probabilities to answers, based on previously observed data.

    Governments and cooperations around the globe try to use these models to automate decisions. One massive problem here is the lack of transparency and human bias in the data.

    For example, when a cooperation uses an AI to determine who should be fired. You get fired, you try to complain, but you just get the answer the maschine had a wide variety of input data & you should have worked harder.

    We experienced in the past, that AIs focus on things we don’t necessarily want them to focus on. In the example from above maybe your job performance was better then your colleges Dave, but you are a PoC and Dave is white. In the past PoCs were more prone to get fired, so the AI decided, that you are the most probable answer to the question ‘Who should we fire?’.

    If a human would have made the decision you could interview him and discover the underlying racism in this decision. Deciphering the decision of an AI is next to impossible.

    So we slowly take away our ability to address wrongs in our burocratic processes, by cementing them into statistical models & thereby removing our ability to improve our societal values. AI has the potential to grind societies progress to a halt & drag easily fixable problems decades or centuries into the future.

    • PsychedSy@sh.itjust.works
      link
      fedilink
      arrow-up
      11
      ·
      10 months ago

      So you eliminate race as a possible input. Now it finds proxies. Non-standard name? Address? What holidays you take off? Maternity leave gaps can signal parenthood. Patterns of time off/FMLA can align with treatments. It’s hard to choose relevant inputs without choosing revealing inputs.

    • abraxas@lemmy.ml
      link
      fedilink
      English
      arrow-up
      9
      ·
      edit-2
      10 months ago

      The real-real issue is how many people don’t bloody understand what AI/ML are, but are making huge decisions about where they are appropriate to use.

      I can’t count how many times I’ve heard “Let’s add AI to this page!” was requested from non-tech execs in the last year, not knowing what does or doesn’t work. Our most successful analytical report runs a simple 10-rule heuristic and nobody is the wiser.

      So yeah, people trying to inject AI into hiring/firing. The people who did inject AI into predicting criminality. It all boils down to negligence, ignorance of your tools.

    • Facebones
      link
      fedilink
      arrow-up
      7
      ·
      10 months ago

      Exactly. The “black box” nature of these systems should be a red flag for any practical usages.

  • gerryflap@feddit.nl
    link
    fedilink
    arrow-up
    23
    ·
    10 months ago

    Who says it ends here? We’ve made tremendous progress in a short time. 10 years ago it was absolutely unthinkable that we’d be at a stage right now where we can generate these amazing images from text on consumer hardware and that AI can write text in a way that could totally fool humans. Even as someone working in the field I was fairly sceptical 5-6 years ago that we’d get here this fast

    • teichflamme@lemm.ee
      link
      fedilink
      arrow-up
      7
      ·
      10 months ago

      Agree 100%.

      We are at the start and the progress is incredibly fast and accelerating.

      Even the way that image generation alone has improved within the last year I wouldn’t have believed.

      • gerryflap@feddit.nl
        link
        fedilink
        arrow-up
        2
        ·
        10 months ago

        Yeah also indeed. Back then I was actually working with image generation and GANs and it was just starting to take off. A year later or something StyleGAN would absolutely blow my mind. Generating realistic 1024x1024 images while I was still bumbling about with a measly 64x64 pixels. But I didn’t foresee where this was going even then

        • newIdentity@sh.itjust.works
          link
          fedilink
          arrow-up
          1
          ·
          edit-2
          10 months ago

          Or remember when everyone was impressed with GauGAN in 2019/2020. We would’ve never guessed that just 2 or 3 years later we would have multiple competing available models.

          Or when Dall-E mini had a hype last year and everyone was impressed with it?

          Now there even are first experiments that make videos out of a prompt and they pretty much look exactly like the earlier iterations of diffusion models

    • abraxas@lemmy.ml
      link
      fedilink
      English
      arrow-up
      1
      ·
      edit-2
      10 months ago

      Where some underrepresented domains have made massive strides, this GPT thing has done relatively little for data science.

      The important thing is that “what AI/ML can and cannot do” is not changing that much. It’s successful application is what’s changed. The idea of making AI libraries more accessible is huge, and leads to stuff like this. But under the hood, OpenAI doesn’t do much different than other AI tools. It’s just easier to use yourself. You can do more faster as computers get faster, but that seems to be limited with the endgame of Moore’s Law anyway.

      OpenAI runs on supercomputers now. It’ll continue to run on supercomputers in the future. Instead of getting better, it has started to get worse at many things. Experts have always had a fairly good grasp of where it’ll end. There are things AI was always expected to do better than humans at. And things it never will.

      I mean, I expected AI image generation and better text quality. But I also expected the limits it currently has. And I’ve only done a little directly in the field.

      • gerryflap@feddit.nl
        link
        fedilink
        arrow-up
        1
        ·
        edit-2
        10 months ago

        But the fact that it can do so much is an awesome (and maybe scary) result in and of itself. These LLMs can write working code examples, write convincing stories, give advice, solve simple problems quite reliably, etc all from just learning to predict the next word. I feel like people are moving the goalpost way too quickly, focussing so much on the mistakes it makes instead of the impressive feats that have been achieved. Having AI doing all this was simply unthinkable a few years ago. And yes, OpenAI is currently using a lot of hardware, and ChatGPT might indeed have gotten worse. But none of that changes what has been achieved and how impressive it is.

        Maybe it’s because of all these overhyping clickbait articles that make reality seem disappointing. As someone in the field who’s always been cynical about what would be possible, I just can’t be anything else then impressed with the rate of progress. I was wrong with my predictions 5 years ago, and who knows where we’ll go next.

  • BetaDoggo_@lemmy.world
    link
    fedilink
    arrow-up
    5
    ·
    10 months ago

    The issue is the marketing. If they only marketed language models for the things they are able to be trusted with, summarization, cleaning text, writing assistance, entertainment, etc. there wouldn’t be nearly as much debate.

    The creators of the image generation models have done a much better job of this, partially because the limitations can be seen visually, rather than requiring a fact check on every generation. They also aren’t claiming that they’re going to revolutionize all of scociety, which helps.