• Margot Robbie@lemmy.world
    link
    fedilink
    English
    arrow-up
    12
    arrow-down
    2
    ·
    1 year ago

    Breakthrough technological development usually can be described as a sigmoid function (s-shaped curve), while there is an exponential progress in the beginning, it usually hit a climax then slow down and plateau until the next breakthrough.

    There are certain problem that are not possible to resolve with the current level of technology for which development progress has slowed to a crawl, such as level 5 autonomous driving (by the way, better public transport is a way less complex solution.), and I think we are hitting the limit of what far transformer based generative AI can do since training has become more and more expensive for smaller and smaller gains, whereas hallucination seems to be an inherent problem that is ultimately unfixable with the current level of technology.

    • cyberpunk_sunbear@lemmy.zip
      link
      fedilink
      English
      arrow-up
      5
      arrow-down
      2
      ·
      1 year ago

      One thing that I think makes AI a possibility to deviate from that S model is that it can be honed against itself to magnify improvements. The better it gets the better the next gen can get.

      • vrighter@discuss.tchncs.de
        link
        fedilink
        English
        arrow-up
        9
        ·
        1 year ago

        that is a studied, documented, surefire way to very quickly destroy your model. It just does not work that way. If you train an llm on the output of another llm (or itself) it will implode.

        • barsoap
          link
          fedilink
          English
          arrow-up
          3
          ·
          1 year ago

          Also at best it’s an refinement, not a new sigmoid. So are new hardware/software designs for even faster dot products or advancements in network topology within the current framework. T3 networks would be a new sigmoid but so far all we know is why our stuff fundamentally doesn’t scale to the realm of AGI, and the wider industry (and even much of AI research going on in practice) absolutely doesn’t care as there’s still refinements to be had on the current sigmoid.