Report finds newer inferential models hallucinate nearly half the time while experts warn of unresolved flaws, deliberate deception and a long road to human-level AI reliability
  • mrfriki@lemmy.world
    5·
    7 hours ago

    See, old dynasties had a similar problem because of inbreeding, which probably says a lot of how this models are maintained.

  • mesa@lemmy.worldEnglish
    61·
    23 hours ago

    I have been noticing the non-open source models seem to be doing some strange things with their filters/accuracy. I wonder if they upped their temperature on the advanced models. Its still very much like a art.

    • Rhaedas@fedia.io
      1·
      22 hours ago

      The public models can be adjusted and altered and it’s free market on what works and what doesn’t. So with time I expect them to work better overall, even with the limitations they have with size and changes to make them work on regular systems. The paid ones don’t have the same diversity or automatic culling of what breaks things, so they’re in some sense throwing darts.