• FlashMobOfOne@lemmy.world
      link
      fedilink
      arrow-up
      9
      arrow-down
      1
      ·
      22 hours ago

      Think of it like making a xerox copy of a xerox copy. The copy of the copy is always shittier.

      Using synthetic data can escalate model collapse, as a model is only as good as its training data, which is partly why these LLM models “hallucinate”, having been trained on a wealth of garbage from Reddit.

      • Nouvellalia@lemmy.world
        link
        fedilink
        arrow-up
        5
        ·
        18 hours ago

        I was going to make a snarky comment about

        “How’s it thinking then?! I was trained by adults and so forth back generations!”

        Then I looked at the quality of people trained by other humans instead of nature, comparing myself to my ancestors, looking at the society around me.

      • HaraldvonBlauzahn@feddit.org
        link
        fedilink
        arrow-up
        3
        ·
        21 hours ago

        So that means when I publish a nifty FOSS project, I should alongside publish one hundred copies which contain stealthy BS LLM modifications which introduce subtly wrong code (like failing invariants or undefined ehavior in concurrent C++ code)? And all dated back to 2020?

        Got it!