• Buckshot@programming.dev
    link
    fedilink
    English
    arrow-up
    20
    arrow-down
    2
    ·
    2 days ago

    How about we don’t include bioweapons in the training data. Then it can’t regurgitate it.

        • Cocodapuf@lemmy.world
          link
          fedilink
          English
          arrow-up
          6
          ·
          21 hours ago

          Doubt all you want. Was the LLM trained with information about toxic or harmful chemicals? Yes. Was the LLM trained with information about weapons? It’s practically unavoidable.

          The LLM is perfectly capable of mashing two concepts together, that’s essentially all it does. So will its concoctions work? We’ll find out.

          • DreamlandLividity@lemmy.world
            link
            fedilink
            English
            arrow-up
            1
            ·
            edit-2
            8 hours ago

            We’ll find out.

            We won’t. The companies don’t sort their training data to an extent where they could reliably exclude something. So they won’t. They will keep adding bypassable filters on top and pretend it’s not their problem people bypass it.

        • BrilliantantTurd4361@sh.itjust.works
          link
          fedilink
          English
          arrow-up
          3
          arrow-down
          1
          ·
          1 day ago

          If humans have the language to express something, an llm can arrive at that conclusion. That doesnt mean each one will be valid or technically feasible but by virtue of how they work they can arrive at a recipe for such things. It is the reason why they are “good” at programming and proofs, it can flow probabilistically from the corpus of data.

          • DreamlandLividity@lemmy.world
            link
            fedilink
            English
            arrow-up
            4
            arrow-down
            2
            ·
            edit-2
            24 hours ago

            Well, ok. But by that logic, so can a monkey smashing keys on a typewriter for long enough. If the LLM is not trained on something, it does not see the connections between the words. So it can generate text that looks like chemistry/bioengineering but chances are it would not be a weapon, just random whatever.