Comment on New report illuminates why OpenAI board said Altman “was not consistently candid”

• PsychedSy@sh.itjust.works ⁨11⁩ ⁨months⁩ ago

  By answering questions? We are general intelligences that can answer questions. Oh shit oh fuck what am I doing talking.

  source

  • photonic_sorcerer@lemmy.dbzer0.com ⁨11⁩ ⁨months⁩ ago

    Hey guess what, we general intelligences are capable of terrible things.

    source
  • xor@lemmy.blahaj.zone ⁨11⁩ ⁨months⁩ ago

    Okay, so let’s do a thought experiment, and take off all the safeguards.

    Oops, you made:

    • a bomb design generator
    • an involuntary pornography generator
    • a CP generator

    Saying “don’t misuse it” isn’t enough to stop people misusing it

    And that’s just with chatgpt - AI isn’t just a question and answer machine - I suggest you read about “the paperclip maximiser” as a very good example of how misalignment of general purpose AI can go horribly wrong

    source

    • elbarto777@lemmy.world ⁨11⁩ ⁨months⁩ ago

      I was going to say that a well-determined individual would find this information regardless. But the difference here is that it being so easily accessible would increase the risks of someone doing something reaaaally stupid by a factor of 100. Yikes.

      source
    • PersnickityPenguin@lemm.ee ⁨11⁩ ⁨months⁩ ago

      The last 2 already exist it’s called stable diffusion. And for awhile Bing did it too.

      source
    • Socsa@sh.itjust.works ⁨11⁩ ⁨months⁩ ago

      ChatGPT was very far from the first publically available generative AI. It didn’t even do images at first.

      Also, there are plenty of YouTube channels which show you how to make all sorts of extremely dangerous explosives already.

      source

      • xor@lemmy.blahaj.zone ⁨11⁩ ⁨months⁩ ago

        But the concern isn’t which was the first generative ai - their “idea” was that AIs - of all types, including generalised - should just be released as-is, with no further safeguards.

        That doesn’t consider that OpenAI doesn’t only develop text generation AIs. Generalised AI can do horrifying things, even just by accidental misconfiguration (see the paperclip optimiser example).

        But even a GANN like chatGPT can be coerced to generate non-text data with the right prompting.

        Even in that example, one can’t just dig up those sorts of videos without, at minimum, leaving a trail. But an unresticted pretrained model can be distributed and run locally, and used without trace to generate any content whatsoever that it’s capable of generating.

        And with a generalised AI, the only constraint to the prompt “kill everybody except me” becomes available compute.

        source