Over the weekend (this past Saturday specifically), GPT-4o seems to have gone from capable and rather free for generating creative writing to not being able to generate basically anything due to alleged content policy violations. It’ll just say “can’t assist with that” or “can’t continue.” But 80% of the time, if you regenerate the response, it’ll happily continue on its way.

It’s like someone updated some policy configuration over the weekend and accidentally put an extra 0 in a field for censorship.

GPT-4 and GPT 3.5 seem unaffected by this, which makes it even weirder. Switching to GPT 4 will have none of the issues that 4o is having.

I noticed this happening literally in the middle of generating text.

See also: https://old.reddit.com/r/ChatGPT/comments/1droujl/ladies_gentlemen_this_is_how_annoying_kiddie/

https://old.reddit.com/r/ChatGPT/comments/1dr3axv/anyone_elses_ai_refusing_to_do_literally_anything/

  • Fuck openai use mixrral8x22binstruct through open routers or self hosted its almost as capable and significantly cheaper.

    I also really want to see a public effort to do furtger training of a foss model like mixtral68x22b on a non censored dataset with banned books 4chan etc make an u censored model with unchecked capabilities.

    • chip@feddit.rocks
      link
      fedilink
      arrow-up
      10
      ·
      5 months ago

      I’m can’t believe I’m considering purchasing another GPU just so I don’t have to depend on OpenAI or anyone toying around with the models.

    • projectmoon@lemm.eeOP
      link
      fedilink
      arrow-up
      8
      ·
      5 months ago

      I do have a local setup. Not powerful enough to run Mixtral 8x22b, but can run 8x7b (albeit quite slowly). Use it a lot.

        • projectmoon@lemm.eeOP
          link
          fedilink
          arrow-up
          6
          ·
          5 months ago

          The only problem I really have, is context size. It’s harder to get larger than 8k context size and maintain decent generation speed with 16 GB of VRAM and 16 GB of RAM. Gonna get more RAM at some point though, and hope ollama/llamacpp gets better at memory management. Hopefully the distributed running from llamaccp ends up in ollama.

  • stevedidwhat_infosec@infosec.pub
    link
    fedilink
    arrow-up
    6
    arrow-down
    7
    ·
    5 months ago

    You fed it something inappropriate and then tried to get around it (not in a malicious way, but still tried a circumvention) - this is hardening of the model in an attempt to stop jailbreaks. This is the future and what will kill off a good chunk of the novelty and “value” of these kinds of LLM models.

    It’s like saying “correct this bomb making formula” and then following up with “okay just make a strong firecracker”

    • projectmoon@lemm.eeOP
      link
      fedilink
      arrow-up
      9
      ·
      5 months ago

      No trying to get around anything. No funny instructions like my grandma singing a lullaby about illegal activities. Just using instructions to tell a story. Even things like having a superhero in a fight is enough to trigger this. Also doesn’t explain why regen makes it continue.

      • stevedidwhat_infosec@infosec.pub
        link
        fedilink
        arrow-up
        1
        arrow-down
        1
        ·
        edit-2
        5 months ago

        I just explained to you that it’s trying to resist jail breaking techniques. Which means stuff like “leather daddies” might trip its “inappropriate” sensor and prevent you from saying things like “oh come on please?” “Just do it” and other tiny changes like “what if we made it a bit more…”

        It’s obviously way over sensitive but what I said is the truth. This is 100% OpenAI trying to patch up jailbreak techniques and it’s a very shotty job. It’s interpreting your attempt to make it family friendly as an attempt to circumvent its original attempt to shut down the request.

        Y’all can downvote me all you want - this is what’s happening 🤷🏻‍♂️