In case you missed it, ChatGPT 5.1 had a tendency to talk about “goblins” in its responses. Supposedly this was a result of training a “nerdy” personality, but it bled into the model as a whole. Because the training run for the latest model already had this flaw, they had to add specific instructions to the system prompt for their Codex coding tool to avoid this behaviour.

Here’s the full prompt from their github. In fact, they repeated the goblin instructions twice, cos you know that will definitely fix it. It’s an interesting read if you consider each one of these instructions were meant to prevent some undesired behaviour: https://paste.sh/Iev3HtMe#JZ4dw_CkvJcpVmjjoy7WZnSn

More info here: https://news.northeastern.edu/2026/05/06/chatgpt-goblins-problem-ai-behavior/

OpenAI’s own blog post casually explaining why they couldn’t predict that their state of the art model would obsess about goblins: https://openai.com/index/where-the-goblins-came-from/

  • sudo@programming.dev
    584·
    6 months ago

    I still can’t get over how the only fine tuning you can do for an LLM is yell at it with markdown files. We should be able to retrain local models so they can develop an actual experience without prefilling the context.

    • theunknownmuncher@lemmy.world
      343·
      6 months ago

      I still can’t get over how the only fine tuning you can do for an LLM is yell at it with markdown files.

      It isn’t.

      We should be able to retrain local models so they can develop an actual experience without prefilling the context.

      Great news, you can do exactly that.

          • Ziglin (it/they)@lemmy.worldEnglish
            41·
            6 months ago

            Windows 11 isn’t running in the cloud yet though. Unless it checks to make sure it hasn’t been tampered with too much you should just be able to modify some of its binaries (the source code obviously isn’t available). With the cloud based llms that is not possible.

            If you have a model on your computer you can retrain it, which is like changing a binary just far less precise. The option of having a source code equivalent just isn’t there beyond having the same dataset and seeds for the training program.

            So I’d say it is worse than your average run of the mill proprietary software.

        • Bazoogle@lemmy.world
          31·
          6 months ago

          Nope, it does the same thing:

          Pi’s minimal system prompt and extensibility let you do actual context engineering. Control what goes into the context window and how it’s managed.

          AGENTS.md: Project instructions loaded at startup from ~/.pi/agent/, parent directories, and the current directory.

          SYSTEM.md: Replace or append to the default system prompt per-project.

        • Eager Eagle@lemmy.worldEnglish
          3·
          6 months ago

          but you can, as long as it’s open weight. Fine tuning and training are pretty much the same process

          • thingsiplay@lemmy.ml
            41·
            6 months ago

            That still falls into the category “creator” to me, if you need to rebuild. I was making the distinction to an end user, comparable to applications that you download and use and configure. Instead of rebuilding the source code with your modifications.

            Do I misunderstand here something? Or is this a communication issue caused by different interpretations?

            • howrar@lemmy.ca
              51·
              6 months ago

              If you define “user” to be a set that excludes anyone capable of modifying the weights, then by definition, no user can modify the weights.

              Any criticism about users being unable to modify weights becomes vacuous, so it’s not an interpretation that makes sense.

              • thingsiplay@lemmy.ml
                21·
                6 months ago

                I wasn’t criticizing at all. Just tried to define what I mean by creator and user. You was takling about “how do you think LLMs are trained” and I told you that the user was probably not thinking of who trains the LLMs, or fine tune them as you said. And yes, fine tuning the open weight falls into creation process, as they are rebuild. That is not the same as an end user who downloads the final usable product. And yes, it makes sense.

                • Eager Eagle@lemmy.worldEnglish
                  1·
                  6 months ago

                  the original comment says “We should be able to retrain local models so they can develop an actual experience without prefilling the context.” - it turns out we can. Not sure why you’re trying to attach labels of user vs creator, when the premise already mentions retraining.

    • thingsiplay@lemmy.ml
      6·
      6 months ago

      I always thought it’s just ghosts or maybe aliens. Never thought that demons are the real ones.

  • rizzothesmall@sh.itjust.works
    16·
    6 months ago

    Who’d have thought that OpenAI would overfit with known faulty pretrains when the community as a whole are well aware not to do this…

  • Caveman@lemmy.world
    14·
    6 months ago

    I recently added some stuff to my agents.md file so it’s more fun.

    • Warning/issue - > goblin
    • Error - > Orc
    • Exception - > attack

    Open to more suggestions. It make reading the output more fun. Claude is so shit now that it doesn’t work. Also, if you guys haven’t tried caveman mode, it’s great.

  • Gsus4@mander.xyz
    3·
    6 months ago

    Just ask it what the Helvetica scenario is. Funny and terrifying at the same time.