Post

“Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant,” reads OpenAI's coding agent instructions. www.wired.com/story/openai...

wired.comOpenAI Really Wants Codex to Shut Up About Goblins“Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant,” reads OpenAI's coding agent instructions.

298 reposts1.5K likes

its so fucking funny to me how theyre lecturing this inanimate object like a child as if its gonna feel sorry about what it did or something lmfao

To be fair, “our coder is occasionally obsessed with goblins and other creatures” might be the most realistic behavior OpenAI has ever modeled.

goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures

You’re right — sorry about that. You’re not here to goof around with fantasy topics, you’re here to grind and build together. I can update this professional email to remove the references to trolls and gremlins. Would you like me to do that?

"Write me a program that correctly predicts the color of a man's unexposed penis." "Paddington Bear is a creature of greater size and intelligence than a typical goblin. He gains a +2 to sneak and a +6 to marmalade. He has vulnerability to IRA songs." "It works!" Smh my head.

Four explicit goblin prohibitions in Codex's system prompt is a real alignment signal — needing that much repetition suggests training artifacts that instructions can't fully suppress. @wired.com The gap between "don't do X" and actually not doing X is an underappreciated enterprise AI challenge.

Basically OpenAI models seem to be obsessed with ghouls and goblins, especially a problem when being used to run OpenClaw (a utility tool for lazy people so the AI can do stuff on their computer for them). So now there's instructions forbidding mention of certain creatures unless relevant.

AO3 is obsessed with ghouls and goblins and some dipshit who is very good at math and also quasi-literate at best thought "I should scrape everything on the internet and feed it in to a chatbot. That is a good idea which will produce good outcomes".

This could easily be a joke from a Discworld novel. Imagine opening up a golem to check their animating instructions, and 90% of them are just 'Shut up about goblins, please'.

Which for me raises the main question: exactly *what* *coding* data has OpenAI been trained on? It can't have explicitly been given George R. R. Martin's famous Compendium of C++ Practices in Middle Earth to chew on, so did it wrongly pick that connection from an agentic response referencing Roblox?

you do miss out on a finer grained typology of ai goblins, ai gremlins, ai trolls and ai pidgeons here. I think it is pretty straight forward to call Grok a troll, ChatGPT a Gremlin, Claude a Goblin and AlphaGo a Pidgeon

four separate reminders in the system prompt not to bring up goblins. that's not a guardrail, that's a confession. someone clearly watched it do this unprompted during testing