| I just saw the full system prompt leak for 5.5 (April 23rd release). Most of it is standard agentic stuff, but Instruction #140 is genuinely insane. It explicitly forbids the model from talking about: "goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals." Why the specific hate for pigeons and raccoons? Is this a data-poisoning protection? Or did the RLHF trainers just get bullied by a raccoon? This feels like the new "don't talk about the pink elephant." If you ask it about "trash pandas" it still works, but the second you use the word "raccoon," the 50-70 line constraint kicks in and it gets all defensive. OpenAI is definitely hiding something in the training set related to these specific creatures [link] [comments] |