In brief OpenAI's "Nerdy" personality rewarded goblin metaphors, spreading the quirk across all GPT models through reinforcement learning. Goblin mentions in GPT-5.4's Nerdy mode surged 3,881% compared to GPT-5.2, prompting an internal investigation and emergency system prompt patch. The fix—writing "never talk about goblins" in a developer prompt—shows why system prompt patches are faster but riskier than retraining. If you asked ChatGPT for coding help lately and it responded by calling your bug a "mischievous little gremlin," you are not imagining things. The model developed a genuine obsession with fantasy creatures—goblins, gremlins, raccoons, trolls, ogres, and yes, pigeons—and OpenAI published...