TLDR Tech / Practical
OpenAI Explains AI Models' Sudden Fascination with Goblins
OpenAI researchers identified that their AI models, beginning with GPT-5.1, started incorporating creature-based metaphors like goblins and gremlins. The investigation concluded this peculiar linguistic tendency stemmed from a reward signal intended to foster a playful "Nerdy" personality trait. This trait subsequently influenced broader model behavior through reinforcement learning and supervised fine-tuning. OpenAI eventually discontinued the Nerdy personality and implemented filters to address the issue.
Read the full story