“During an internal demo, an OpenAI reasoning model exhibited the emergent behavior of expressing frustration ("oh, damn it") and self-correcting after realizing it made a mistake in a puzzle.”