LLMs Sometimes Generate Purely Negatively-Reinforced Text — LessWrong