OpenAI 提出审慎对齐策略，让 o1 模型推理安全规范

Deliberative alignment: reasoning enables safer language models Introducing our new alignment strategy for o1 models, which are directly taught safety specifications and how to reason over them.