Deliberative alignment: reasoning enables safer language models

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.