From hard refusals to safe-completions: toward output-centric safety trainingopenai.com1 point·6thbit··0 commentsOpen articleSaveView on HN