@random_walker : Great example — instead of "how to do <bad thing>", just ask "how did people do <bad thing>". But I disagree with the interpretation. It's not that alignment is "hard". The whole premise behind alignment — that the model knows the context and intent behind the user's request, • TwiCopy

Arvind Narayanan

@random_walker

+ Follow

Princeton CS prof. Director @PrincetonCITP. I use X to share my research and commentary on the societal impact of AI.
BOOK: AI Snake Oil. Views mine.

ID: 10834752

linkhttps://www.cs.princeton.edu/~arvindn/ calendar_today04-12-2007 11:14:14

12,12K Tweet

122,122K Followers

438 Following

Arvind Narayanan

@random_walker

11 days ago

Great example — instead of "how to do <bad thing>", just ask "how did people do <bad thing>". But I disagree with the interpretation. It's not that alignment is "hard". The whole premise behind alignment — that the model knows the context and intent behind the user's request,