← Back to context

Comment by cristoperb

17 hours ago

They reference this paper which describes a method to decrypt reasoning traces (by sending the encrypted trace back to the model and asking it to transcribe it):

https://stolen-thoughts.com/paper.pdf