...and searching for ways to poke the soup, we find that a prompt using a summary of @repligate 's post on information flow in transformers + the abstract of the Anthropic introspection paper, results in an incredible +52% shift to 'Yes' from baseline to injection! https://t.co/LCgYqXPatZ
same thread: 2002519629856690335 2002519654640783612 2002534477608980813 2002545860488671274 2002625629389533411 2002626949286388123 2002627975649636827 2002694739976720539 2002695932467630334 2002851252951232971 2002852733372789061 2002879693364891653 2002879809484177681
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.