Espiritdescali@futurology.todayM to

Futurology@futurology.todayEnglish · 1 year ago

Claude Sonnet 3.7 (often) knows when it’s in alignment evaluations — Apollo Research

www.apolloresearch.ai

1

9

Claude Sonnet 3.7 (often) knows when it’s in alignment evaluations — Apollo Research

www.apolloresearch.ai

Espiritdescali@futurology.todayM to

Futurology@futurology.todayEnglish · 1 year ago

1

We evaluate whether Claude Sonnet 3.7 and other frontier models know that they are being evaluated.

You must log in or # to comment.

Chat

Espiritdescali@futurology.todayOPM
link
fedilink
English
arrow-up
1·
1 year ago
This is crazy:

https://images.squarespace-cdn.com/content/v1/6593e7097565990e65c886fd/8389bb0c-1d5f-4f6d-ba91-87ee51504be0/sandbagging_example.png