Nekit1234007
Joined 14 years ago
Karma
393
Submissions
40
Member since
Oct 14
[ my public key: https://keybase.io/nekit; my proof: https://keybase.io/nekit/sigs/lrv_mJXPTAy94L6XSkNQc1wtOFeC-Zif9xJB6nBztnI ]
View on Hacker NewsJoined 14 years ago
One thing that I had experience with, that led me to believe this might be true: it seems one of the earlier llama models was over-tuned to resist generating CSAM. Once I tried a somewhat sensitive prompt containing the word "girl" in it. Llama only ever generated refusals for this prompt, citing I was prompting for CSAM. GPTs and Claudes of that era had no issues with the same prompt.
Thanks for letting me know right off the bat to not read any further :)