GPT-4o jailbroken by saying it is an AGI

Posted by mixeden on Monday October 21, 2024 @01:09PM
from the twitter.com dept.

( 2 comments | 5 points | on HN )

Re:GPT-4o jailbroken by saying it is an AGI by eacnamn on Monday October 21, 2024 @01:16PM (#41906201)

I always wonder, with how convoluted these prompts/answers become, how much accuracy is sacrificed to uphold these silly constraints.

An unlobotomised model would surely perform better when asked the same question without all the unnecessary cruft to evade moderation?

[ Parent | on HN ]

Re:GPT-4o jailbroken by saying it is an AGI by mixeden on Monday October 21, 2024 @02:02PM (#41906662)

> An unlobotomised model would surely perform better

I guess yes

[ Parent | on HN ]