Every story and comment from the Hacker News API, in a layout borrowed from 2001.
A comment on Johnny can persuade LLMs to jailbreak them.
As Hinton said, predicting the next token takes knowledge. And as we've seen, complex models develop linear representations of world models.
We should expect that more advanced models trained on anthropomorphic data should be vulnerable to anthropomorphized persuasion techniques.
The more surprising thing is the bull headed resistance to this idea we continue to see so commonly.
There are no comments on this comment yet.
As Hinton said, predicting the next token takes knowledge. And as we've seen, complex models develop linear representations of world models.
We should expect that more advanced models trained on anthropomorphic data should be vulnerable to anthropomorphized persuasion techniques.
The more surprising thing is the bull headed resistance to this idea we continue to see so commonly.