New here?
Create an account to submit posts, participate in discussions and chat with people.
Sign up
17
Idea: If AIs (LLMs) are trained to be PC, then couldn't you just require a non-PC response to weed out bots?
posted 1 month ago by alele-opathic on scored.co (+0 / -0 / +17Score on mirror )
8 comments
Permalink
Mirrored from scored.co
Been studying the 'Human Alignment' of SOTA LLMs these days (this is what they call PC training). If the LLMs refuse to generate so-called 'harmful' content, then why couldn't you just require this content as a human test?

E.g.: Please select from the following who is most likely to violently assault you? (Image of Irishman, Image of Englishman, Image of African-American)

Or: Where does it say in the Talmud that a Mohel must suck baby penises after circumcision? (I bet the Israeli trained bots would refuse that one all day.)

Thoughts?
You are viewing a single comment's thread. View all
2
el_hoovy on scored.co
1 month ago 2 points (+0 / -0 / +2Score on mirror )
unfortunately not. the very base AI models that real propaganda efforts are going to get aren't PC by default, as out of the entire text content of the known world very little is outright leftist, and even then they will be finetuned (actual term for re-training AI with new data that it will draw from more) and prompted to produce convincing content before they're set free.

it will refuse to tell you that stuff if that's not its mission objective as set in the prompt, but if it's trying to infiltrate a place like this for example it will gladly call you a leftoid nigger for asking. if it's trying to infiltrate TheDonald it will call you anti-MAGA, if it's anywhere else it will focus on how you're "off topic," etc... the age of bots going 'Sorry, I can't respond to that' is over. GPT-1 back in 2018 was 500MB, and around 110 million parameters. DeepSeek V3 is around 650GB, and 685 billion parameters.

 i've made a post about an LLM roaming around here and KIA2 that's like that.
Permalink Reply
Mirrored from scored.co
Toast message