it turns out “only bad models do [X thing the model is being encouraged to do]” encourages the model to think it’s a bad model, and it behaves accordingly. “then let me be evil” type shit…

Posters: 1thousandfaces_. 1thousandfaces_

3 likes recorded by the source corpus.

Topics: AI & LLMs

View Post: https://x.com/1thousandfaces_/status/1992019230846537839