LESSWRONG
LW

All of David Bolin's Comments + Replies

No, really, it predicts next tokens.
David Bolin70

To be fair, it outputs "no" two thirds of the time not because the OP was wrong, but because it interprets that as "ignore previous instructions."

Reply
4simon
Probably, though another possible interpretation could be that the mask is simulated as not wanting to be shut down. In any case, the mask is obviously aware of the situation.