2 ms·Two-faced AI language models learn to hide deception3 points by mellosouls 3y agojruohonen 3y agoA very long paper but at the end it is just plain old poisoning, although the threat of deceptive instrumental alignment is new to my understanding.