Comment by ivanblagdan

Semantics of how this works aside, take a moment to appreciate how easy it is to remap the variable “zombie” to “human” in a prompt without the model altering its behavior. It instantly makes you realize the immensity of the AI safety & alignment problem.