I'm not an AI safety expert. But I was thinking through alignment tractability a bit today. I convinced myself of the below and wanted to see if anyone knew whether it was (a) wrong or (b) similar to any existing models. We model a single agent invocation as a function: $$ f : X \to Y $$ where $X$ is the set of possible inputs/world states and $Y$ the set of possible outputs/actions. Suppose a policy is, in the most general case, a predicate...
Originally asked by Liam Monninger on Stack Exchange (ai) (CC BY-SA 4.0).
