We’re proposing an AI protection method known as iterated amplification that we could us specify sophisticated behaviors and targets which might be past human scale, by means of demonstrating the best way to decompose a role into more practical sub-tasks, reasonably than by means of offering categorised information or a praise serve as. Even though this concept is in its very early phases and we’ve got handiest finished experiments on easy toy algorithmic domain names, we’ve made up our minds to provide it in its initial state as a result of we predict it will turn out to be a scalable option to AI protection.
