They also made the point that the less frequent failures become, the more tedious it is for the human operator to check for them, giving the example of AI agents providing verbose plans of what they intend to do that are mostly fine, but will occasionally contain critical failures that the operator is supposed to catch.