I am not an expert (not even an amateur) in the area, but I wonder if the AI containment work would be futile without corrigibility figured out, and superfluous once it is? What is the window of AI intelligence where it is not yet super-human (too late to contain), but already too smart to be contained by the standard means?