Daniel Kokotajlo comments on AI companies should be safety-testing the most capable versions of their models