“The long-held "alignment tax" theory, which posited that safe AI would be less capable, appears to be false, as building safer models is proving to be more useful and beneficial for capability.”