“Pausing AI development to focus on alignment research at a fixed level of compute would not yield significant progress, as scaling compute is necessary to make new discoveries in alignment.”