“The typical policy scenario proposed by AI doomers will on net increase x-risk. If you ban all AGI development (or, as some have proposed, even 2024-level open-source models) outside of a single, highly secured, perhaps state-run Manhattan Project with no contact with the outside world, you are much more likely to end up with a misaligned superintelligent singleton than if you develop AI in a broad and multipolar way, in which case misaligned AGIs can be counteracted and defeated by aligned AGIs”