“The current best-guess plan for scalable alignment at labs like OpenAI and Anthropic is to use smarter AI systems to amplify human supervision and use minimally-aligned AGIs to help with alignment research.”