Anthropic:Anthropic details using AI agents to accelerate alignment research on “weak-to-strong supervision”, where a weak model supervises the training of a stronger one — Large language models' ever-accelerating rate of improvement raises two particularly important questions for alignment research.