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 (Anthropic)