AI writes its own test questions to uncover hidden behaviors

Original: Discovering Language Model Behaviors With Model Written Evaluations

Source: anthropic.com ↗