Anthropic is proposing that top AI companies coordinate a way to pause the development of advanced AI systems if they become too dangerous.