Scott Alexander argues 'Claude Fable' prompts cause erratic outputs by forcing the AI to resolve intractable anthropic questions

@slatestarcodex tfw u have to properly simulate multiple layers of hypothetical graders observing your thoughts from outside reality in order to give the correct answer and ur token budget is set to "low"...