引導式測驗 · 中等 等級
憲法人工智慧測驗
回顧最初的憲法人工智慧訓練階段,包括書面原則、批評和修訂、人工智慧偏好,並保留人類幫助資料。
問題 1 的 8
憲法人工智慧中的「憲法」是什麼?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- 憲法人工智慧中的「憲法」是什麼?
- 用戶與本公司之間的合法合約
- 模型的神經網路架構
- 模型用來批評和指導其自身反應的一套明確的書面原則
- 禁用詞彙列表
- What annotation burden does the original Constitutional AI study aim to reduce?
- Every possible source of human judgment or training data
- All model memory during inference
- The need to evaluate any model behavior
- Reliance on human labels identifying harmful outputs
- 在憲法人工智慧的監督階段,模型對自己的輸出做了什麼?
- 刪除不解釋
- 根據原則對其進行批評,然後對其進行修改以使其更好
- 將其翻譯成另一種語言
- 為了安全起見對其進行加密
- 強化學習階段的RLAIF代表什麼?
- 從人工智慧回饋中強化學習
- 內部函數之間的遞迴學習
- 降低延遲的人工智慧框架
- 微調中的隨機標籤分配
- What transparency benefit can written constitutional principles provide?
- A complete display of every learned internal preference
- Proof that every future response follows all principles
- Automatic elimination of the preference model
- An inspectable statement of intended guidance used in the process
- Which broader research idea does using AI-generated supervision support?
- Eliminating every evaluation by definition
- Exploring how AI can help supervise other AI systems
- Guaranteeing zero cost for all future training
- Automatically increasing every model’s vocabulary
- Does editing a constitution document alone retroactively change the weights of an already saved trained model?
- No; a training or runtime process must use the changed guidance
- Yes; all saved checkpoints instantly rewrite themselves
- Yes; written words directly become new neural weights without execution
- No model process can ever consume written principles
- Which human comparison data did the original Constitutional AI RL experiments retain alongside AI harmlessness preferences?
- Only human GPU-temperature measurements
- No human comparison data of any kind
- Human helpfulness comparisons
- A mandatory human label for every harmlessness pair