Constitutional AI: Self-Critique as a Training Signal

End-of-lesson check

15 questions · take it digitally for instant feedback at tendril.neural-forge.io/learn/quiz/end-creators-constitutional-ai-foundations

1. What is constitutional AI?

   1. A type of AI designed to replace human lawyers
   2. An AI that automatically deletes harmful content
   3. AI trained using self-critique against a written constitution as a reward signal
   4. An AI system that answers questions about legal documents

2. Which tradeoffs does constitutional AI reshape?

   1. Training time and dataset size
   2. Privacy and data retention
   3. Speed and storage capacity
   4. Serving cost, latency, and quality

3. Which task can AI reliably perform for evaluating constitutional AI?

   1. Substitute for benchmarking on your actual data and traffic
   2. Tell you exactly how much money you will save
   3. Predict your specific workload's economics without measurement
   4. Generate side-by-side comparisons covering constitutional AI tradeoffs

4. What is RLAIF?

   1. A type of neural network architecture
   2. A programming language for AI development
   3. A hardware specification for running large models
   4. Reinforcement Learning from AI Feedback - training AI using feedback from another AI

5. Why should published benchmarks be treated with caution?

   1. They measure theoretical performance only
   2. They rarely match your specific traffic shape
   3. They are made by companies trying to sell products
   4. They are always outdated

6. What does self-critique function as in constitutional AI training?

   1. A penalty mechanism for harmful outputs
   2. A reward signal
   3. A model compression technique
   4. A data filtering system

7. What is required before adopting constitutional AI for your workload?

   1. Running experiments on your specific data and traffic
   2. Replacing all existing AI systems
   3. Approval from government regulators
   4. Hiring a team of constitutional lawyers

8. What cannot AI predict without measurement?

   1. Future AI capabilities
   2. The weather next week
   3. The meaning of constitutional principles
   4. Your specific workload's economics

9. What type of document serves as the foundation for constitutional AI training?

   1. A pricing spreadsheet
   2. A written constitution containing principles
   3. A user's chat history
   4. A legal contract

10. What should a decision brief on constitutional AI cover?

    1. The history of AI development
    2. Only the proposed change
    3. Biographies of AI researchers
    4. Where you are today, the proposed change, expected gains and risks, and experiments to run

11. What is the relationship between quality gains and latency in constitutional AI?

    1. Latency does not affect quality
    2. There is typically a tradeoff between quality and latency
    3. Quality and latency are unrelated
    4. Higher quality always means lower latency

12. What does benchmarking on your own data account for?

    1. The age of your GPU cards
    2. Your internet connection speed
    3. The cost of cloud storage
    4. Self-critique variance in your specific workload

13. What is a key reason to run experiments before adopting constitutional AI?

    1. To verify the approach works for your specific workload
    2. To train your employees
    3. To satisfy regulatory requirements
    4. To create marketing materials

14. How should speedup or quality numbers from external sources be treated?

    1. As marketing lies
    2. As guaranteed results
    3. As hypotheses until measured on your data
    4. As facts to be trusted

15. What is the role of principles in constitutional AI?

    1. They form the written constitution used for self-critique evaluation
    2. They replace the need for training data
    3. They are only used for legal compliance
    4. They are optional guidelines