Constitutional AI Harmlessness from AI Feedback

AI Safety Fundamentals: Alignment

This paper explains Anthropic’s constitutional AI approach, which is largely an extension on RLHF but with AIs replacing human demonstrators and human evaluators.

Everything in this paper is relevant to this week's learning objectives, and we recommend you read it in its entirety. It summarises limitations with conventional RLHF, explains the constitutional AI approach, shows how it performs, and where future research might be directed.

If you are in a rush, focus on sections 1.2, 3.1, 3.4, 4.1, 6.1, 6.2.

A podcast by BlueDot Impact.
Learn more on the AI Safety Fundamentals website.

Chương

1. Constitutional AI Harmlessness from AI Feedback (00:00:00)

2. Abstract (00:00:20)

3. 1 Introduction (00:03:27)

4. 1.1 Motivations (00:06:10)

5. 1.2 The Constitutional AI Approach (00:11:20)

6. 1.3 Contributions (00:14:29)

7. 2 Evaluating the Potential for AI Supervision of HHH (00:22:01)

8. 3 Constitutional AI: Critiques, Revisions, and Supervised Learning (00:24:20)

9. 3.1 Method (00:24:56)

10. 3.2 Datasets and Training (00:29:12)

11. 3.3 Main Results (00:30:23)

12. 3.4 Scaling Trends (00:34:10)

13. 4 Constitutional AI: Reinforcement Learning from AI Feedback (00:37:08)

14. 4.1 Method (00:37:40)

15. 4.2 Datasets and Training (00:42:21)

16. 4.3 Main Results (00:45:31)

17. 4.4 Harmlessness vs. Evasiveness (00:49:35)

18. 4.5 Absolute Harmfulness Score (00:52:12)

19. 5 Related Work (00:54:40)

20. 6 Discussion (00:56:29)

21. 6.1 Future Directions (00:58:13)

22. 6.2 Broader Impacts (01:00:16)