Safety in LLMs Using Chain-of-Thought: Lessons From Constitutional AI
How Anthropic used chain-of-thought prompting to improve the quality of LLM responses.
BERNEASE HERMAN
Senior Data Scientist
WhyLabs
In this webinar, we will better understand how Anthropic used chain-of-thought prompting to improve the quality of LLM responses, reducing harm and improving safety. After the LLM responds, send additional prompts to measure how the LLM response differs from the intended rules (or constitution) to generate a higher-quality response. While this technique is built into Anthropic's Claude LLMs to reduce harm and toxicity, it can be applied to any LLM model for a number of rules. We'll explore (1) additional performance improvements achieved by chain-of-thought prompting; and (2) how you can establish your own constitution to both measure and improve the safety and quality of your LLM application.
You'll learn the following in this webinar:
- Understanding chain-of-thought prompting
- Comparison of RLHF (human feedback) and RLAIF (AI feedback) preference LLM models in practice
- Use cases and best practices for chain-of-thought prompting to improve LLM response quality and safety