Helpful Harmless Honest
Domain: LLM Concepts
🎯 Motivation
The three HHH goals of alignment: the model should be helpful (solve the user’s problem), harmless (not cause damage), and honest (not deceive).
📋 Concrete Example
Helpful: gives a useful answer. Harmless: refuses to help with weapons. Honest: admits uncertainty instead of making things up. Sometimes these conflict - harmless may require refusing a helpful request.
🔗 Analogy
A good doctor - helpful (treats your illness), harmless (won’t prescribe dangerous drugs), honest (tells you the truth about your condition). All three are essential; balancing them when they conflict is the challenge.