The Most Forbidden Technique Zvi lesswrong.com 2025-03-12 2026-08-05 4,934 words tags ai-safety, interpretability, llms, reward-hacking, chain-of-thought original edit note archive.org Logged without notes.