We can have 1B context windows, but is it worth if we end up having a NIAH benchmark result that looks more red than green?