Not important, but I would appreciate it if someone could provide intuition as to why on Qwen2-7B-Instruct nearly flawlessly handles contexts up to 128k in leng

Not important, but I would appreciate it if someone could provide intuition as to why on Qwen2-7B-Instruct nearly flawlessly handles contexts up to 128k in length, the inaccuracies occur around context width = 40K. If seems counterintuitive to me that if I happen to have a context of size about 40K tokens, then if I add some noise to get the context away from 40K then there are no inaccuracies. Thanks.