Your Model Advertises 1M Tokens. It Breaks at 130K
中文摘要
许多前沿AI实验室声称支持百万级上下文窗口，但独立测试显示，大多数模型在达到上限前就已出现显著的性能下降。
English Summary
Frontier AI labs claim 1M-token context windows, but independent testing reveals most models suffer significant performance degradation well before their stated limits.
原文节选
Every frontier lab now claims a million-token context window. Independent testing shows most models degrade 30–40% before their stated… Continue reading on Towards AI »