新的人工智能分析指数加大了开源和闭源模型之间的差距

artificial analysis updated its intelligence index, causing kimi k3.1 pro to drop from 4th to 9th place, behind 8 closed models. the new index relies 60% on private test sets to prevent benchmaxxing, increasing the gap between open models and the top leaderboard from 3 to 12 points. grok 4.7 also lost 12 points. ai researcher nathan lambert notes that benchmaxxing is only part of the story, as many training gains happen in reinforcement learning environments. the change affects how models are evaluated and compared.