AI模型评测平台Arena负责人Peter Gostev于北京时间10月16日发布国际象棋实验复盘，指出当前大模型仅靠反复练习、保存记忆和延长上下文，仍难以实现稳定的持续学习。 实验先让GPT-6 Astra和Claude Opus 5.5分别与Stockfish棋类引擎对弈200局，自主选择难度、记录经验并调整策略。Opus最终获得约1760的估算Elo评分，高于Astra的1400（Elo评分是一种衡量棋手相对竞技水平的评分体系，广泛应用于国际象棋等竞技项目）。 但在随后的30局直接对抗中，Astra反而以24.5比5.5大胜。另一项对照测试显示，GPT-6.1 Sol使用约80万Token长上下文时，表现反而逊于约30万Token版本。 Gostev认为，模型在长时间运行中可能形成固定策略或反复改变战术，却未必能有效积累经验。实验反映出，增加上下文容量和交互次数，并不必然转化为持续学习能力。

据悉英国向乌克兰提供更多“风暴阴影”导弹以打击俄罗斯。

The 28th General Conference on Weights and Measures on the 15th in Versailles, France adopted a resolution to implement "continuous Coordinated Universal Time" from 20 May 2027. The change means UTC will no longer be adjusted by adding or subtracting leap seconds and could remain continuous for several centuries. UTC is currently the only time scale recommended as an international reference; it runs at the same rate as International Atomic Time while historically using leap seconds to stay close to Earth-rotation-based Universal Time. Since 1972, UTC has added or removed one second when the monitored difference reached 0.9s. Under the resolution the maximum permitted difference going forward is set at 3,600 seconds (one hour).