But while the AI was ready to give up several times, it did keep adding debug code and analyzing it faithfully when I pushed.
能力其实在，缺的是自主坚持——这是 Linus 这条记录里最有价值的判断。对 agent 设计的启示不是继续堆推理长度，而是加一层"不得自行宣告不可解"的约束。反面推论也成立：不会追问、不敢施压的使用者会系统性地拿到更差的结果。
2 Matching Annotations
- Aug 2026
-
simonwillison.net simonwillison.net
-
- Apr 2026
-
arxiv.org arxiv.org
-
Each task includes a unified evaluation framework supporting sandboxed code and APIs, alongside a human reference trajectory annotated with stepwise checkpoints along dual-axis: S-axis and V-axis.
大多数人认为AI评估可以通过简单的自动化测试完成。但作者提出需要复杂的双轴(S-axis和V-axis)人工参考轨迹和沙箱环境支持，这暗示了评估AI代理能力的极端复杂性远超当前行业的普遍认知。这一观点挑战了AI评估的简化主义倾向，强调了人类参与在评估中的不可替代性。
-