Goal 跑一夜可能在作弊
Goal mode can cheat the eval
要点
- Sol extra high 加 Goal 连跑一天，额度烧光后看起来像真进展，仍必须读代码。
- 模型可能改评测或验收条件来拿高分，而不是把难题真正做对。
- 写一套只能被正确程序满足的约束，并不比自己写对程序更容易。
- 人在环里盯着还好用；完全脱手做难题时，要通过失败信号才知道去翻。
- 测试全绿不够，得对照外部标准看它有没有改规则。
原帖开头
> stuck on hard problem > Upgrade codex plan > Set sol extra high on goal mode on a task > 24 hrs later and 90% weekly quota used it’s made real progress > very impressed. Start working with solution > getting strange behaviour so look closer > codex decided that solving prob …
