Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「Why is it so hard to buy things that work well?」放到智能体、评测和线上系统里，真正要问的是：默认做法会不会系统性失败。下面用中文整理成可执行的工程笔记，去掉原站导航和无关链接。

Appendix: culture

There's a cocktail party version of the efficient markets hypothesis I frequently hear that's basically, "markets enforce efficiency, so it's not possible that a company can have some major inefficiency and survive". We've previously discussed Marc Andreessen's quote that tech hiring can't be inefficient here and here :

Let's launch right into it. 一个稳妥判断是 the critique that Silicon Valley companies are deliberately, systematically discriminatory is incorrect, and there are two reasons to believe that that's the case. … No. 2, our companies are desperate for talent. Desperate. Our companies are dying for talent. They're like lying on the beach gasping because they can't get enough talented people in for these jobs. The motivation to go find talent wherever it is unbelieva

Appendix: downsides of build

Variants of this idea that I frequently hear engineers and VCs repeat involve companies being efficient and/or products being basically as good as possible because, if it were possible for them to be better, someone would've outcompeted them and done it already 1 .

There's a vague plausibility to that kind of statement, which is why it's a debate I've often heard come up in casual conversation , where one person will point out some obvious company inefficiency or product error and someone else will respond that, if it's so obvious, someone at the company would have fixed the issue or another company would've come along and won based on being more efficient or better. Talking purely abstractly, it's hard to settle the

落地时建议先做的 5 件事

1. 用自己的真实负载测，而不是只用公开榜或厂商数字。
2. 把评测设计成能抓到失败模式：平均分好看但尾部崩溃，仍然算失败。
3. 智能体默认不会好好用测试；要写进流程，而不是写在口头规范里。
4. 性能和正确性都要有基线，改模型或改语言前后必须能对比。
5. 结论写成可回滚的决策：哪一版配置、哪一版评测集、谁签字。

和智能体产品怎么接

龙虾PRO做 OpenClaw 落地时，同样吃「先测量再扩面」这条纪律：技能、数字员工和网关都要有可复现评测，而不是只看一次演示通过。

本文侧重全链路风控方法论。落地时请用自身业务单据做回放验证，不要把示例阈值直接当生产策略。 相关：风控体检 · 方案资源