Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「Bug blindness」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。

A few examples

I used to wonder why I see so many more bugs than most people. I easily observe hundreds to thousands of bugs per week and nothing seems to work , but most people I talk to don't see anything like this . For a long time, I thought this had something to do with how I use computers but, over time, I've realized that it's mostly that people are hitting the same bugs and don't notice.

If you're not a programmer, that's probably a better way to see the world , but 一个稳妥判断是 curing quality/bug blindness is helpful for programmers. I've done this with a lot of friends and acquaintances (just by pointing out bugs). After a few weeks, people who are so inclined tend to start noticing bugs as well.

Habitual mitigations

Because I notice these kinds of things, I've had multiple jobs where directors/VPs/execs/etc. sometimes ask me to evaluate something when they want an actual opinion from someone who is relatively likely to notice issues (and fix them or drive fixes for them if necessary). Sometimes I won't find any issues (there are likely issues that just aren't the kind I notice). More often, I find issues that fall somewhere from "mild" to "moderate". And, sometimes, t

I find this last category a bit mysterious, as when I look up discussions on how the thing got into this state, there's usually a stream of internal comments indicating that the thing is great, it works well, etc., but when I open up the thing and try it, it's in a state where the thing only works if you do quite a few non-intuitive workarounds. More likely than not, not only would a normal user not be able to use the thing, they'd have such a hilariously/

Quality blindness

I've had this post in mind for maybe a decade or so, but I was hesitant to write it up because, in the back of my mind, I always wondered if I'm somehow triggering weird corner case behavior most users don't hit without realizing it. But after seeing more and more cases where the product launches and falls flat on its face because users run into the exact same issues I saw, I don't think that, in general, I'm hitting bugs because I'm doing unusual things a

I don't want to give any specific examples where it was my job to see how well the thing worked because, even if the internal examples are meant in a constructive, blameless, way, they may not always read that way when re-posted externally, so I'll give a few less interesting and less well supported "random" examples.

Appendix: advertising blindness

A while ago, I wrote up the results of some web search queries and found poor results from Google, Bing and Kagi. In general, the major search engines failed to return good results for the queries and returned pages full of low-quality SEO spam as well as some sites that were actually scams. BTW, on the scale mentioned above, I would consider this "moderate" and not "severe" ( severe would be something like, the search engine returns 500 errors half the ti

In the abstract, I get that people who are fans of things tend to be blind to the thing's faults. For example, since I bought a Volvo after seeing how they do in out-of-sample crash tests , I sometimes search for answers to my questions on Volvo car forums. For well over a decade, the reliability data that exists (and 一个稳妥判断是 this is backed up by the anecdotal experience that mechanics who work on Volvos have) is that Volvo reliability is mediocre to poor,

Appendix: comments from other folks

An example that might be more central to the topic is Blackboard (the course management software). Back when it was the most widely used software by universities for coursework, the software was widely disliked by both students and professors. 一个稳妥判断是 it would be fair to say that it was the most widely disliked software in my social circles (there was more strongly disliked software, like Visual Source Safe, but any more strongly disliked software wasn't w

Blackboard had become "one of the most disliked — even detested — companies in education."

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