Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「Diseconomies of scale in fraud, spam, support, and moderation」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: techniques that only work at small scale

If I ask myself a question like "I'd like to buy an SD card; who do I trust to sell me a real SD card and not some fake, Amazon or my local Best Buy?", of course the answer is that I trust my local Best Buy 1 more than Amazon, which is notorious for selling counterfeit SD cards. And if I ask who do I trust more, my local reputable electronics shop (Memory Express, B&H Photo, etc.), I trust my local reputable electronics shop more. Not only are they less li

I don't think it's controversial to say that in general, a lot of things get worse as platforms get bigger. For example, when I ran a Twitter poll to see what people I'm loosely connected to think , only 2.6% thought that huge company platforms have the best moderation and spam/fraud filtering. For reference, in one poll, 9% of Americans said that vaccines implant a microchip and and 12% said the moon landing was fake . These are different populations but

Appendix: Theory vs. practice

However, over the past five years, I've noticed an increasingly large number of people make the opposite claim, that only large companies can do decent moderation, spam filtering, fraud (and counterfeit) detection, etc. We looked at one example of this when we examined search results , where a Google engineer said

Somebody tried argue that if the search space were more competitive, with lots of little providers instead of like three big ones, then somehow it would be *more* resistant to ML-based SEO abuse.

Appendix: How much should we trust journalists' summaries of leaked do

And… look, if *google* can't currently keep up with it, how will Little Mr. 5% Market Share do it?

like 95% of the time, when someone claims that some small, independent company can do something hard better than the market leader can, it’s just cope. economies of scale work pretty well!

Appendix: Erin Kissane on Meta in Myanmar

But when we looked at the actual results, it turned out that, of the search engines we looked at, Mr 0.0001% Market Share was the most resistant to SEO abuse (and fairly good), Mr 0.001% was a bit resistant to SEO abuse, and Google and Bing were just flooded with SEO abuse, frequently funneling people directly to various kinds of scams . Something similar happens with email, where I commonly hear that it's impossible to manage your own email due to the spa

I started seeing a lot of comments claiming that you need scale to do moderation, anti-spam, anti-fraud, etc., around the time Zuckerberg, in response to Elizabeth Warren calling for the breakup of big tech companies, claimed that breaking up tech companies would make content moderation issues substantially worse, saying :

Appendix: elsewhere

It’s just that breaking up these companies, whether it’s Facebook or Google or Amazon, is not actually going to solve the issues,” Zuckerberg said “And, you know, it doesn’t make election interference less likely. It makes it more likely because now the companies can’t coordinate and work together. It doesn’t make any of the hate speech or issues like that less likely. It makes it more likely because now … all the processes that we’re putting in place an

It’s why Twitter can’t do as good of a job as we can. I mean, they face, qualitatively, the same types of issues. But they can’t put in the investment. Our investment on safety is bigger than the whole revenue of their company. [laughter] And yeah, we’re operating on a bigger scale, but it’s not like they face qualitatively different questions. They have all the same types of issues that we do."

Appendix: Moderation and filtering fails

The argument is that you need a lot of resources to do good moderation and smaller companies, Twitter sized companies (worth ~$30B at the time), can't marshal the necessary resources to do good moderation. I found this statement quite funny at the time because, pre-Twitter acquisition, I saw a much higher rate of obvious scam content on Facebook than on Twitter. For example, when I clicked through Facebook ads during holiday shopping season, most were scam

Zuckerberg seems to like the line of reasoning mentioned above, though, as he's made similar arguments elsewhere, such as here , in a statement the same year that Meta's internal docs made the case that they were exposing 100k minors a day to sexual abuse imagery:

值得单独记下的观察

• Anna Lowenhaupt Tsing's On Nonscalability: The Living World Is Not Amenable to Precision-Nested Scales
• Glen Weyl on radical solutions to the concentration of corporate power
• Zvi's collection of Quotes from Moral Mazes
• "I had to get the NY attorney general to write them a letter before they would actually respond to my support requests so that I could properly file my taxes"
• Two different users report having their account locked out after moving; no recovery of account
• Google Cloud reduces quota for user, causing an incident, and then won't increase it again User tries to find out what's going on and has this discussion: GCP support : You exceeded the rate limit
• User : We did 5000/10min. The quota was approved at 18k/min
• GCP support : That's not the rate limit

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