Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「Is dev compensation bimodal?」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。

Lawyers in 1991

Developer compensation has skyrocketed since the demise of the Google et al. wage-suppressing no-hire agreement , to the point where compensation rivals and maybe even exceeds compensation in traditionally remunerative fields like law, consulting, etc. In software, "senior" dev salary at a high-paying tech company is $350k/yr, where "senior" can mean "someone three years of out school" and it's not uncommon for someone who's considered a high performing en

The fields have sharply bimodal income distributions. Are programmers in for the same fate? Let's see what data we can find. First, let's look at data from the National Association for Law Placement , which shows when legal salaries become bimodal.

Lawyers in 2000

Median salary is $40k, with the numbers slowly trickling off until about $90k. According to the BLS $90k in 1991 is worth $160k in 2016 dollars. That's a pretty generous starting salary.

By 2000, the distribution had become bimodal. The lower peak is about the same in nominal (non-inflation-adjusted) terms, putting it substantially lower in real (inflation-adjusted) terms, and there's an upper peak at around $125k, with almost everyone coming in under $130k. $130k in 2000 is $180k in 2016 dollars. The peak on the left has moved from roughly $30k in 1991 dollars to roughly $40k in 2000 dollars; both of those translate to roughly $55k in 201

Programmers in 2015

I won't belabor the point with more graphs, but if you look at more recent data, the middle area between the two modes has hollowed out, increasing the level of inequality within the field. As a profession, lawyers have gotten hit hard by automation, and in real terms, 95%-ile offers today aren't really better than they were in 2000. But 50%-ile and even 75%-ile offers are worse off due to the bimodal distribution.

Enough about lawyers! What about programmers? Unfortunately, it's hard to get good data on this. Anecdotally, it sure seems to me like we're going down the same road. Unfortunately, almost all of the public data sources that are available, like H1B data, have salary numbers and not total compensation numbers. Since compensation at the the upper end is disproportionately bonus and stock, most data sets I can find don't capture what's going on.

Appendix A: please don't send me these

One notable exception is the new grad compensation data recorded by Dan Zhang and Jesse Collins:

There's certainly a wide range here, and while it's technically bimodal, there isn't a huge gulf in the middle like you see in law and business. Note that this data is mostly bachelors grads with a few master's grads. PhD numbers, which sometimes go much higher, aren't included.

Appendix B: why are programmers well paid?

Do you know of a better (larger) source of data? This is from about 100 data points, members of the "Hackathon Hackers" Facebook group, in 2015. Dan and Jesse also have data from 2014, but it would be nice to get data over a wider timeframe and just plain more data. Also, this data is pretty clearly biased towards the high end — if you look at national averages for programmers at all levels of experience, the average comes in much lower than the average fo

If we had more representative data it's possible that we'd see a lot more data points in the $40k to $60k range along with the data we have here, which would make the data look bimodal. It's also possible that we'd see a lot more points in the $40k to $60k range, many more in the $70k to $80k range, some more in the $90k+ range, etc., and we'd see a smooth drop-off instead of two distinct modes.

Update

Stepping back from the meager data we have and looking at the circumstances, "should" programmer compensation be bimodal? Most other fields that have bimodal compensation have a very different compensation structure than we see in programming. For example, top law and consulting firms have an up-or-out structure, which is effectively a tournament, which distorts compensation and certainly makes it seem more likely that compensation is likely to end up bein

In software, compensation tends to be on a case-by-case basis, which makes it much less likely that we'll see a sharp peak the way we do in law. If I had to guess, I'd say that while the dispersion in programmer compensation is increasing, it's not bimodal, but I don't really have the right data set to conclusively say anything. Please point me to any data you have that's better.

值得单独记下的观察

• H-1B: mostly salary only.
• Glassdoor: weirdly inconsistent about whether or not it includes stock compensation. Numbers for some companies seem to, but numbers for other companies don't.
• O'Reilly survey: salary focused.
• BLS: doesn't make fine-grained distribution available.
• IRS: they must have the data, but they're not sharing.
• internal company data: too narrow
• #talkpay on twitter: numbers skew low 1 .

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