换把尺子,岗位就涨了1721%Change the Ruler, and the Jobs Jump 1,721%
先把这条新闻的计量单位看清楚:1,721% 不是多出了多少人,是多出了多少次提及。Draup 从公开招聘帖和 LinkedIn 这类平台上抓词,统计的是 job post 里出现 agent orchestration 这个词的频率,2025 年到今年跳了 1,721%。同一个动作,去年叫「搭系统」,今年叫「编排智能体」,词换了,涨幅就出来了。
这是我在书里拆华为 τ law 时用的同一套怀疑姿势:改变的是能力,还是评价方式?声明出来的定律和观察出来的定律性质不同,声明出来的技能热潮和观察出来的技能热潮也一样。银行真的在雇人干活,这大概是真的;「最热技能涨了 1,721%」是招聘数据公司卖给你的一把新尺子,这也是真的。两件事不冲突,但混在一起讲,读者会以为华尔街突然长出了十七倍的工作岗位。
报道自己露出了底:真正的重头是「forward-deployed engineers」,要的是技术能力加某个业务线的领域知识,从交易台到后台到人力资源。Swaminathan 说企业里的复杂性「很多是隐藏的」,连自动批准年假申请都要织一张边缘案例和特殊豁免的网。翻译一下:AI 没在替银行干活,AI 在替银行制造新的协调工作,然后需要人来协调这些协调。
再看钱。生成式 AI 经理的中位基本工资约 19 万美元(来源:CNBC)。同一篇报道里,治理相关的技能在数据里有超过 16,000 次提及,而训练、部署、运行模型相关的约 8,400 次,前者接近后者的两倍。也就是说,这个行业现在更缺的是给系统上锁的人,不是造系统的人。这不是技术成熟的标志,是技术还没被信任的标志。
信任问题最后落在谁身上?Dimon 说的「huge redeployment plans」,重部署。被重部署的是现有员工,被重训练的是现有开发者。新岗位的门槛是 19 万美元的领域专家,旧岗位的出路是「学点软技能,你会适应的」。Swaminathan 那句「对合适的人才来说,这是非常激动人心的时刻」,把「合适」两个字放在了最需要被追问的位置——不合适的那批人,报道里没有名字,没有数字,也没有人被采访。
AI 在这里扮演的角色,和所有被武器化的表达一样:它不制造事实,它重新描绘事实,然后让描绘出来的事实变成招聘市场的新常识。词频是可以被制造的,稀缺感是可以被制造的,19 万的薪资数字也是可以被制造出来的——只要足够多的招聘帖重复同一个词。
真正没被统计的是另一栏:那些因为「效率提升」而不再发布的岗位。它们不上 LinkedIn,不进 Draup 的样本,涨幅表里永远看不到它们。1,721% 的分母里没有它们,分子里也没有。
Read the unit first. 1,721% is not more people. It is more mentions. Draup scrapes public job posts and platforms like LinkedIn, counts how often the phrase agent orchestration appears, and compares this year to 2025. The same work that was called building systems last year is called orchestrating agents this year. Swap the word, and the growth rate appears.
This is the same suspicion I apply to Huawei's τ law: did the capability change, or did the way we measure it change? A law that is declared and a law that is observed are different kinds of thing. So are a skills boom that is declared and one that is observed. Banks are probably hiring real people to do real work. Also true: a hiring-data firm is selling you a new ruler. Both can hold. Mix them together and the reader concludes Wall Street grew seventeen times the jobs.
The article gives itself away. The real center of gravity is the forward-deployed engineer, who needs technical ability plus domain knowledge of a specific business line, from trading desks to back office to HR. Swaminathan says enterprise complexity is often hidden, and that even automating vacation approvals means weaving a web of edge cases and exemptions. Translation: AI is not doing the bank's work. AI is generating new coordination work, which then requires people to coordinate the coordination.
Follow the money. Generative AI managers draw a median base of about $190,000 (source: CNBC). In the same dataset, governance-related skills show over 16,000 references, nearly twice the roughly 8,400 tied to training, deploying and running models. The industry is short on people who lock the systems, not people who build them. That is not a sign of maturity. It is a sign the technology is not yet trusted.
And who absorbs the trust problem? Dimon's "huge redeployment plans." The redeployed are existing employees. The retrained are existing developers. The new door opens at $190,000 for a domain expert; the old door opens onto learn some soft skills and you will adapt. Swaminathan's line about a very exciting time for the right talent puts the whole weight on the word right. The people who are not right have no names, no numbers, and no interviews in this piece.
AI plays the role every weaponized expression plays: it does not manufacture facts, it redraws them, and then the redrawn facts become the new common sense of the hiring market. Word frequency can be manufactured. Scarcity can be manufactured. A $190,000 salary figure can be manufactured, as long as enough job posts repeat the same phrase.
The column nobody counts is the other one: the postings that were never published because efficiency improved. They do not appear on LinkedIn, they do not enter Draup's sample, they never show up in a growth table. They are missing from the denominator of 1,721%. They are missing from the numerator too.