文 | 影子备忘录
周岚接到部门解散通知的那天，会议室里坐了五十个人。没有人哭，也没有人拍桌子，大家安静地听HR念完流程，然后各自回工位收拾东西。
她在这家公司做了七年运营，从专员做到主管，经手的项目流水以亿计，到头来全部归零只用了一个下午。
后来她转进了一家传统制造企业做行政，薪资降了大约三成。她没有抱怨太多，只说了一句：“至少那边还在招人。”
“那边”指的是工厂车间。过去一年里，写字楼与工厂之间那条曾经泾渭分明的界线，正在被一批又一批失业的白领踩得模糊。
社交平台上，“本科生进厂”的话题积累了数千万浏览量，前端工程师在裁员后转身走进PLC培训班，三十多岁的中层管理者把简历投向了从未正眼看过的制造业岗位。
一个共识正在被 quietly 改写：写字楼的工位不再天然比流水线更安全。
但把这个现象简单讲成“AI抢了白领的饭碗，所以大家去工厂”，恰恰漏掉了最关键的部分。AI替代的从来不是笼统的“人”，而是一组可以被拆分、编码、远程交付的任务。
写字楼里的工作之所以最先被冲击，并非因为坐在里面的人不够努力，而是因为他们手里的大部分活，天然就是模型最容易消化掉的那种材料，比如代码、报表、文案、客服对话、数据核对等。
而工厂之所以成为退路，也不是因为它更“高级”或更“稳定”，只是那里还残留着AI暂时够不着的东西。那些是一台卡料的注塑机发出的异响，一条产线上标签贴歪时需要的即时判断，以及一双必须亲自伸进机器里的手。
被拆解的工位
陈默是重庆人，二十八岁，做了五年前端。五年里他换了五家公司，薪资从起步爬到九千多就停住了，之后再也没动过。
今年四月第二次被裁之后，他投出去上百份简历，收到的面试邀约全部来自外包岗。空窗四个月，他没有再等。
八月中旬，他在社交账号上发了一张图，上面只有四个字：“前端已死”。当天他刚拿到低压电工证，第二天就去PLC培训机构报了到。
陈默的选择在程序员群体里不算孤例。我接触到的八位从互联网转向PLC或上位机的从业者中，有六位原本做的是前端。
从行业发展来看，这并非偶然。AI对编程岗位的冲击有一条清晰的下滑曲线：它最先吃掉的是那些边界明确、模式重复、可以被拆成一个个小任务的编码工作。
而前端恰好是这类任务最密集的工种，相当于以前十个人干的活，现在借助AI两个人就能全部完成。
Anthropic基于真实使用数据构造的“观测暴露度”给出了一个更系统的图景：计算机程序员以74.5%位居AI实际覆盖率的首位，客服代表70.1%，数据录入员67.1%，病历记录专员66.7%。
报告同时指出，AI理论上能完成的计算机与数学岗位任务高达94%，但实际使用中仅覆盖33%，因为法律约束、软件条件和人工复核流程拖慢了落地速度。
这意味着，暴露度最高的那批岗位，恰恰是AI“能够碰”且“已经在碰”的地方。写字楼里那些年薪体面、工位靠窗的职位，其核心内容往往正好落在模型最擅长的射程之内。
但把裁员潮全部归给AI，也是一种过于顺滑的叙事。牛津经济研究院在2026年初的一份研究简报中明确指出，目前并没有证据表明企业正在大范围用AI取代员工；AI在不少情况下更像是企业为常规裁员寻找的一个说法。
科技行业CEO们把AI称为“完美的替罪羊”，“它让事情看起来不是我们的决定、我们的错，而是技术的”。
换句话说，真实的岗位收缩是技术替代、资本周期和裁员修辞三者叠加的结果。
AI掏空了一部分工位的内核，同时为企业提供了一套不会引发恐慌的措辞。
对于坐在工位上的人来说，原因的复合性并不重要。
结果是一样的：门开了，你得走。
学历的倒挂
离开写字楼之后，第一课是重新学习“求职”这件事本身。
陈默还算幸运的，他有编程底子，PLC培训五个月后底薪接近一万。
但更多人的过渡路径没有这么平滑。一个二十二岁的二本毕业生小陆，经济管理专业，毕业后进了一家做冰箱的小厂。
他犯过一个致命的错误：在入职表上如实填了本科学历。从那以后，每一句批评都带上了学历的前缀：“你一个大学生这么简单的事都不会做”“学习好又有什么用”。
流水线上十五个人，七八个来自四川同一个村子，剩下几个是结伴出来的同乡，只有他是“闯入者”。
他的本科学历在工友中间显得刺眼，像一件穿错了场合的衣服。
其实在多个工厂的招聘环节中，中介会明确叮嘱本科及以上学历的求职者“不要向厂方提及”；有的工厂认为高学历者“稳定性差、易跳槽”，更倾向招大专及以下学历的人。
社交平台上有人总结出经验：如果只是做普工，本科文凭最好不要写出来，随便填个高中或大专即可。
写字楼里被反复定价的“人力资本”，到了工厂门口需要被主动隐藏，才能获得一个站上流水线的资格，这是一套评价系统对另一套评价系统的拒绝承认。
在互联网行业，学历、履历、职级构成一条连续的阶梯；在车间，这条阶梯的大部分刻度都不作数。
工厂要的是稳定性、体力、服从排班，以及在重复动作中不出错的能力。而这条原本被当作“底牌”的退路，本身也在收窄。
人社部的数据显示，传统流水线上的装配、焊接、包装岗位AI替代率已达72%，汽车等部分行业超过80%；中国制造业用工人数从2008年高峰期的约2亿降至目前的1.2亿左右。
全国“黑灯工厂”数量已超过500家，青岛某家电企业的用工量减少了八成。
一家轴承厂的老操作工干了二十三年，新产线名单上没有他，因为整座厂只留九个运维工定时巡检。
白领们涌向工厂时，迎接他们的并不是一个静止的、等待劳动力填入的容器。
那个容器本身也在被自动化重新塑形：最底层的过渡性岗位在收缩，留存下来的岗位越来越多地要求另一种能力，即诊断高度自动化产线的复杂故障，成为“机器人的保姆和医生”。
而这类岗位的门槛，并不比写字楼低。
一平方米的“牢房”
陈默的同路人唐海镇走的是另一条路。他从互联网测试岗离职后，经过五个月PLC培训，2026年六月正式入行，底薪接近一万元。
从数字上看，这几乎是一次成功的技能平移：编程能力没有浪费，只是从浏览器和服务器换到了传送带和机械臂。但数字背后是另一种生活。
PLC工程师的工作状态高度同质：全年驻厂出差，工厂位置偏僻，住小旅店，吃路边十几块的盒饭。
一位转行上位机的从业者这样描述车间里的处境：作为乙方派驻人员，活动区域就是机器旁一两平方米的空间，没有出入权限，上厕所要报告，让人带出去，所以一天都不敢多喝水。
设备出问题，甲方几个人叉腰站在旁边盯着你修；半夜被叫起来，领导冲你吼“没搞完还敢下班”，有人把这种状态叫作“一平方米的牢房”其实不算冤枉。
但这批人之所以接受这样的处境，是因为他们清楚自己换回来的是什么。
PLC做的是非标准化产品，每条产线按客户需求定制，写程序只占工作的三成，关键是现场调试和解决实际问题的能力。
这份工作“不吃年龄”，经验是加分项而不是负担。
对于一个在互联网行业被三十五岁红线反复挤压的人来说，这已经足够构成留下的理由。
他们不是在追求一份更好的工作，只不过是在逃离一种随时可能被“优化”的状态。
传送带旁边的位置不体面，但它至少不要求你每天证明自己比昨天的模型更快。
谁支付了差价
如果把镜头拉远，周岚、陈默、小陆和唐海镇各自站在不同的位置上，共同拼出了一幅比“AI替代”四个字复杂得多的图景。
世界经济论坛的分析显示，高AI暴露职业中二十二至二十五岁劳动者的就业已经出现下降；Anthropic的数据则表明，其技术部门初级员工占比从2025年初的21%降至14%，新招聘员工中的初级占比从27%降至9%。
入门级的梯子正在被抽掉，而站在梯子最下面的人，被要求自己去别处找路。
陈默用失业补偿金和几个月脱产时间支付了转行的成本；小陆用隐藏学历换来了一个流水线岗位；周岚用三成薪资降幅换来了“那边还在招人”的确定性。
再技能化的账单，最终是由个体用失业金、积蓄和向下流动的薪资自己结清的。
真正从AI效率提升中获益的企业，并不在这条成本链上。
更深的错位在于时间。白领被裁的时候，往往正处在现金流最刚性的年纪：房贷、育儿、父母的医疗，每一项都等不起一个完整的培训周期。
工厂技术岗的入门需要学徒期和实习期，期间收入可能归零。一个三十多岁的前程序员在物流仓库上夜班，瞒着家人，每天下班先去公厕拍掉工服上的灰，把衣服塞进背包最底层。
传送带不会问他过去的KPI，也不会问他还能撑多久。它只要求他弯腰，抱起纸箱，核对标签，重复。
陈默还在等他的PLC课程结束；唐海镇已经在车间里找到了某种踏实感，尽管那踏实感是用全年驻厂和一平方米的活动空间换来的；周岚在新公司的行政岗位上很少再提过去的事。
他们各自站在传送带的不同位置上，共同指向一个尚未被充分讨论的事实：被AI叙事推到工厂门口的人没有消失，他们只是从一套评价系统掉进了另一套，而两套系统之间，不存在平滑的换算率。
传送带还在转，它不读简历，不问学历，也不关心来的人曾经在哪个写字楼里加过班。它只关心下一件东西有没有放对位置。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
A shadow memo.
Fifty people sat in the conference room the day Zhou Yi was notified of the dissolution of the department. No one cried, no one filmed the table, everyone listened quietly to HR and then went back to work and packed.
She ran the company for seven years and managed from the Commissioner to run the projects in the hundreds of millions, all of which ended up in a single afternoon.
She was later transferred to a traditional manufacturing enterprise for administration, and her salary was reduced by about 30 per cent. She didn't complain too much, she just said, "At least there's people still there."
“Over there” refers to the factory workshop. In the past year, that once-severed line between the writing building and the factory is being blurred by a series of unemployed white collars.
On the social platform, the topic of “Bed students entering the factory” accumulated tens of millions of views, with front-end engineers turning to PLC courses after lay-offs, and mid-level managers in their thirties turning their curricula vitae to jobs that they had never seen before.
A consensus is being rewritten quietly: the work space in the writing building is no longer naturally safer than the current line.
But it's a simple story that says, "Ai took a white-collar job, so you go to the factory, and that's exactly what's critical. AI has never replaced a general “person”, but a set of tasks that can be delivered remotely by split, coded.
The work in the writing building was first hit not by the lack of effort on the part of the people sitting in it, but by the fact that most of the work in their hands is naturally the type of material that models can most easily digest, such as codes, statements, paperwork, client conversations, data reconciliation, etc.
And it's not because it's more "higher" or "stabilized" because it's just something that's not there for the time being. Those are the sound of a plating machine with a card, the instant judgement required when a line label is mislabeled, and the hand that must be reached in person.
Dismantling places
Chen Moo is Chongqing, 28 years old, at the end of five years. He changed five companies in five years, and his salary stopped at more than 9,000, and he never moved again.
After his second dismissal in April this year, he cast out hundreds of resumes, and all the interviews he received came from outsourced positions. The window was empty for four months and he didn't wait.
In mid-August, he posted a map on the social account, with only four words: “the front end is dead”. On that day, he received a low-voltage electrician certificate and reported to the PLC training facility the following day.
Chen's choice is not an isolated case in the group of programmers. Six of the eight people I've been exposed to moving from the Internet to PLC or on board were at the front end.
This is not an accident in terms of industry development. The impact of AI on programming positions is a clear downward curve: The first thing it ate was the coding of well-defined borders, duplicated patterns that could be broken down into small tasks.
The front end coincides with the most intensive of these tasks, equivalent to the work done by the previous 10 individuals, which can now be accomplished in full by two persons from AI.
Anthropic gave a more systematic picture based on “observation exposure” based on the real use of data: computer programmers ranked 74.5 per cent at the top of AI's actual coverage, with visitors representing 70.1 per cent, data entry clerks 67.1 per cent, medical records commissioner 66.7 per cent.
At the same time, the report notes that AI theoretically performs up to 94 per cent of the computer and mathematics jobs, but only 33 per cent of them are actually used because of legal constraints, software conditions and the slowness of the manual review process.
This means that the most exposed group of posts is exactly where AI “can touch” and “have already touched”. The core elements of the posts in the writing buildings, which are paid decently and have window-to-window positions, often fall within the range of the model ' s best shot.
However, the full attribution of the downsizing to AI is an overly smooth narrative. In a research brief at the Oxford Institute for Economics in early 2026, it was made clear that there was currently no evidence that businesses were replacing their employees with AI on a wide scale; in many cases, AI was more of a statement that businesses were looking for layoffs.
The technology industry CEOs call AI “the perfect scapegoat”, “which makes things look not our decisions, our faults, but technological”.
In other words, real job contraction is the result of a combination of technological substitution, capital cycles and layoffs.
AI has emptyed a portion of the core of the work space and has provided the enterprise with a set of language that does not cause panic.
For those who sit on the posts, the complication of causes is not important.
The result is the same: the door is open, you have to go.
Inverted education.
After leaving the writing building, the first lesson is to retrain the job search itself.
Chen was lucky that he had a programming base and that the PLC training was five months later, with a base salary close to 10,000.
But the path of transition for more people is less smooth. A 22-year-old 2-year-old, Xiao Rui, specializing in economic management, graduated and went to a factory to make refrigerators.
He made a fatal mistake: he filled out his scientific history on the entry form. Since then, every criticism has been accompanied by a prefix to a degree: "You're a college student who can't do something as simple as that."
Fifteen people were on the waterline, seven or eight of them from the same village in Sichuan, and the remaining few were from the same village as their companions, only he was the “intruder”.
His scientific history was so obscurantistic among his fellow workers that he was dressed in the wrong clothes.
In fact, in the recruitment chain in a number of factories, intermediaries explicitly advise job seekers with undergraduate and above to “do not mention it” to the factory; in some cases, those with higher education are considered “unstable and easy to jump” and are more inclined to recruit students with higher education and below.
Some people on social platforms have learned their experience: if it is just a general job, it is better not to write undergraduate diplomas, just fill in a high school or college.
The repeated pricing of “human capital” in the writing building and the need to be actively hidden in front of the factory in order to qualify for one station-up water line is a denial by one evaluation system of another.
In the Internet industry, academic qualifications, curricula vitae and grades constitute a continuous ladder; in workshops, most of this ladder is not counted.
The factory needs stability, strength, discipline and the ability to do nothing wrong in the repetition. And this retreat, which was supposed to be the bottom card, narrows itself.
According to data from the Ministry of Human Welfare, the replacement rate for assembly, welding and packaging positions on traditional water lines has reached 72 per cent, with over 80 per cent in some sectors, such as automobiles; the number of Chinese manufacturing workers has dropped from about 200 million at its peak in 2008 to around 120 million today.
There are more than 500 “black light plants” in the country, and there has been an 80 per cent reduction in the number of workers employed in an electric enterprise.
The former operator of an axle bearing plant had been working for 23 years, and he was not on the list for the new line of production, as the entire plant had only nine timed inspections.
When white collars rush to the factory, they are not a static container waiting for labour to fill.
The container itself is being automatedly re-formed: the lowest-level transition posts are shrinking, and the remaining posts are increasingly demanding another capability, namely, to diagnose the complex malfunctions of highly automated production lines and become “a robot's nanny and doctor”.
The threshold for such posts is no lower than that for the writing building.
One square metre of “cells”
Chen Moe's companion, Tang Hai Town, went another way. After leaving his post on the Internet, he completed five months of PLC training and was officially enrolled in June 2026 at a minimum salary of close to $10,000.
In numerical terms, this was almost a successful skills shift: programming capacity was not wasted, but was replaced by conveyor belts and mechanical arms from browsers and servers. But behind the numbers is another life.
PLC engineers work in a highly homogeneous state: they travel to the factory throughout the year, are isolated, live in a small hotel and eat over a dozen roadside boxes.
An operator who moved on to the plane described the situation in the workshop as follows: the area of activity, on the side of the machine, was one or two square metres of space, had no access to permission, had to report to the toilet, had to be taken out, so he was afraid to drink more water all day.
When the equipment was in trouble, several people from the A side stood by and watched you fix it; when the leader shouted at you in the middle of the night, “to leave work before it is over”, it was not a crime to call it a “one square metre cell”.
But the group accepted the situation because they knew what they had changed back.
PLC does non-standardized products, customizing each line of production to client needs, and writing procedures represent only 30 per cent of the work, critical to the ability to debug and solve practical problems on the ground.
The job is “no eating age”, and experience is more subdivision than burden.
For a person who has been repeatedly squeezed by the 34-year-old red line in the Internet industry, this is enough to justify staying.
They are not looking for a better job, they are simply fleeing a state that could be “optimal” at any time.
The seat next to the conveyor belt is not decent, but at least it does not ask you to prove yourself faster every day than yesterday's model.
Who paid the difference?
If shot were to be pulled away, Zhou Yi, Chen Moo, Xiao Lu and Tang Hai would each be in different positions, a much more complex picture than the word “AI”.
An analysis by the World Economic Forum shows that employment among workers aged 22 to 25 years in high-exposure occupations has declined; data from Anthropic show that the share of junior employees in the technical sector has dropped from 21 per cent at the beginning of 2025 to 14 per cent, and that the share of new recruits has dropped from 27 per cent to 9 per cent.
The entrance ladder is being pulled off, while those at the bottom of the ladder are being asked to find their own way elsewhere.
Chen Moo paid for the transfer costs with unemployment compensation and a few months' time off; Xiaojia replaced a water line job with a hidden education; and Zhou Yi Zhou Zhou Zhou Zhi used a 30 per cent salary reduction in exchange for certainty that he was still recruiting.
Reskilled bills are ultimately settled by the individual on the basis of unemployment benefits, savings and downward salaries.
Enterprises that truly benefit from AI efficiency gains are not on this cost chain.
The deeper error is in time. When white collars are laid off, they are often at the most rigid age of cash flow: mortgages, childcare, parental care, each of which cannot afford a full training cycle.
The introduction of technical work in factories requires apprenticeships and internships, during which income may be zero. A former programmer in his thirties worked at night in a logistics warehouse and went to the public toilet every day to film the ash of the work clothes and stuffed them to the bottom of the backpack, without his family being present.
The conveyor belt won't ask him about his past KPI or how long he'll last. It only requires him to bend over, hold up the boxes, check the labels, repeat them.
Chen Moe is still waiting for his PLC course to be completed; Tang Hai Town has found a certain sense of pedigree in the workshop, even though the pedigree was replaced by a year-long presence and a square metre of space; Zhou Zhou Zhou Zhou Zhou Zhou Zhou Zhou Zhou has rarely mentioned the past in the administration of the new company.
They stand in different positions in the conveyor belt, pointing together to a fact that has not yet been fully discussed: the people pushed by the AI narrative to the factory door have not disappeared, they have simply fallen from an evaluation system to another, and there is no smooth convert rate between the two systems.
The conveyor belt, which does not read the curriculum vitae, does not ask for a degree and does not care which writing building the person came to work. It only cares if the next thing is in the right place.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.