Foxconn workers flow to AI labeling factory

Original source: Times Finance

Image credit: Generated by Unbounded AI‌

In the first half of this year, the technology circles of first-tier cities were activated by artificial intelligence.

Large-scale models such as Wenxin Yiyan, Tongyi Qianwen, and Light Years Away are sending waves of enthusiasm. The bigwigs with top-level resources are standing at the core of this grand event. They have sent out hero posts one after another. The battle for big-model talent is also on the table.

As a link closely linked with the AI industry chain, the Taiyuan Data Industry Base in Shanxi, 500 kilometers away from Beijing, is calm. Thousands of artificial intelligence labelers have gathered here. The topics they talked about stayed on the project progress, the rate of drawing the box to the standard, and three meals a day.

"The excitement is theirs, and we only have countless boxes." A data labeler told Times Finance.

In the memory of He Qing, the owner of the data labeling company, the excitement happened four or five years ago. For the first time, the spring breeze of artificial intelligence blew to this midwestern city. Sensitive businessmen began to draw up territory and recruit people, providing a steady stream of "nutrition" for artificial intelligence. Most of them are laymen of artificial intelligence, because of the sudden huge labor demand, they have a wonderful connection with cutting-edge technology.

"At that time, many bosses brought all their family members here, and they could make money by just moving their fingers." He Qing has heard a lot of exciting news in the data labeling industry-some people earn millions within three months, and others The orders that are grabbed can be queued to the second year.

But such good days are gradually fading away.

Decent "pipeline"

At 8:45 in the morning, a dense crowd of people blocked the elevator entrance. Only one-third of the people squeezed into the first elevator, and everyone's final destination was on the sixth floor.

The elevator door opened slowly, and the crowd spread out in all directions, and walked into offices where there was no difference. The space of about 100 square meters was filled with hundreds of computers. feet.

"As long as you follow the house number and ask one by one, it will all be marked with data." The vendor on the ground floor of the park described it this way.

This data labeling base, which has attracted nearly a thousand people, is like a hidden Internet cafe hidden in the park. People sitting in front of the computers skillfully click on the keyboard and mouse, and the desks of about one square meter are occupied by huge computers.

Source of data labeler at work: Photo by Times Finance

The only thing that can show their individuality is the colorful headphones worn on their heads. They have a common identity: data labelers.

The mouse clicked the left and right buttons back and forth, and the pictures on the screen zoomed in and out, and the cursor quickly drew frames of different sizes... After the repeated actions lasted for half an hour, Meiling twisted her neck slightly, and the bones in her spine creaked creaking sound.

"It's enough for a newcomer to get through the first week, and it's quick to get used to it if you're proficient." Mei Ling told Times Finance while still staring at the screen. As many as 30% of people gave up in the first week.

Every two weeks, Ms. Zhou, the foreman, will lead more than a dozen newcomers to start their apprenticeships. Such repetitive and boring work has dissuaded many young people.

Two years ago, Meiling transformed from a kindergarten teacher to a data labeler. In her hometown, Luliang, there are few jobs, and telemarketing is one of the more respectable destinations. Now, under the influence of the wave of artificial intelligence, data labelers provide another choice for women in the county.

Six months ago, due to the change of the labeling base, Meiling moved from her hometown to Taiyuan, the provincial capital. "Automatic driving or face recognition doesn't require the participation of large-scale labelers." She showed a proud look. In the eyes of her family, sitting in an office and operating a computer, with a monthly income of more than 3,000 yuan, the treatment has exceeded most of the county. Already working.

In 2005, Zhu Songchun, a computer vision expert, returned to his hometown of Ezhou, Hubei from the United States, founded the Lianhuashan Research Institute, and formed the earliest big data labeling team in China. Subsequently, data labeling factories gradually took root in second- and third-tier cities, and industrial clusters appeared in Hebei, Henan, Shandong, Shanxi and other regions.

Through repeated label training, artificial intelligence can reach the moment of "awakening". In Meiling's view, this is the same as the previous work of kindergarten teachers.

Data labeling is the first link in the birth of artificial intelligence products, followed by model training and optimization, model management, reasoning applications, etc. Feeding artificial intelligence products requires hundreds of millions of data, which will first flow to the computers of the "beauties".

However, Meiling's fantasy of "high technology" was shattered little by little by the repetitive sound of the mechanical mouse. She has calculated that 1,500 frames is the limit of the daily workload. Once this warning line is crossed, the eyeballs will be sore.

After get off work, even when facing the TV, what she sees is a mosaic mosaic, which looks like fuzzy pictures that need to be marked after zooming in.

"There are always unfamiliar faces in the next seat, and there is little communication between colleagues." After working for a year and a half, Wu Xia, who works in the same base, has not yet gotten used to the silence in the office.

After graduating from junior college, she originally entered the factory with her classmates, but because of project changes and classmates leaving, she became a "lone ranger". As soon as the work started, the office became a "workshop" where the automated assembly line started, with a cold industrial atmosphere and little human touch.

One of the characteristics of the data labeling industry is individual piece counting and no need for teamwork, which forms a management method different from that of ordinary white-collar workers.

Here, the labelers do not have a fixed position, but randomly assign hundreds of people to the flow direction according to project changes. The longest project is 2-3 months, and the short-term project is only 2-3 days. A project team of more than a dozen people has an administrator to keep an eye on everyone's work progress.

Annotators will not spend their energy on managing the relationship between colleagues. The piecework type pays attention to efficiency and concentration, and time and money are linked. To complete an average of 1,000 frames means that an average of 2 frames must be completed per minute.

"When you talk to others, you will lose a few boxes of money." Meiling said.

Foxconn workers flow to labeling factory

In the data labeling park, there are also scattered technology research institutes and entrepreneurial bases for overseas students. In Meng Ran's view, these "high-end" positions are far away from him.

Before entering university, he never left his hometown Linfen. After graduating from university, his family hoped that he would not leave the province. Two kilometers away from the base is Foxconn Taiyuan Science and Technology Industrial Park. This factory area absorbed the most active local workers. At its peak, nearly 60,000 people were active on the assembly line of the factory area.

No matter how hard he tried, college student Meng Ran’s job hunting radius never exceeded 5 kilometers. He once moved from the second phase of the data labeling base to the third phase; before officially becoming a data labeler, Foxconn next door was the place where he burned his youth.

The source of the Foxconn park less than 2 kilometers away from the data labeling base: Photo by Times Finance

Meng Ran once entered the factory for two consecutive vacations to make money, and each time he left in a hurry after receiving a salary of several thousand yuan.

Every winter and summer vacation, the entrance of Foxconn campus is full of college students with big bags and small bags, and everyone’s goal is to get the highest rebate and hourly fee in the whole year. "Everyone comes here to make quick money, and they pack up and leave as soon as the peak season is over. The factory is too busy to work, and it is difficult to stick to it for a long time."

Meng Ran didn't like the working atmosphere at Foxconn. Before entering the workshop, the electronic equipment must be handed in, and the only thing left to face every day is the crowd of rushing workers in similar clothes and the bleak and cold factory building. When you meet a grumpy team leader, it is common for you to be verbally abused every day.

With the roar of the production line starting, workers need to install a certain part continuously, and such actions often last for more than 10 hours. In a completely enclosed space, even trance is a luxury. Meng Ran didn't dare to have a few words with the workers around him until the foreman relaxed his management a bit.

In 2018, after the completion of the nearby data labeling base, Meng Ran had a second choice for his work. Just a block away, there's a more comfortable job at your fingertips.

Faye Wong used to be a recruiter for Foxconn. The factory's off-peak season and personnel changes, coupled with ambiguous rebates and frequent changes in income, make her often fall into endless conflicts with migrant workers. Annotators are a better choice for her.

"In the past few years, the data labeling threshold was low and the unit price was high. I could maintain a monthly income of 4,000 yuan, and the projects I did were all related to large factories, which was relatively secure." Faye Wong has seen many skilled workers leave the base to find another Out of the way, but back again in a circle.

Many annotators have similar work trajectories as Meng Ran. The work experience in electronics factories is the common point of their resumes, and the data annotation factory has become their next stop after leaving the electronics foundry.

The common features of large number of workers, considerable income, and simple operation have built a two-kilometer bridge virtually, connecting the two super factories together.

Disappeared projects and companies

For labelers, an intuitive feeling is that the good times are coming to an end.

The project with a unit price of a few cents disappeared, and the price of a label box was reduced to a few cents; the simple plane drawing point drawing box disappeared, replaced by a point cloud project that required multi-dimensional labeling; regular employees gradually left the project team, and the cost-effective Taller interns supported more than half of the workload.

He Qing, the owner of the data labeling company, has not been to the base for half a year, and she has gradually reduced her investment in the company.

Since the second half of last year, her team has never been able to receive projects with high customer orders, and the customer bill period has been delayed from three months to half a year. "Many small factories with insufficient cash flow and no ability to advance capital have closed down, and our team members have lost one-third."

Three years ago, Li Wei's enthusiasm was ignited by the callout frame. She was slow and not good at communication, and she felt that she had found a "chosen" job.

Li Wei took over the project with a unit price of 0.25 yuan. When the efficiency was high, she could draw 1,200 frames a day and earn nearly 8,000 yuan a month. "In order to make more money, someone bought a host and started work at home. If you become proficient, your income will increase.”

Like everyone else, Li Wei vaguely felt that the gold rush era was over.

The company has launched a brand-new project. What is presented in front of us is no longer a real-world road map, but a model map composed of thousands of green, purple, and blue points. A completed picture includes There are nearly a hundred marked boxes, and a set of questions is composed of dozens of pictures with only subtle differences.

Complicated operation interface diagram source: provided by the interviewee

"It is necessary to repeatedly switch between the plan view and 3D. Some blocked images have to be supplemented by brain, and the accuracy of the frame is also required to be controlled at 0.01 meters. The cost performance of the work is getting lower and lower." As long as the deviation from the required range is 1 mm more, They will be ruthlessly beaten back by the review.

Data, computing power, and algorithms are the three cornerstones of artificial intelligence. The greater the quantity and the higher the quality of data, the more mature large models can be trained, which is manifested in the work of annotators who are constantly improving their accuracy.

"The rules are being adjusted again in the past few days, and the accuracy requirement has been increased to more than 80%." Accuracy has become the "death spot" of labelers, and it is also a high-frequency vocabulary that appears when they complain.

A marked picture has to go through 2-3 steps such as review and quality inspection, otherwise it cannot enter the settlement cycle.

Sometimes, Wu Xia felt like she was trapped in a complicated maze, and she couldn't get out no matter what. She had been wracked by a new project for nearly a week—while submitting the questions, she was constantly called back, which made her fall into anxiety. "If the question is returned too often, it will be assigned to other people, and the previous energy will be in vain."

Meng Ran's anxiety was another kind. Since August last year, his work has become more relaxed. In the past 5 minutes, tens of thousands of data have been accumulated, and now there will be no load red line for half an hour.

"It may be that the amount of data on the platform has decreased, or it may be that the efficiency of machine review has increased." Meng Ran's sense of insecurity was quickly confirmed. Due to the forced reduction in his workload, his daily income dropped from one to two hundred yuan to a few hundred yuan. Tens of dollars.

A knockout race spread among the major agencies. Meng Ran has seen a team disbanded overnight, and more than a dozen employees who were owed wages sued the company to the labor bureau; if the situation was a little better, they would be transferred to the next agent along with the computer and employees.

"To be on the safe side, go to a team with more than 30 people." This is Meng Ran's advice to newcomers.

Annotator is exiting the stage of history

After a week of training and a half-month novice period, in May of this year, Xiaoting, who was in Hunan, finally adapted to being a data labeler, but witnessed the rapid decline of the company until its demise.

"After one month of employment, the company can't survive. The boss invites everyone to have a breakup meal, but the salary will have to wait for a few months." In Xiaoting's view, the current data labeling industry is full of "landmines", and the risk is far greater than income.

Whether it is a data labeling entrepreneur or tens of thousands of labelers, there is no way to avoid the fact that manual data labeling is gradually becoming insignificant on the stage of large-scale models.

What is different from Meiling's preschool teacher's job is that students will not take the teacher's job so quickly. Today, the large model technology nurtured by labelers is rapidly feeding back the data labeling process.

Taking Tesla as an example, it has continuously developed automatic labeling technology since 2018, from 2D manual labeling to 4D space automatic labeling. The advancement of technology has devoured the operating space for manual labeling. In 2021, Tesla's manual labeling team will exceed 1,000 people, and in 2022, more than 200 employees will be laid off.

Other car companies, including Xiaopeng Motors and Momo Zhixing, have also launched automatic labeling tools. Gu Weihao, CEO of Momo Zhixing, publicly stated that currently, to obtain lane lines, traffic participants and traffic light information, the cost of manual labeling is about 5 yuan per picture, while the cost of Momo DriveGPT is only 0.5 yuan.

In 2019, Wu Di, an AI data trainer in a first-tier city, had a premonition of the ceiling of his career. His company is responsible for developing the smart customer service project of the e-commerce platform. The progress was faster than he had imagined. In less than a year, the 10-person data labeling team he was in charge of was cut off, and only sporadic operators remained.

"The day the project continues to mature is when we are no longer needed."

The evolution of the large model is like a rushing river, always making a surprise attack at a certain moment, leaving the artificial team behind.

In a survey report by the University of Zurich in March this year, researchers found through actual measurements that ChatGPT's processing ability in 15 labeling tasks was higher than that of crowdsourcers.

At the beginning of April this year, Li Jie, a medical student at school, completed the text labeling of a large factory in the medical field within one month. This project will be used to provide intelligent diagnosis and dialogue services. This also made Li Jie feel the evolution of large models for the first time. speed.

"At the beginning, we kept feeding the platform with classified medical terms, and in the second week, the system was able to automatically realize the basic noun classification, and the accuracy rate exceeded 90%."

In Taiyuan, Shanxi, Ms. Zhou, the foreman of the base, began to persuade newcomers to take over more difficult projects, because it was difficult for the company to bear the pressure of projects being shelved again and again. "At present, the simpler the labeling business, the thinner the profit, and some projects will be yellow if half completed, and the labor cost cannot cover the project's income at all."

A recruiter in the data labeling industry told Times Finance that since this year, the recruitment threshold has gradually shifted from junior college students to undergraduate students. "In the past, there was basically no experience requirement for labelers. Now many companies hope that new employees can start working on projects directly, which can reduce the initial training costs."

At present, intelligent labeling can roughly capture the basic shape and position of objects, but in terms of accuracy, it still lags behind professional labelers.

No one knows when smart labeling will usher in a big explosion, but Li Wei is always accompanied by a sense of insecurity. Whenever she opens a new project page, the red box representing smart labeling always pops up first, as if reminding people in front of the screen all the time:

One day, it will take her place.

(The interviewees in this article are all pseudonyms.)