Source: Tech Planet

Image source: Generated by Wujie AI

Zheng Wen still remembers an afternoon a few months ago when she earned only 20 cents in an hour. She graduated from a vocational school in Hunan and works as a large model data annotator, with a daily task of adding labels to the raw data she receives, such as images, videos, and text.

However, large models have high requirements for data quality. That day, a picture had to be modified 8 times before it was approved, taking a whole hour. This means that she only earned 20 cents in that hour, while under normal circumstances she could earn 12 yuan by annotating 600 frames. "Money is not easy to earn," she emphasized repeatedly.

This is almost a consensus among all data annotators. On one end, these workers earn less than 5000 yuan per month, serving as the cornerstone for the construction of large models like an army of ants. On the other end are the AI dreams of internet giants, hoping to surpass Chat GPT 4.

Data annotation uses a piece-rate system for wages and does not involve office politics. The only concern is the overly monotonous nature of the work, making it difficult for most to persist for 3 months. Almost everyone told Tech Planet, "You'd better not go."

What they don't know is that before long, most of them may be replaced by AI in this monotonous work.

From 50 cents to 4 cents, prices plummet

In 2017, Lin Shuang made a "quick buck": over 6000 yuan in 15 days. For Lin Shuang, who graduated from a vocational school, this income was quite substantial. At that time, people had high expectations for AI, and almost no one doubted its future. All investment institutions firmly believed that billion, tens of billions, or even hundreds of billions of enterprises could be born here.

Almost all AI technologies are based on competition in algorithms, computing power, and data. The huge data is the underlying factor for the quality of the technology. Programmers with bright backgrounds sit in offices in "Beijing, Shanghai, and Guangzhou," outlining AI blueprints through code iteration, while vocational school graduates, housewives, and others in third- and fourth-tier cities handle massive data packages containing images, text, and voice in cubicles.

ChatGPT is no exception. An employee of Baidu's Wenyin Yiyu project team said that large models themselves do not have any new technology or high technical barriers. The key issue is the parameter barrier formed by the computing power barrier.

Data annotators in the era of large models are not significantly different from before. The few differences may be a more comfortable working environment and higher requirements for annotation quality. An industry practitioner introduced to Tech Planet that when they first entered the industry, they would form a team of about 10 people, with one person responsible for quality inspection. If it did not pass, the employee would have to redo it. The quality of the data determines the quality of the large model.

Data laborers do not care about new branches of AI technology; they are more concerned about the unit price, as this is piece-rate pay.

"At that time, when the unit price was high, I could earn over 10 cents for drawing a 2D frame. The most I earned was over 600 yuan in a day after working over 10 hours," Lin Shuang recalled. However, this was not the highest. Another annotator said that the price for early 2D frames could reach as high as 50 cents.

Drawing frames is a common operation in data annotation, where annotators mark objects in images, such as vehicles, traffic lights, and obstacles, according to requirements. Drawing frames is divided into 2D and 3D, with the latter being more expensive.

However, this heat did not last long. With more and more people entering the industry and the overall development of the AI industry not going smoothly, the unit price for annotating a picture has become lower and lower. Lin Shuang said that the lowest price now is only 4 cents.

"If it's drawing frames, the average unit price in the industry is around 0.15 yuan, but it depends on the project. If you can get orders directly, the minimum requirement for a first-hand order should be around 100 new employees, which is quite large in scale. The 3D frames could possibly reach 30 cents each, but it's rare to reach 50 cents."

Of course, if you have professional knowledge in fields such as medicine and finance, the unit price will be higher. For example, many medical large models require annotators to have clinical expertise and relevant work experience.

Most practitioners earn less than 5000 yuan per month, but there are also a few lucky ones. Yang Shuo used to run a clothing store in Sichuan, but the pandemic affected his business. He transitioned to data annotation for large models this year and now earns 8000 yuan per month. "I signed a contract with the company and paid a joining fee of 9500 yuan. The contract stipulates a minimum monthly income of 7000 yuan."

Who is making money after all

Companies like Alibaba, Tencent, ByteDance, as well as automakers like SAIC and Lynk & Co, are the sources of data annotation business distribution. In order to obtain orders directly from the source at the best price, data annotation companies need to have a certain scale.

An employee of a data annotation company told Tech Planet that they receive orders directly from large companies, but these companies require them to have 500 employees. Therefore, they choose to meet the personnel requirements through franchising or setting up subsidiary companies.

The difference between the two is that franchising is suitable for newcomers to set up studios. If a subsidiary company is established, there is generally only one in a region. A novice studio needs to pay a franchise fee of 25,000 or 30,000 yuan. A subsidiary company is the exclusive agent for a region and needs to pay a fee of 50,000 yuan. They can guarantee sufficient orders within three years and are responsible for technical training within three years. These studios or subsidiary companies form a large union, ranging from hundreds to thousands.

The above-mentioned employee of a data annotation company said that the popularity of large models has once again pushed the data annotation industry into a boom, with almost someone visiting their company every day.

However, running a data annotation company is not easy. Data annotation companies will tell you that the first 1 to 2 months in this industry are difficult because employees need a learning curve. Initially, only 5-8 people are needed, and even middle-aged women in their 40s have no problem.

Stability is the most important factor for data annotation companies or studios. However, most of the data annotators contacted by Tech Planet often "speedily" leave within 3 months due to the lack of interest in the work. New employees cannot immediately start practical work, and the high turnover results in unstable data annotation quality and cycles. Data annotation studios prefer to recruit cash-strapped housewives.

"Part-time work is definitely not feasible; there will be downtime. With rent and computer investment, you will lose money. The best way is for all employees to work full-time," said Wei Ming, who has run a data annotation studio, to Tech Planet.

The repayment period for most data annotation companies starts at 3 months and can be up to half a year, but they need to pay employees on a monthly basis, requiring a certain level of financial reserves. "For one person, 3500 yuan; for 100 people, it's 1.05 million in 3 months."

Zhang Jian once joined a union with over 200 employees. In the first year, they caught the industry's boom period, with the unit price for 2D frames reaching as high as 50 cents. That year, his union made over 4 million yuan.

However, in the second year, the market took a sharp downturn. The unit price for annotation decreased, employee turnover increased, and downtime increased. In addition, two major projects did not settle, and after a whole year, they lost over 3 million yuan. "The bosses all said they would absolutely not touch data annotation in the short term," Zhang Jian said. "They are currently in a legal battle with the upstream."

This is a low-profit business. Haitian Ruisheng is currently the first listed company in the data annotation industry. Last year, the company had a revenue of 263 million yuan, with a profit of only 29.45 million yuan and a net profit margin just over 10%. However, in the first half of this year, due to a decrease in the number of customers, the company fell into a loss.

"Screw" that could be replaced at any time

Relying on the accumulation of Kenyan workers' ant-like efforts, OpenAI's language dialogue large model capabilities have emerged. These ordinary people, known as data laborers, have supported Sam Altman's (OpenAI founder) AI dream. However, most of the work in their hands will soon be replaced by new products they have helped create, if nothing unexpected happens.

In foreign countries, Anthropic, founded by former OpenAI employees in 2021, has raised $5.15 billion this year, more than 7 times the total amount raised in the past two years. The company provides a new method to train models with less human involvement.

This year, the AI startup refuel launched an open-source tool called Autolabel, which can annotate datasets using mainstream large models available on the market. The company's test results claim that Autolabel's annotation efficiency is 100 times higher than manual annotation and costs only 1/7 of the manual cost.

In China, a company called Vision Future is also developing annotation large models. In an interview, they stated that some projects have already been delivered using GPT, achieving an accuracy of over 80%, close to manual annotation.

However, Haitian Ruisheng believes that AI will never achieve fully automated annotation because if machines want to continue to evolve and become closer to human judgment and understanding, they will definitely need humans to guide them.

Almost all individuals who have been involved in data annotation have revealed the same viewpoint to Tech Planet: data annotation is a job with no threshold, requiring only proficiency in using a computer.

In fact, if simple annotation can be done by AI, human involvement will shift to more challenging data selection and standard work, which also means that the industry's threshold will continue to rise, especially for large language models like ChatGPT and Wenyin Yiyu.

As a comparison, even before ChatGPT became popular, OpenAI had assembled a dozen doctoral students to "annotate." Baidu's data annotation base in Haikou has hundreds of full-time large model data annotators, with a 100% undergraduate rate among annotators.

The characteristic of these large language models is that annotators need to have a certain knowledge base and logical analysis ability. According to "Caijing Eleven," annotators need to assess the type of question, then score and rank 5 answers, with a score range of 0-5. If the score is below 3, they also need to annotate the specific reasons, such as "off-topic (0 points)," "severely off-topic (1 point)," "logical problems, factual errors, a small proportion gives 2 points," and so on.

Another popular area for data annotation is autonomous driving. According to a report by Deloitte, the annotation demand in the autonomous driving field accounted for 38% of the entire downstream AI application in 2022, and is expected to rise to 52% by 2027. Compared to large language models, the educational requirements for simple drawing operations in the autonomous driving field are still relatively lenient.

Annotators are the cornerstone of the transition from the mobile internet era to the artificial intelligence era. Most of the practitioners contacted by Tech Planet are unaware of the changes that AI will bring to them and are unaware of the contributions they have made to the development of AI. They are just the new generation of screws in the internet era, and they could be replaced at any time.

(Note: All names in the article are pseudonyms.)

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。