Summary of Key Points
This article, through the part-time experience of the author's friend "Dada," reveals the "hidden aspects" behind the development of embodied intelligence (AI robots that can perform physical tasks in the real world). To teach robots to perform real actions such as folding clothes and picking up cups, a large number of ordinary people are required to wear specialized equipment and repeatedly perform these tasks to collect data. Although this seems like a cutting-edge technology job, the hiring process is quite traditional: the barriers to entry are extremely low (no educational background or experience required), the pay is daily (200-250 yuan/day), the work is repetitive and monotonous, and during interviews, the size of the candidate's hands is even considered to ensure they can fit the equipment. The article concludes by pointing out that the progress of AI is actually built on the physical effort and time invested by countless ordinary workers, yet these very workers could potentially be replaced by the robots they help train.
Detailed Analysis
1. Why do AI robots need human "teachers"? What's lacking is "real-world experience"
Embodied intelligence is different from large models like ChatGPT, which can be trained using readily available text and images from the internet. However, for robots to perform tasks in the real world (such as serving drinks or folding clothes), what is needed is "physical interaction data"—information about how much force to apply with fingers, the trajectory of arm movements, and when to grasp objects. This type of data has never been collected on a large scale in human history and must be gathered from scratch. As of early 2026, there were only 500,000 hours of high-quality data globally for this purpose, while training a robot capable of household chores requires tens of millions of hours of data, representing a 20-fold gap. Therefore, humans are needed to wear the equipment and repeatedly perform these actions to convert this "experience" into a format that robots can understand—essentially acting as "mentors" for the robots.
2. What do the data collectors actually do? Two types of tasks:
There are two main roles for data collectors:
- Remote operation: People wear equipment and use game controllers to direct the robot's movements (e.g., sorting building blocks or folding paper cups). Every time a joystick is moved or a button is pressed, the robot's actions are recorded. It might seem like playing a game at first, but after a while, the wrists become extremely sore, and the robot may move in a manner similar to that of someone with Parkinson's disease.
- No robot guidance: Without remote control, people wear VR glasses and bracelets and repeatedly perform the tasks themselves (e.g., folding clothes or trousers). The actions must be performed identically over and over again, as if a "repeat button" has been pressed. Some collectors also work in different environments (such as McDonald's or supermarkets), staying in one place for two to three months before moving on to another.
3. How low are the barriers to entry for this job? No education required; interviews focus on hand size
When Dada applied for the job, he found that there were no educational background or experience requirements, and they didn't even need to know what embodied intelligence was. The interview process was quite unusual:
- Online group interviews: More than 20 people participated in a Tencent meeting, where HR asked about height and weight because the data collection gloves come in fixed sizes; they also asked if the candidate had ever used VR glasses (to ensure they wouldn't feel dizzy).
- High acceptance rate: Almost everyone passed the interview, except those who voluntarily withdrew. Dada compared it to the interviews for summer internships in factories during high school—as long as you're physically fit, you can apply.
In terms of pay, part-time jobs offer a daily salary of 200-250 yuan, while full-time positions come with social security but require more overtime and shift work. Many people choose night shifts to combine it with their daytime jobs for extra income.
4. Behind the cutting-edge technology is traditional labor?
The most striking contrast in this article is that the most advanced AI industries use the most traditional hiring methods. Dada's experience was very similar to working a part-time job in a factory: daily pay, repetitive tasks, low barriers to entry, and a high acceptance rate. For example, the workplace he visited had just been renovated (with a strong smell of paint), and the work involved collecting phones; he worked from 9 a.m. to 6:30 p.m. with almost no breaks. This model of "low-end labor supporting high-end technology" is not much different from the assembly line workers of the past, except that now the robots are being served by these workers, rather than manufacturing products.
5. The hidden concern about AI progress: Could the teachers of robots be replaced by the robots they train?
The article raises a concerning issue: These data collectors use their time and effort to teach the robots, but once the robots become capable of doing everything, their jobs might disappear. The development of large models, for example, also relied on countless people who annotated and cleaned data; currently, the underlying data for embodied intelligence is still being provided by these ordinary workers earning a daily salary of 200 yuan. However, as AI advances, these jobs could be automated, just as the article suggests: "We cheer as AI moves forward, but we don't know if anyone will remember that its first lessons were taught by people earning a daily wage of two hundred yuan."
Final Thoughts
This article shows us that AI doesn't become intelligent out of nowhere; every step of its progress is thanks to the efforts of ordinary workers. However, the future of these workers could be impacted by the technology they help create. This is both an inevitable consequence of technological development and a question we need to consider: how can we ensure that these "AI mentors" are not left behind as technology advances?