In an office at dusk, a stack of exercise books sits on the corner of the desk, and a red pen leaves the wrist sore after hours of marking. This is an evening many teachers know well. Today, feeding a whole class's homework into a machine takes only a few minutes: scanning, grading, leaving marks on the original pages and generating a report. The time saved ought to go to the students, not to repeated red ticks.
This scene took place in Wuzhen, Zhejiang. On November 2, the Light of the Internet Expo of the World Internet Conference opened, under the theme of intelligent connectivity and an intelligent future. According to Xinhua News Agency, exhibitors from many countries and institutions took part, and more than a thousand new technologies were on show. iFlytek moved a classroom and a clinic, just as they were, in front of visitors.
A Classroom Moved onto the Stand
What drew parents most was the smart grading machine. Staff put a pile of homework in, and the machine scanned it quickly. Multiple-choice and fill-in-the-blank questions, even geometry proofs and essays, could be annotated step by step, with marks left on the original pages. According to China Science Daily, it can grade a whole class's homework in as little as five minutes.
Even when the order is shuffled, it matches each student accurately, without changing the old habit of writing on paper. Once grading ends, the system automatically produces a class learning report, showing at a glance which questions were missed most and which knowledge points were not grasped. Teachers are spared the manual tallying, so review can happen the same day and mistakes do not sit overnight.
Coverage Across Zhejiang's Schools
The novelty on the stand only counts once it reaches daily life. According to China Science Daily, in Zhejiang, products such as smart classrooms and the smart grading machine already cover 90 districts and counties and more than 5,000 schools across the province, serving millions of teachers and students. This is not a prototype; it is infrastructure used in class every day in many schools.
Here a caution is in order. The figure of more than 5,000 schools refers to one province, Zhejiang, and to the caliber disclosed at the Wuzhen event at that time. Widen the lens to the whole country and the company, at the World Artificial Intelligence Conference in Shanghai the previous July, gave a figure of more than 60,000 schools serving over 100 million teachers and students. One is a province, one is the whole country, and the time points differ too, so the two cannot simply be added together.
For ordinary families, the change is quite plain. Children do a few fewer repetitive exercises, teachers lose a few fewer nights to marking, and a class in the mountains can share a lesson with one in the city. Whether the technology is any good is not shown by how flashy the launch event is, but by whether that screen in the classroom lights up as usual the next morning.
Why the Numbers Do Not Match
The numbers on the medical side are easier to misread, and need to be taken apart more carefully. The Wuzhen event disclosed two calibers: Xiaoyi, a health assistant aimed at individuals, has handled more than 220 million AI consultations; Zheli Zhiyi, rooted in 90 counties and districts of Zhejiang, has delivered 461 million instances of assisted diagnosis and treatment.
The two differ greatly. One is health Q&A on a phone, the other is an assistance system that lends a hand to doctors in grassroots clinics; one serves users nationwide, the other counts only Zhejiang. Adding 220 million and 461 million together, or treating them as the same thing, would be a misreading.
There is a third caliber. According to a disclosure the company made at the earlier World Artificial Intelligence Conference in Shanghai, Zhiyi Zhuli, an assistant for grassroots doctors, has completed 1.23 billion assisted diagnoses nationwide, serving more than 300,000 doctors. This is a nationwide figure, for a different product, at a different point in time; it must not be mixed into one number with the two Wuzhen sets.
That the three books are kept separately shows precisely how far AI has reached into different links of the chain. The roots of this capability go back further: according to China News Service and iFlytek Healthcare's official website, in 2017, Zhiyi Zhuli, developed by iFlytek together with Tsinghua University, passed the written part of the national medical licensing examination with 456 points.
The score was more than 90 points above the passing line, and drew considerable attention at the time. The developers stressed repeatedly that the robot played only an assisting role in diagnosis and treatment, that it would not take the place of doctors in seeing patients independently, and still less replace people. That restraint was later written into the product's positioning: offer a basis for judgment, do not make the conclusion for the human.
From the Clinic to the Mountains and Islands
Bring the lens back to Zhejiang. A shortage of doctors in mountain areas and on islands is an old problem. According to China Science Daily, Zheli Zhiyi has reached 90 counties and districts, extending assisted diagnosis and treatment to grassroots clinics: organizing health records before a visit, prompting on key diagnostic points and helping write medical records during it, and following up afterwards.
It does not make the final call for grassroots doctors; it puts the easily missed clues in front of them. When a case is hard to judge, the system offers a basis for judgment and the doctor decides using experience. In places short of staff, one more pair of tireless eyes often matters more than it would in a big city.
On the consumer side is Xiaoyi in the phone. If you cannot read a check-up report, are unsure which department to register for, or want to ask about medication contraindications, you can put the question to it first. It catches these scattered, high-frequency little questions, extending the patience hospitals have least of to beyond the clinic.
The Technical Base Behind the Scenes
What lets these applications run is a less visible layer of support underneath. According to China Science Daily, iFlytek, on the Feixing platform, a domestically built ten-thousand-card computing base, completed an upgrade of the Spark large model X2.6, supporting a one-million-token context window and markedly improving training efficiency for ultra-long contexts.
Closer to daily life is device-cloud collaboration. The on-device model decides for itself: small matters involving privacy stay local, complex tasks go to the cloud, and where cooperation is needed the two work together. As the company disclosed at an earlier developer festival, device-cloud routing accuracy has risen from just over 80 percent to 96.1 percent.
Only Essential Scenarios Count as Real Adoption
The four-day exhibition is short; the stands come down, but the story does not end. The next morning, teachers light up the blackboard as usual, village doctors open the assistance system as usual, and engineers on night shift watch over computing power no one can see. Real change never happens under the spotlight, but in these repeated everyday moments.
For ordinary people, the standard for judging whether an AI is useful can be very simple: has it entered those unavoidable, draining, essential tasks? Marking homework, reading scans, reviewing bid documents, driving long distances all belong here. If it saves time and prevents one mistake, that is worth more than memorizing a few more specs.
Of course, the wider the rollout, the more clearly the boundaries need to be stated. Assisted diagnosis is assistance after all; the doctor still makes the final call. The grading machine offers a reference on class learning; how to explain, how to coax a child, still rests with the teacher. Freeing people from repetition, rather than erasing them from the process, is the bottom line this round of adoption most needs to hold.
A stack of exercise books and a machine that never rests: what lies between them is more than a few minutes. What it measures is that technology has finally been willing to bend down and catch the most trivial, most wearing, and most overlooked parts of daily life. That is probably the plainest shape AI takes when it enters ordinary life.
Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.