Has the era of educational big models completely arrived?
Recently, New Oriental released the HarmonyOS native version of its online app and launched the "Huixue System", an intelligent learning system for postgraduate entrance exams that loads big model technology. At this point, all education companies in China with a market value exceeding billions have released large-scale model products. New Oriental, with the largest market value and currently exceeding 80 billion yuan, was the last company to be announced.
What is more noteworthy than New Oriental's delayed arrival is that its first large-scale model product comes from the postgraduate entrance examination business, rather than the K12 business, which is more commonly used in the industry. Why has New Oriental's postgraduate entrance examination business become a pioneer in the application of big models?
For the industry, the addition of New Oriental has made the trend of technology application more apparent. Although the era of education big models has arrived, it is still in its infancy, and there are still several hurdles to overcome in its development.
Online automatic correction and proposition function
New Oriental is not considered a technology leading company, and now embracing big model technology has actually fallen behind the industry.
As early as July 2023, NetEase Youdao released the first education model in China, "Zi Yue". Afterwards, Homework Help, TAL Education, Xiwo, and Yuanfudao all released their own large-scale models, and some even developed more than one set of models. Among the education companies with a market value exceeding billions in China, Zhonggong Education, Doushen Education, and Huatu Education have all released large-scale model products. Top companies with a market value of less than 10 billion yuan, such as Gaotu, Chalk, Xueda Education, and Squirrel Ai, have all applied big models.
New Oriental has actually applied big model technology in its business. At the Alibaba Cloud AI Leaders Summit in April this year, Zhang Chen, the head of the P-end product department of New Oriental University, introduced that New Oriental applies big model technology to analyze the communication content between students and teachers, and then proposes communication improvement suggestions to enhance student satisfaction; Apply big modeling techniques to analyze courses and learning materials, and then break down and analyze knowledge structures to improve students' classroom efficiency.
Overall, the exploration in this stage is to apply big model technology to the internal production processes of the company to improve efficiency. Now, the Huixue system provides new features for students.
Firstly, automatic correction. It is reported that the AI correction function of the Huixue system supports objective questions for all subjects in the postgraduate entrance examination as well as subjective questions in English, especially essay correction.
Writing is an important single item for students to improve their grades, and correcting essays has become a concentrated requirement for candidates during exam preparation. The speed and quality of teachers' grading of compositions directly affect students' satisfaction.
Hong Zhe, General Manager of the Online Domestic Business Unit of New Oriental University Student Learning and Development Center, introduced that in the past, it used to take about 10 minutes for teachers to grade a mock exam paper, but now human-machine collaboration only takes about 2 minutes to complete.
Next is automatic proposition. The number of real exam questions for postgraduate entrance examination is relatively small, but candidates need a lot of practice and simulated exams. The big model technology can automatically generate simulation questions based on the core knowledge points of the discipline and the difficulty of the examination.
For example, in the Western medicine comprehensive exam subject with a limited question bank, New Oriental directly generated 4000 questions in the first batch, matching the entire subject's 6000 core knowledge points.
The challenge of this technology lies in balancing the accuracy and applicability of automatic proposition, that is, it cannot generate a large number of "water questions" in order to pursue the accuracy of the questions.
Hong Zhe stated that this requires teachers to enter annotated and layered knowledge points in the system in advance, so that the system can dynamically match different aspects and difficulty levels of knowledge points during automatic proposition.
Why did you first implement the postgraduate entrance examination business
Why did New Oriental take the lead in implementing large model products in its postgraduate entrance examination business?
Firstly, this is related to the internal business structure of New Oriental. In New Oriental's main education business, there are a large number of non subject training students, with approximately 484000 applicants in the latest fiscal quarter as of August 31 this year. However, non disciplinary training businesses are still in the early stages of development, with a wide range of categories and significant differences, which poses certain obstacles to the large-scale application of AI technology. Preparing for overseas exams is also an important business, but resources for practicing and testing TOEFL, IELTS, and other exams are already mature and abundant.
The postgraduate entrance examination business is also huge in scale. Li Lin, the director of the postgraduate entrance examination project at New Oriental University Student Learning and Development Center, introduced that the number of students in New Oriental's postgraduate entrance examination business exceeded 400000 this year, which is almost one tenth of the national postgraduate entrance examination registration number in 2025. Such a large user base provides application scenarios for big model technology.
Secondly, this is related to the demand of New Oriental's postgraduate entrance examination users. Li Lin introduced that in recent years, the diversified characteristics of New Oriental's postgraduate entrance examination users have been obvious. In addition to recent college graduates, the number of previous graduates and working population has significantly increased. The working population has accounted for 20% of New Oriental's postgraduate entrance examination business users, "Li Lin told 21st Century Business Herald reporters. The user profile can be further subdivided, for example, the working population can also be subdivided into groups such as "mothers" and those who have taken multiple postgraduate entrance exams.
The complex user base brings about complex user needs, and students' academic foundations, study habits, and schedules vary greatly, requiring the use of AI technology to provide customized and personalized services.
Thirdly, this is related to external competition in the postgraduate entrance examination training industry. In recent years, the market pattern of postgraduate entrance examination training has undergone drastic changes. Li Lin believes that in the past, the market for postgraduate entrance examination training was in an inverted triangle, with large national institutions having a larger market share. Now, it is in a trapezoidal shape, with good growth momentum for waist institutions.
The so-called "waist mechanism" refers to small and medium-sized institutions that perform well in the subdivision field of postgraduate entrance examination. For example, some institutions focus on art and MBA postgraduate training, with a small user base but annual revenue of over 100 million yuan. Even small training institutions specialize in providing training for candidates applying for graduate studies at a certain university.
These types of institutions are difficult to compete with large institutions in terms of brand influence and risk resistance, but they focus on differentiation advantages in vertical fields. Large institutions also need to leverage AI technology to provide customized and personalized services in vertical fields.
New Oriental's decision to start applying big model technology from the segmented business of postgraduate entrance examination is actually an industry trend. Comparable is the public examination training industry, where the three giants of Chalk, Huatu Education, and Zhonggong Education have all released large model products, but their application scenarios are different. Chalk focuses on answering questions and providing guidance, Huatu Education focuses on coaching during the interview process, and Zhonggong Education combines big models with digital human technology for teaching assistance.
An education industry insider told 21st Century Business Herald reporters, "For the educational application of large models, the more vertical they are, the more professional they may be, and the higher the efficiency
Model as application "is an inevitable direction
For the education industry, as all leading companies implement large-scale model products, the development path of education large-scale models is becoming increasingly clear.
Recently, at the 2024 Artificial Intelligence Big Model Industry Conference, the China Enterprise Evaluation Association and the Zhongguancun Smart Artificial Intelligence Industry Alliance jointly released the "2024 China Big Model Enterprise TOP50 Development Report". Among them, the Xiwo Big Model and the Xueersi Nine Chapters Big Model, two major education models, entered the TOP50.
The report points out that the technological level of large models is moving from billions of parameters to trillions of parameters. At present, the number of enterprises with parameter scales exceeding 100 billion has reached 25, and there are 7 enterprises with parameter scales exceeding one trillion.
The parameter level of the education big model is difficult to compare with the universal big model that has already reached the trillion level, which means that "model as application" is the inevitable direction of the education big model.
Most education companies do not need to develop base models, just as they do not need to develop their own cloud services. This standardized base service will only need to be purchased from a few technology giants in the future. What education companies need to do is to adapt and apply their own data on top of the base, "said the education industry insider.
For education companies, the future development of "model as an application" still needs to overcome several obstacles.
Firstly, choose ecology. Nvidia has a market share of up to 80% in the artificial intelligence chip market, mainly due to the development of a GPU based general-purpose parallel computing platform and programming model CUDA, which has built the foundation of Nvidia's software ecosystem, "said the education industry insider.
It is worth noting that the New Oriental postgraduate entrance examination big model product was first implemented in the HarmonyOS system. There are multiple reasons for this, including New Oriental's early investment in research and development within the HarmonyOS ecosystem, New Oriental's smart education business also developing a product line based on open-source HarmonyOS, and an increasing number of New Oriental postgraduate students reporting that they have migrated their phones to the HarmonyOS system.
But currently, Hongmeng Ecology is still in the early stages of development. To promote its large model products to a larger user base, New Oriental also needs to expand its product coverage to larger ecosystems such as Android.
As for the more directly related ecosystem of large models, it is still far from clear who will win and become the most widely used base model in the future.
Secondly, it is to reduce costs. The business model of "model as application" has already been implemented in the field of education, such as offering membership based subscription services, but the balance between input and output still needs to be explored.
Taking the postgraduate entrance examination as an example, this is a small market with an annual scale of only about 30 billion yuan. According to third-party research, less than 20% of users are willing to bear a unit price consumption of over 5000 yuan. Due to factors such as declining consumer power and a decrease in the number of applicants for postgraduate entrance exams, education companies need to carefully consider costs and benefits when developing and launching large-scale model products.
The education industry insiders mentioned above believe that this will create increasingly strong tension in future development. On the one hand, considering cost and practical application effects, low-cost and mature technology based large-scale model products can achieve cost reduction and efficiency improvement, and assist teaching; On the other hand, big model technology is considered to bring about significant changes that can disrupt the current education model. In this process, new technologies may need to constantly replace existing ones, which not only requires advanced investment but may also result in waste. However, it is a necessary stage in the development process.