Written by Wang Huiying

Edited by Ziye

A wave of enthusiasm for large model technology triggered by ChatGPT continues.

Compared to looking up at the starry sky of technology last year, the core proposition surrounding AI large models now revolves around scenario implementation, that is, finding scenarios in reality where large models can truly shine.

In various industries, education has a natural fit with large models and is also a scenario that players in large models have high hopes for implementation. From this perspective, players in the education sector have also become the earliest batch of enterprises to deploy large models.

Starting last year, NetEase Youdao's "Zi Yue" large model, iFlytek's "iFlytek Spark" large model, TAL Education's MathGPT, Zuoyebang's "Galaxy" large model were successively launched, and this year Yuanfudao's "Kanyun" large model passed the filing... In the era of digital transformation in education, various large models are flourishing.

These enterprises all have profit demands, and large models + education are not only to empower technological change in the education industry but also to find new growth. Therefore, educational smart hardware, which is more noticeable for C-end users, has become a must-compete area for players.

For a long time, a big problem in education has been "teaching students in accordance with their aptitudes," and personalization, high quality, and large scale have been seen as an "impossible triangle" in the industry. However, the emergence of large models, through the generalization capabilities of AI technology such as generation and understanding, embedding teaching resources into large models, seems to be able to change parents' or users' needs for "teaching students in accordance with their aptitudes." Reshaping smart hardware using AI large models is also a direction for players to compete.

Essentially, the emergence of educational smart hardware is to help students improve quality and efficiency. Data released by iiMedia Consulting shows that in 2023, China's educational smart hardware market reached 80.7 billion yuan, an increase of 29.53% year-on-year, and it is expected that China's educational smart hardware market will exceed 100 billion yuan in 2025.

With the emergence of various smart hardware, the booming market has also brought fierce competition, with significant homogenization issues, low market recognition, and a lack of industry standards. It is also important to find a balance between technology and commercialization.

Undoubtedly, large models have injected new variables into the education industry, and industry players have all boarded the technological express train. How fast they can ultimately run depends on their comprehensive strength.

1. What directions are the various players focusing on in applying large models in education?

All industries deserve to be redone by AI. Since the advent of ChatGPT, this statement has been repeatedly mentioned and applied, and the education industry is no exception.

In the era of large models, education companies are eager to board this technological express train, solve industry pain points, and improve their strength, leading to a "Hundred Models War" in the education industry.

The earliest layout was made by technology company iFlytek. In May last year, when iFlytek released the iFlytek Spark cognitive large model, it also released educational products based on this large model. According to iFlytek, "large model + AI learning machine" allows AI to grade compositions like a teacher and have realistic conversations like a spoken language teacher.

In addition to general large model players like iFlytek, vertical players in the education sector have also rushed into the market to launch vertical large models in the education field.

In July last year, NetEase Youdao launched the country's first educational large model "Zi Yue" and upgraded it to the "Zi Yue" 2.0 version in January this year, using search enhancement technology to significantly improve the accuracy of model answers and reduce hallucination issues. In September last year, Zuoyebang released its self-developed Galaxy large model, which is a tailor-made educational large model covering multiple subjects, stages, and scenarios according to the official introduction.

Almost simultaneously, TAL Education chose a vertical large model specializing in the subfield of mathematics, "MathGPT," which mainly targets global math enthusiasts and research institutions, with problem-solving and lecturing algorithms as its core. When users use MathGPT, they can upload math problems in text or image format to receive conversational answer feedback, or they can generate random math problems and receive solutions from the system by pressing the "Randomly Generate a Problem" button.

Not long ago, Yuanfudao also made moves in the direction of large models. On May 15, "Netinfo Beijing" announced new progress in AI filing, and Yuanfudao's Kanyun large model passed the large model filing.

One significant shift is that if last year various companies were competing in "modeling" technology, this year, after the upgrade of various large models, the competition point has been placed on the implementation and application level of large models, as large models have now entered the "model-as-application" era.

The consensus in the industry is that education is an industry with special needs, requiring high interactivity and personalization. In educational scenarios, large models do not play the role of a "tool person" providing correct answers but are more like teachers imparting knowledge and dispelling doubts. Through companion learning and growth, AI can solve the challenge of "teaching students in accordance with their aptitudes" in education.

Therefore, companies have unanimously placed the focus of applying and implementing educational large models on personalized analysis and guided learning scenarios.

For example, from the AI scenario test video released by Yuanfudao's large model, we can see that after students submit their homework to the large model, AI immediately analyzes the homework situation and points out possible errors. Then the large model guides students step by step through questioning, pointing out errors that were not discovered during the problem-solving process, and guiding the child to arrive at the correct answer.

Yuanfudao Large Model AI Scenario Test, Source: Feixiang Planet Video Channel

NetEase Youdao focuses on two scenarios: English learning and family tutoring. Along with the upgrade of Zi Yue 2.0, NetEase Youdao released three AI applications, namely AI homework tutor "Little P Teacher," Youdao Speed Reading, and virtual human speaking tutor Hi Echo 2.0. Last year, during the Zi Yue 1.0 period, NetEase Youdao launched "LLM Translation," "Virtual Human Speaking Coach," "AI Composition Guidance," "Grammar Intensive Course," "AI Box," and "Document Q&A."

iFlytek aims at the application scenario of smart classrooms. Earlier this year, Spark V3.5 was officially released, and iFlytek launched products such as smart blackboards, smart classrooms, personalized learning manuals, and English listening and speaking classrooms. Among them, Spark Smart Blackboard has four major functions: multimodal understanding and recommendation, fully natural interaction, virtual human assistant learning, and intelligent recording and sharing.

Zuoyebang is more inclined towards writing scenarios. Currently, the Galaxy large model supports AI problem-solving, multi-language AI Q&A, and other capabilities, boasting proficiency in poetry, prose, and textbook knowledge; it also supports AI writing functions that can be used to improve writing skills, optimize writing structure, and provide article refinement, grammar correction, and creative inspiration.

This year marks the first year of large model application and is also a crucial year for the implementation of vertical large models in education. For a long time, one-to-many has been the norm in education, and teaching students in accordance with their aptitudes has become the ultimate goal of the education and training industry. We cannot be sure how much the era of large models will change, but we can be sure that it is already on the path to change.

2. AI large models reshape smart hardware, becoming the direction pursued by players

Analogous to 3C electronics, which have gained new imagination and competitive dimensions after being reshaped by large models, educational smart hardware is also being upgraded and replaced by large models.

In fact, educational smart hardware has a history of over 30 years in China, from the initial English dictionary Wenquxing to the Bubugaoxi series of learning machines and point-reading machines from BBK, to smart learning machines under internet technology. The educational smart hardware track is constantly evolving with technological progress.

With the explosion of AI technology last year, various educational enterprises focused their attention on smart hardware while studying large model technology.

The emergence of large model technology has released greater market opportunities for educational smart hardware.

Among them, learning machines with the most complete functions and the richest applicable scenarios are the first choice for parents. Under the influence of large models, learning machines have also become one of the most obvious categories of educational smart hardware that AI empowers change.

According to a report by iiMedia Consulting, nearly 70% of Chinese consumers attach great importance to the AI functions of smart learning machines, and over 50% of respondents attach great importance to the learning resources of smart learning machines.

The combination of market demand and AI technology upgrades has attracted many educational manufacturers and technology companies to deploy.

iFlytek is considered one of the first technology companies to benefit from AI learning machines. As early as 2021, iFlytek announced that iFlytek Smart Learning Machine would be renamed iFlytek AI Learning Machine and released its AI Learning Machine T10 on the spot.

Last year, after the release of the iFlytek Spark cognitive large model, iFlytek's first AI learning machine T20 Pro equipped with a large model was launched. Relying on the cross-domain multi-task human-like understanding and generation capabilities of the large model, it launched three major functions: AI 1-on-1 math interactive tutoring, AI 1-on-1 English speaking practice, and AI 1-on-1 Chinese and English composition grading.

iFlytek AI Learning Machine T20 Pro, Source: iFlytek AI Learning Machine Weibo

More importantly, the emergence of large models has brought new growth to smart hardware products. According to iFlytek's 2023 interim report, the GMV of iFlytek AI learning machines equipped with the Spark cognitive large model increased by 136% and 217% in May and June, respectively.

On the other hand, companies specializing in education, such as TAL Education and NetEase Youdao, have also upgraded their learning machines with the empowerment of large models, leveraging their own educational resources and experience.

In January this year, NetEase Youdao launched the AI learning machine X20 equipped with the Zi Yue large model, with family tutoring as its main selling point, and also incorporated AI applications such as Little P Teacher, Hi Echo 2.0, and Youdao Speed Reading.

As NetEase Youdao CEO Zhou Feng said, in a downturn cycle, the biggest opportunity for smart hardware lies in technological innovation, and large models are expected to drive hardware sales growth.

TAL Education launched the TAL Learning Machine equipped with the Jiuzhang large model, whose biggest highlight is its self-developed intelligent system and curriculum content, providing users with customized learning plans. Yuanfudao also released its ink screen tablet Xiaoyuan Study and Practice Machine last May, providing an AI study and practice integrated solution to help improve learning efficiency. Zuoyebang has successively launched learning machines, learning pens, dictionary pens, and other products equipped with the Galaxy large model based on its original error printer.

Reflecting on market data, with the addition of AI large model functions, personalized and customized features have spurred more demand, also helping educational smart hardware move towards high-end.

According to data from Runto Technology, in the first quarter of 2024, the online sales of Chinese learning tablets (learning machines) in all channels (including Pinduoduo, Douyin, and Kuaishou platforms) reached 689,000 units, an increase of 79.9% year-on-year. At the same time, ultra-high-end learning machines priced above 6,000 yuan are being rapidly accepted by the market.

For a long time, the emergence of educational hardware products has been to help students learn with higher quality. Hitting the market demand, hardware products have long become an important revenue pillar for companies involved in the education sector. However, in recent years, along with the consumption cycle of electronic products, the market has fallen into a downturn period, and now the emergence of large models has pushed educational smart hardware products onto a fast track.

However, it is undeniable that with many entrants and fierce market competition, a mature educational system, high-quality educational resources, and advanced large model technology are all indispensable for doing well in educational smart hardware and grabbing a slice of the pie.

3. The era of "intelligent emergence" has arrived. What is the prospect of technology and commercialization going hand in hand?

The emergence of large models has not only sparked a "Hundred Models War" inside and outside the industry but has also opened a new era of "intelligent emergence."

In the past, artificial intelligence was about "teaching machines whatever skills humans want them to learn." However, the intelligent emergence initiated by ChatGPT has opened a new paradigm of "human-like" interactive learning through natural language, achieving "skills that humans have not taught, but machines can learn."

After more than a year of development, compared to competing technically, manufacturers' competition points have long advanced to commercialization itself. In other words, players not only face technological gaps but also how to cover large model R&D costs through scenario implementation and commercialization.

Among them, the natural particularity of the education sector puts higher demands on industry players.

First, from the perspective of scenario application, education itself is a scenario that requires high rigor, scientific rigor, and accuracy of content, which poses a challenge to manufacturers in the development process. Many manufacturers are working hard to improve big data quality, optimize training algorithms, and establish verification and validation mechanisms.

At the same time, it is necessary to continuously iterate technical capabilities and prepare for long-term investment. "When we keep up with the international cutting-edge level in general capabilities, we can achieve breakthroughs in specialized fields such as education, automotive, and healthcare," said Liu Qingfeng, Chairman of iFlytek.

Looking at the market, most manufacturers extend the application of large models in their advantageous vertical fields, trying to create differentiation. For example, iFlytek and NetEase Youdao have accumulated profound expertise in machine translation, text image recognition, and other areas, and their large models have ultimately derived spoken language teachers in product form. Yuanfudao extends to scenarios such as homework grading based on resources like Xiaoyuan Search and Yuan Question Bank, while TAL Education's MathGPT focuses on the field of mathematics and excels in mathematics and scientific reasoning.

Compared to differentiated scenario applications, from the perspective of smart hardware products, with more and more players entering the market, players face the challenge of homogenization.

A survey conducted by iResearch showed that the three most important dimensions for parents when purchasing educational smart hardware products are product functionality, educational content, and eye protection design.

Taking AI learning machines as an example, most learning machines on the market currently focus on functions such as companion learning, family tutoring, and AI teacher interaction, and eye protection mode is also a must-have option. Although AI learning machines from different brands may differ in some details, they generally lack innovation and differentiation.

Perhaps the only way to demonstrate differentiation is through educational content, that is, the accuracy of AI functions and the quality of built-in learning resources under large model applications, including the age range and subject scope of applicable content, resource richness, whether it is exclusive content, and whether it can be continuously updated.

Essentially, combining AI technology with products on the smart hardware side is not difficult, but what is truly