Author | Periarthritis
Edit | Yunpeng
Tomorrow, the 2023 World Artificial Intelligence Conference will officially open, and it is reported that more than 30 large models will be unveiled, including the iFLYTEK Xinghuo cognitive large model, NetEase Youdaozi big model, Baidu Wenxin Yiyan big model, etc. In the context of the "100-model war", large models are also stirring up the education market.
Recently, the second Xiaodu Qinghe mobile phone passed the 3C quality certification, only more than 40 days after Xiaodu officially released the last mobile phone, in the current cold smartphone market, Baidu into the mobile phone has attracted the attention of the outside world. Baidu did not enter the game to "fight" with mobile phone manufacturers, but chose the smart education hardware track, and its first mobile phone is also a mobile phone equipped with AI learning large models. Prior to this, iFLYTEK has connected its iFLYTEK AI learning machine T20 Pro to the Xinghuo cognitive big model.
In addition to these smart education hardware products that have landed, NetEase Youdao also launched a spoiler video of the "Ziyue" large model and AI speaking teachers; Xueersi also said that it is self-developing the mathematical large model MathGPT; Homework Gang also began internal testing of the education model based on the Chinese market last month. Half a year after the launch of ChatGPT, players are frequently showing new actions in the field of education, which is suitable for one of the landing scenarios of large models. As Huachuang Securities pointed out, the large model with multi-round dialogue and multi-language understanding ability is naturally suitable for education.
Not only that, the education sector also continued to rise, on the first trading day after the release of the Spark Cognitive Model, iFLYTEK's stock price rose again, closing at 63.86 yuan, in addition, since the beginning of this year, iFLYTEK's share price has repeatedly hit a record high. Online education platform Gaotu rose 19.54% on June 8, a new one-day increase since March this year.
CICC pointed out in the research report that the market has continued to be high so far this year, and the last round of market catalysis mainly focused on the breakthrough progress of large models, and after the emergence of the model layer, the next stage of AI industry trends will be more catalyzed by the application side. Northeast Securities even mentioned that 2023 will become the year of the outbreak of AI+ education applications. Compared with the domestic situation, overseas online education platforms such as Duolingo and Quizlet have not only launched intelligent education assistants, but also explored a more mature business model.
In response to the impact that the big model may have on the education industry, the president of Tongji University said in an interview with the media that the large model represented by ChatGPT will change the way humans obtain knowledge, which used to be a binary structure of teachers and students, and today, machines are better than teachers in many fields, so the future teaching method may be expanded to a three-dimensional structure of teachers, machines and students.
So, which players have entered the game to promote large model landing education, how will generative AI empower the whole process of education, what problems and challenges will large model landing education face, we try to track the latest actions of various players in the current market to find the answers to these questions.
First, the upgrade of "AI tutoring", online education platform to promote the transformation of humanized tutoring
At present, three players, such as online education platforms, Internet technology companies, and AI startups, are participating in the landing of large models in the education scene. However, the most participating companies are online education platforms, which generally access the large models of other companies, optimize existing products, and rarely launch smart hardware products. Judging from the AI tutors and AI assistants launched by various companies, the future of online education may develop in the direction of human-like tutoring.
In February this year, a foreign online education platform called Quizlet launched Q-Chat, an AI tutor connected to ChatGPT. Subsequently, Khanmigo, launched by another foreign online education platform, Khan Academy, also connected to ChatGPT, and also has a function similar to AI tutoring. In fact, before accessing generative AI, some online education platforms have launched the so-called "AI tutoring" function, but most of those AI tutors in the past rely on recommendation algorithms to recommend recorded teaching videos and problem-solving methods, and cannot give answers to certain details in the videos and methods.
How to understand "humanized" AI tutoring? To put it simply, in the past, students could not use software to search for math problems, at most they could find answers with solution steps, and then there was an explanation video, but how much is the connection between the explanation video and the original question? Or can you clarify this type of math problem? This is something that previous learning software cannot guarantee.
And AI tutoring is not "an answer" or "an explanatory video". When students do not know a certain topic, the AI tutor does not give the answer first, and then guides the students to write by themselves, and when the students write incorrectly, the AI tutor will explain where the students will not. In the past, most AI tutors relied on recommendation algorithms to recommend recorded teaching videos and problem-solving methods, but could not give answers to certain details in the videos and methods.
It is worth mentioning that in addition to the AI tutoring function on the student side, Khanmigo launched by the foreign online education platform Khan Academy has also launched a "teaching assistant" on the teacher side, and after opening the "teacher mode", Khanmigo will guide teachers to make lesson plans, and at the same time help teachers collect materials, etc., thereby saving teachers' lesson preparation time and allowing teachers to spend more time interacting with students.
▲ AI tutoring Q-Chat launched by foreign online education platforms
Compared with the AI tutoring of all disciplines, the foreign language learning website Duolingo has launched a more segmented generative AI service based on its own brand characteristics. In March, it launched its Duolingo Max subscription, adding two new features, Explain My Answer and Roleplay.
Duelingo Max subscription service adds two new functions: Explain My Answer and Reloplay
Explain My Answer This function is the same as the literal meaning, when the user encounters learning problems and cannot understand the answer, the intelligent learning assistant will explain the problem to the user and provide users with more examples. In the past, online platforms without access to large models could not carry out such one-on-one concrete counseling.
Reloplay can be understood as the user interacting with various AI characters in Duolingo, so that users not only do not need to hire a speaking teacher to guide their speaking, but also do not need to go to the "speaking corner" to find different types of people to practice.
▲ Users use the Reloplay function to interact with AI characters
Xueersi is an online education platform that does not currently have access to other companies' large models. Recently, it announced the development of MathGPT, a large model for mathematics enthusiasts around the world. The leader of the AI team pointed out that in order to put large models into the field of mathematics, AI needs to have more accurate, clear and powerful logical reasoning capabilities, and the error tolerance rate is low, but the performance of general large language models cannot make breakthroughs at present.
Xueersi also made such a vivid analogy, "The universal language model is like a liberal arts student, it has excellent performance in tasks such as language translation, abstraction, comprehension and generation, but it has obvious deficiencies in solving and explaining mathematical problems." "More specifically, solving math problems is often wrong, and some math problems can be solved but the method is more adult-oriented, and cannot be adapted to the knowledge structure and cognitive level of school-age children."
At present, MathGPT has been led by Chief Technology Officer (CTO) Tian Mi, and the project has started corresponding team building, data, computing power preparation and technology research and development before the Spring Festival in 2023, and is ready to recruit outstanding AI experts from around the world to join.
Second, science and technology enterprises self-developed large models, "grabbing" the intelligent education hardware market
In addition to these online education platforms that have accumulated a large number of users in the "education track", another player who promotes "big model landing education" is Internet technology companies. These Internet technology companies have previously tried in the education track, and at the same time have their own advantages in the field of large models, and they have also launched intelligent education hardware products one step ahead of the online education platform.
The first Internet technology company in China to announce the launch of a self-developed large model and landing in the field of education is NetEase Youdao. At the end of March, it said it would release the "Ziyue" big model, and open up the AI speaking teacher and Chinese and English grading demos developed based on the "Ziyue" large model. At the 6th Artificial Intelligence Conference, NetEase "Ziyue" big model achievement AI speaking teachers will also make their debut.
▲NetEase Youdao AI speaking teacher spoilers
Although the progress of iFLYTEK and Xiaodu's large model is slightly behind NetEase Youdao, both companies have launched smart education hardware products.
Taking the iFLYTEK AI learning machine T20 series as an example, it has carried out large-model application in major disciplines such as language, mathematics and English, and launched the functions of Chinese and English composition personal correction, mathematics personalized and accurate learning human interaction auxiliary learning, and English language sparring functions.
▲iFLYTEK learning machine T20 series
In the past, when the learning machine corrected Chinese and English compositions, its functions were very limited, and it could only perform grammar correction, vocabulary, sentence editing, and example sentence recommendation.
After accessing the large model, this learning machine can be "transformed" into a real teacher, first of all, it depends on whether the student meets the writing requirements in the question, and the previous learning machine cannot understand the meaning of the topic, which can easily cause the problem of writing off topic. Secondly, the learning machine will also understand the content of the students' articles, and take the students to analyze the layout and expression of the articles.
Finally, the previous learning machine is more like a "scorer", and after carrying a large model, the learning machine is more like a "guide", which will take students step by step to find out the problems in their own articles, and then continue to train writing ideas. Finally, in the past, when students used the learning machine, they always wondered, "What if I can't write similar essays?" "Now with the addition of large models, the learning machine can generate recommended sample texts in real time to help students accumulate writing materials and better consolidate and improve."
▲Chinese Composition type correction function
Xiaodu, an intelligent technology brand under Baidu Group, has also recently launched a smart education hardware product equipped with a large model - Xiaodu Qinghe learning mobile phone. In fact, before Xiaodu, BBK, Xiaobawang and Hisense all launched learning mobile phones, but they did not make too much splash.
▲Xiaodu launched Xiaodu Qinghe learning mobile phone
Judging from the official information, this smartphone is equipped with a small AI learning large model, which can provide interactive AI lectures, multidisciplinary AI homework assistants, AI speaking practice and other functions. These functions are not much different from other enterprises that layout large-scale model landing education. But the matter of Xiaodu making a learning mobile phone is worth analyzing.
Why did Xiaodu fancy the field of education? A large part of the reason is that Xiaodu already has a relatively strong competitiveness in the education track. According to data from market research institutions IDC and Canalys, Xiaodu Learning Tablet has achieved the first place in China's learning tablet market in 2022, which means that Xiaodu has a certain user base and differentiated advantages in this education field.
Li Zegang, chief analyst of Omdia, said in an interview with the media that Baidu's hardware may still be based on its own AI capabilities, but only using smart phones as a carrier, and because he has not heard the news of Baidu's mobile phones from the supply chain, he estimates that Baidu mobile phone hardware development will still seek ODM (original design manufacturer) cooperation.
With the launch of smart education devices equipped with large models, it can be seen that whether it is iFLYTEK's learning machine or Baidu's Xiaodu Qinghe mobile phone, compared with traditional intelligent education hardware products, they can teach students how to think with the help of generative AI, making up for the lack of only providing reference answers.
More importantly, most of the intelligent education hardware products on the market are loaded with the original teaching and training content in hardware form, and in essence, these companies are still "selling courses". However, after the large model is connected to the intelligent hardware, these intelligent education hardware can combine the actual learning situation of students to recommend suitable difficulty goals for students, and truly achieve customized services of "thousands of people and thousands of faces".
In fact, in 2021, Lu Yun, former vice president of Tencent Education, pointed out in an interview with the media that the development of intelligent education hardware should rely on AI to solve problems, rather than relying on "selling courses" to solve problems. Only then there were no big models like ChatGPT.
According to Morgan Stanley industry data, the domestic smart learning device market is expected to reach 48 billion yuan in 2023, an increase of 20% over 2022. So how to understand this scale? Taking the smart wearable market as an example, according to the data of China Business Intelligence Network, the domestic smart wearable market is expected to reach 93.47 billion yuan in 2023, which is about twice the smart learning device market, which is enough to show that the current smart hardware market is relatively small. But why are some players still stake? A large part of the reason is: the education industry has a large market space, parents are willing to pay, and the business scene is clear.
Three. Embedded in private datasets, AI startups are committed to solving specific discipline problems
Although there is no user base like online education platforms, nor does it have the advantages in large models like Internet technology companies, AI startups such as Chain Reaction Technology and Xizhi Intelligence have embedded private datasets or added other models in specific subject areas after accessing ChatGPT, which improves the ability to solve specific subject problems, so they have also become important players in the process of "big model landing education".
When I was a student, a "slight" word in the math answer may disappoint many people, and Xizhi Intelligent directly turned the paper version of the answer into a two-dimensional beautiful girl AI tutor. When ChatGPT came out, the company connected the computable language model (CLM) in the original automatic problem-solving process with the large language model (LLM) to form a Joint Model model, so that the large language model provides NLP capabilities and CLM provides logical reasoning capabilities, so that the explanation of the two-dimensional beautiful girl AI tutor is greatly improved.
▲The function display of the AI tutoring App launched by Xizhi Intelligence
Some people may be curious, how did the image and voice of this two-dimensional beauty AI tutor do it? This has a lot to do with the aforementioned "access to the big model", because the big language model not only improves the NLP ability of the model behind this AI tutor, but also helps a lot in multimodal interaction.
In addition to making efforts in subject counseling, Aiwei Cat has also launched interest talent mining, time management and emotional management functions. The intelligent algorithm of iZhaocat will generate subject knowledge portraits, hobbies portraits and psychological portraits for students based on their personal computer interaction data, and generate quantitative analysis summary reports.
In the future, there may be more AI start-ups participating in the implementation of large models in the field of education, and more special functions will be developed, thereby reducing the serious problem of homogenization of intelligent learning software.
Fourth, the technical reserves are limited, the content sector is not strong, the funds are insufficient, and the large-scale model landing education is "obstructed and long"
However, in the current education of big models, online education platforms, Internet enterprises, AI start-ups and other three-way players still face different problems and challenges.
For online education platforms, in addition to the self-developed large model planned by Xueersi, other online education platforms such as Quizlet, Duolingo, Khan Academy, and Gaotu are all general large models that access other companies. However, the universal large model is not suitable for all scenarios.
Tian Mi, CTO of Xueersi, mentioned on the Douyin number: "The idea of a large model to do everything may not be the most economical solution, the model parameters are too large, and the cost of training and reasoning will increase by orders of magnitude." He believes that a general large model as an interactive interface, and then use the idea of Toolformer (a language model that can learn to use tools by yourself) to automatically call the corresponding domain model to obtain results according to specific tasks, which is more cost-effective.
▲ Tian Mi, CTO of Xueersi, talks about general large models VS field big models
However, if these online platforms launch self-developed large models in specific scenarios in a short period of time, it is not realistic, the killer feature of these online education platforms is high-quality teaching content and previously accumulated user groups, but they are not as good as those Internet technology companies with such technical accumulation in AI large models.
Can Internet technology companies rely on the "self-developed big model" and their own huge user groups to win the education track? Just two years ago, when betting on the education track was ByteDance. At that time, it announced that the new stone laboratory formed by the original hammer technology team was merged into the intelligent education hardware team led by the original founder Yang Luyu Musical.ly, and the merged hardware team will focus on the education field.
In September 2021, Guagualong, a brand of ByteDance Vigorous Education, announced the launch of smart writing boards, and three months later, the cumulative sales of Vigorous Intelligence exceeded 1 million units. In August last year, a person from ByteDance's education department said in an interview that ByteDance's business cut 80% of employees, and the size of the team became about 1,000 people after the layoff, and 50% of the intelligent hardware team of Vigorous Education left, and Vigorous Intelligent Learning no longer released new products. The important point is that ByteDance has not done any business in educational hardware before.
▲The powerful intelligent learning lamp launched by ByteDance's Vigorous Intelligence
Secondly, the current smart education hardware market is not very profitable, taking Shulang who has been "crawling" in the smart education hardware market for 23 years as an example, its gross profit margin in 2022 exceeds 22%, but the net profit is only 0.8%. This means that today, companies building smart education hardware can no longer focus solely on the device itself, but should focus on the content itself.
One of the key reasons why large models are popular with these technology companies is that they are eager to use AI to form a new content ecology, and then compete in the homogeneous smart hardware market.
At present, in the process of promoting large-scale model landing education, these Internet companies also need to accumulate intelligent education hardware, and at the same time ensure that the content section is high-quality and meets the real demands of users.
For the last type of AI startups, it does not have such a large user base as online education platforms and Internet technology companies, and it is also a problem in terms of R&D investment. AI startup projects are very expensive, from the US market research institute Trendforce can be seen that processing ChatGPT training data requires about 20,000 GPU chips, with the further development of OpenAI and other GPT models of commercial applications, its GPU demand will exceed 30,000. Moreover, the AI big model is still in its early stages, and it will be a long time before it is commercialized, which means that these AI startups will not be able to get a return in a short period of time.
At the same time, it should be noted that the AI large model has a lower fault tolerance rate in the field of education, and the physics teacher of the People's Congress Affiliated Middle School once asked ChatGPT to answer the 2021 Beijing College Entrance Examination questions, and the results showed that 5 physics questions answered 3 questions correctly, 10 biology questions answered 6 questions correctly, and 13 physics questions answered 8 questions correctly, which shows that the "wisdom brain" is far from reaching 100% correctness, which will also put forward higher requirements for the three-way players entering the game.
Conclusion: Big model landing education, or ushered in the "ChatGPT" moment
Under the wave of generative AI, online education platforms, Internet technology companies, AI startups, etc. are all promoting the landing of large models in the field of education. Judging from the various AI assistants launched by online education platforms, the big model is promoting the development of the previous "AI tutoring" to the direction of "human-like tutoring"; At the same time, the interaction ability of intelligent education hardware products connected to large models has also been enhanced, from the past "scorer" to the role of "guide"; Finally, whether it is the launch of the "two-dimensional beauty tutor" by Xizhi Intelligence or the interest mining function launched by Aizhao Cat, it can be seen that the big model is also promoting the development of education in the direction of personalization.
However, at present, these three players are also facing different problems in the process of promoting the landing of large models, which also requires online education platforms, Internet technology companies, AI startups, etc. to continuously improve their technical reserves, enhance their strength in educational content, and continuously promote the accuracy rate of large models in solving practical problems.