The "sprint for college entrance examination" of the big model, the education industry has taken the lead in handing in the papers?

Source: Laika Think Tank, Author: Xin, Editor: G3007

Image source: Generated by Unbounded AI tool

From the perspective of commercialization, due to different training data, different algorithms, different legal regulations, etc., it is more difficult to make products like ChatGPT in China, but in terms of vertical industry large models, there are More opportunities for divergence. Previously, the industry had predicted that in the future, there would only be 2-3 general-purpose large-scale models, but vertical large-scale models will flourish, and rich industry scenarios are the biggest attraction of ChatGPT. At present, many education companies have released the latest educational large-scale model products, and education has become one of the fastest-growing scenarios for AI. We believe that the industry model of education has three major points of interest:

**1. "Education is the easiest field for ChatGPT to implement" is a prophecy. **

**2. While taking the lead in the large-scale model business in the education industry, we also discovered a new problem: that is, when enterprises get involved in the large-scale model business, should they use self-research to take the first step, or open up the interface for external procurement after maturity? Judging from the first wave of the "100-model war" in China, major manufacturers have basically chosen the path of self-development. **

**3. Under the trend of ChatGPT, there are infinite possibilities in educational technology. **

Today, the 2023 National College Entrance Examination officially kicks off. However, the "preparation for the college entrance examination" of domestic large-scale models has gradually entered the "vertical competition", and the education industry has already taken the lead.

In the past month, Netease Youdao, Xueersi, and HKUST Xunfei have successively released the latest educational large-scale model products. Among them, Youdao is an AI spoken language teacher developed based on the "Ziyue" large model, Xueersi official publicity mathematics large model MathGPT is developed, and HKUST Xunfei is an AI learning machine based on the Xinghuo cognitive large model. The latest news shows that Yuanfudao has set foot in large-scale models by launching intelligent hardware — Xiaoyuanxue training machine, and Zuoyebang is also internally testing large-scale education models based on the Chinese market, involving tools App, intelligent hardware, books, etc. .

**As OpenAI co-founder Sam Altman once said, with the support of ChatGPT, there will be excellent copywriting business, education service or other companies. After the college entrance examination proposition composition was released today, many people tried to use AI to write the college entrance examination composition. **

So, in the "sprint for the college entrance examination" of this hundred-model competition, what are the highlights behind the fact that education companies are the first to submit the papers?

Aspect 1: "Education is the easiest field for ChatGPT to land" is a prophecy

Since the release of ChatGPT, the field of AI large-scale models has continued to be popular. In the past year, major domestic Internet companies, artificial intelligence companies and even academia have successively announced their entry into the field of large-scale models. At present, Baidu, Alibaba, SenseTime and other companies have successively released their own large-scale models and application scenarios, but whether it is Baidu's "Wenxinyiyan", Ali's "Tongyi Qianwen", SenseTime's "Consultation ", or the large-scale model moss developed by Fudan University, these are all conversational general-purpose large-scale model products, and they have not been opened for use so far.

**Compared to general-purpose large-scale models, some vertical large-scale models choose to take small steps and speed up their implementation. For example, "Education is the easiest field for ChatGPT to implement" has become a hot topic since the beginning of the year. **In February of this year, Netease Youdao, HKUST Xunfei, Xueersi, and Zuoyebang and other education companies all disclosed the products and technology layout related to ChatGPT. Recently, these companies also announced the latest research and development progress and results.

According to media reports, on May 5, Netease Youdao showed a video of an AI oral teacher developed based on the "Ziyue" model. Users engage in multiple rounds of dialogue. Then, on May 6, HKUST Xunfei released a large model of Xunfei Spark, which will be used in products such as educational intelligent hardware AI learning machines. In addition, Xueersi also announced that the self-developed mathematical model is called MathGPT, and Zuoyebang reported that it is internally testing a large educational model based on the Chinese market, including multidisciplinary problem solving, Chinese and English composition correction, multilingual dialogue, etc. Educational application scenarios.

At the same time that the "Hundred Models War" is in full swing and a new round of start-up financing boom is being set off, why is the education industry uniformly taking the lead this time? Laika Think Tank believes that there are three main points:

**First, from the perspective of the purpose of education, education itself is an important way for human beings to acquire and learn knowledge, and technological progress can greatly improve the efficiency of this learning. **The education field requires a large amount of knowledge storage and transmission, and ChatGPT, as an artificial intelligence model, can quickly process and integrate this information, and at the same time, through training and learning of a large amount of data, accumulate a large amount of knowledge and language expression ability, and finally achieve external Output.

This is like an upgraded iteration of the way we acquire knowledge. In the early 1.0 stage, we mainly read and lectured by teachers, and the 2.0 Internet stage brought us into the era of information explosion. Now 3.0 represented by ChatGPT has emerged, making information and knowledge acquisition more efficient. From passive instillation to active acquisition, to precise interaction, ChatGPT has more practical needs in the field of education.

**Secondly, from the perspective of the characteristics of education, education itself is highly interactive, and it is necessary to complete preaching, teaching and dispelling doubts in this process. **ChatGPT can provide a more natural communication experience, such as voice interaction, image recognition, etc., which can help students better understand and memorize knowledge. For example, for AI speaking teachers, ChatGPT can use its natural language processing capabilities and deep learning algorithms to analyze students' problems in pronunciation, intonation, speech speed, and voice fluency, and provide targeted advice and training to help students improve their oral English. In Chinese composition correction, ChatGPT can also analyze students' composition content, language expression, grammatical structure and other problems, and provide targeted suggestions and guidance.

Based on such industry attributes, ChatGPT has a stronger match with education. For example, the "AI Speaking Teacher" released by Youdao, from the actual pronunciation experience, is no longer the monotonous mechanical pronunciation of AI in the past, but more like a real "teacher" talking to students, and the dialogue experience is real and interesting. Of course, the accuracy and reliability of ChatGPT in the application of AI oral teachers and Chinese composition correction still need to be continuously improved and optimized.

**Third, from the perspective of the development stage of education, as the domestic education industry gradually enters the era of intelligent learning and education digitization, education companies also need to find new singularities, and ChatGPT is the breakthrough. **Especially driven by next-generation information technologies such as artificial intelligence, big data, and 5G, the continuous integration of key technologies and educational scenarios and the deepening of intelligent education will be promoted, and online education, AI education, and personalized education will be further popularized. This is also an important reason for major education companies to increase investment in the layout.

It is worth noting that in foreign countries, ChatGPT has begun to reshape the education industry. According to the survey of Study.com abroad, 21% of foreign teachers have started to use ChatGPT to assist their teaching work. Recently, Chegg, a listed US education and training company, also stated that after launching the auxiliary learning platform CheggMateAI in April this year, it has already affected the user growth of its original business, because students’ interest in ChatGPT has increased greatly, and they no longer need to obtain relevant learning materials through Chegg.

Aspect 2: Self-research first or external acquisition? A New Proposition for Large Models

At the same time as the education industry took the lead in large-scale models, we also discovered a new problem: that is, when enterprises get involved in the large-scale model business, should they use self-research to take the first step, or open up the interface for external procurement after maturity?

Previously, many netizens questioned that some companies launched their own large-scale models first in response to hot spots, but they were actually launched on the basis of GPT released by OpenAI, and their self-developed degree has a lot of moisture. Some Zhihu netizens bluntly said: The service launched by a domestic team seems to directly use a large amount of ChatGPT question and answer data to train their own network, and the input questions are highly similar to the answers obtained through ChatGPT.

In terms of external procurement mode, ChatGPT abroad has clearly pointed to three methods: API, subscription system and strategic cooperation (embedding Microsoft Bing, Office and other software). For example, in the field of education, Duolingo, Quizlet, Khan Academy and other products mainly cooperate with OpenAI to make fine-tuning and interface calls on the GPT large model to enhance the original product experience.

** Judging from the first wave of the "Hundred Models War" in China, the major manufacturers have basically chosen the path of self-development. **

Regarding the problem that Chinese companies choose self-developed ChatGPT-like products and technologies, Zhang Jiong, founder of Xingtubit, once analyzed that there are three reasons: the first is the market side. Currently, ChatGPT is not open to domestic users. It is impossible to use it at will like overseas users, and from the market level, there will inevitably be a Chinese ChatGPT; the second is the technical side. The large models currently available on the market, such as the basic model of ChatGPT GPT-3.5, are basically English-speaking For training, the natural language processing effect of the Chinese benchmark is still slightly inferior to that of the English benchmark, so the appearance of a Chinese benchmark model will further improve the user effect; the third is data security, AI content generation itself is based on a large amount of data training As a result, OpenAI has gradually changed from a public welfare project to a market-oriented project. In the future, there may be uncertainties due to factors such as interests. In addition, mainland China requires all data to be stored locally. Without a team, localized storage and maintenance of data is difficult to meet regulatory requirements.

Self-study is good, but it is not easy to do. Generally speaking, the research and development of general artificial intelligence requires continuous investment of funds and top talents, and is limited by the requirements of computing power, data specifications, historical investment, top talents, and financial strength. According to Business Insider's estimate, the current monthly maintenance cost of ChatGPT is 3 million US dollars. Considering the labor cost, the cost is estimated to double. OpenAI's products use Microsoft's Azure enterprise cloud server, and some investment bank analysts estimate that the annual cost may range from 25 million US dollars to 1 billion US dollars. At the same time, artificial intelligence also faces the problems of long investment cycle and slow output time. For example, it took OpenAI 7 years from the establishment of the company to the launch of ChatGPT.

** Compared with general artificial intelligence, vertical artificial intelligence may be able to run faster based on its own accumulation in professional fields. This time, HKUST Xunfei and Youdao took the lead in landing on the AIGC technology self-research path in the education industry, and also confirmed the proverb "If you don't have diamonds, don't do porcelain work". **

For example, Liu Cong, dean of the Xunfei Research Institute of the University of Science and Technology of China, said that based on years of deep accumulation in deep learning algorithms, large-scale model technology, industry big data, knowledge graphs, multi-modal perception, system engineering and other technologies, Xunfei At the same time, relying on the three national-level platforms of the National New Generation Artificial Intelligence Open Innovation Platform, the National Key Laboratory of Cognitive Intelligence, and the National Engineering Research Center for Language and Speech Information Processing, it already has the confidence for comprehensive research and development of cognitive intelligence large models and long-term continuous breakthroughs and ability.

The same is true for Youdao. On the one hand, Youdao has covered lifelong learning content from children to teenagers to adult education, and has enough learning materials and user data; on the other hand, Youdao itself has accumulated in AI technology and applications, such as having a neural network The four core technologies of translation engine, computer vision, intelligent voice AI technology and high-performance computing cover translation, education, voice assistant, writing and other fields, and have been installed in star products such as Youdao dictionary pen and learning machine superior.

There is also Xueersi, which started as a mathematics extracurricular tutoring. It has accumulated 20 years of mathematics teaching experience and accumulated a huge amount of mathematics-related data. These data are necessary materials for MathGPT training. There is also Yuanfufudao. As early as 2014, it established an AI research institute covering five laboratories including vision laboratory, natural language understanding laboratory, speech laboratory, audio and video laboratory and basic support laboratory, which also provided sufficient support for it. technical support.

** On the whole, both self-research and external mining have their own advantages and disadvantages. For enterprises that choose the self-research route, its biggest advantage lies in "tailor-made customization", which can meet and match the needs of its own business to the greatest extent, but it also requires enterprises to have sufficient comprehensive strength and accumulation, and at the same time have to bear certain burdens. The cost of trial and error is not something ordinary small businesses can afford. Enterprises that choose outsourcing have the advantages of low cost and many choices, but the disadvantage is that the special needs are not flexible enough and not customized enough. Maybe when they use it themselves, their peers are already ahead by a long way. **

It can be said that under the wave of continuous evolution of large language models, the choice of different technical routes and different application methods is still to be discussed and verified, and the answer needs to be found in practice.

Aspect 3: Under the ChatGPT trend, the infinite possibilities of educational technology

At present, the application of ChatGPT and other related technologies in the field of education mainly focuses on Chinese and English oral and composition guidance. This is closely related to the large AI model generated based on massive data behind ChatGPT. It uses available human books, academic papers, news, and high-quality information as learning content, and strengthens learning based on human feedback. The breakthrough of this technology also makes ChatGPT's dialogue closer to human beings, and the logical correlation between sentences is significantly improved.

Moreover, ChatGPT at this stage is still limited in its ability to update data in real time. Obviously, many areas and scales that can answer questions have been sealed. With the iterative update of technology, there are more possibilities in the future, which also gives domestic ChatGPT Provided new market space.

** In the field of education, in addition to specific subject exercises and knowledge quizzes, ChatGPT can also help students and teachers achieve a more personalized, efficient and intelligent learning and teaching mode, including intelligent tutoring, intelligent answering system, language learning, chat robots , smart tutors and other application scenarios. **

For example, in the teaching process, teachers are often faced with interdisciplinary teaching, and some students' questions require teachers to think for a long time before they can answer them. However, if ChatGPT is used, it can provide students with quick and easy answers to questions in the first place. Interactive "testing" of learning results, online questions, or quizzes in class, to evaluate learning effects through user feedback, and further formulate learning plans; at the same time, for some repetitive questions, ChatGPT can directly Replacing the teacher's answer saves the teacher a lot of time and allows the teacher to exert greater value in thinking and creativity.

In addition, the application of ChatGPT can also promote educational innovation. In the traditional education mode, teachers often impart knowledge one-way, while ChatGPT can interact with students, stimulate students' interest in learning, and cultivate students' independent learning ability. In addition, ChatGPT can also use data analysis and artificial intelligence technology to provide teachers with more comprehensive and accurate student evaluation services. In this way, education can be carried out more scientifically, and the learning effect of students can be improved more obviously.

It is worth noting that through this continuous model training, ChatGPT can also be integrated with other education and training. Whether it is preschool education, vocational education, certification or training, this model can help them make training better. Well, make it more efficient.

In short, the application of educational technology represented by ChatGPT is of great significance to personalized education, educational fairness and educational innovation. In the future, we can look forward to the wide application of ChatGPT in the field of education, providing students with more intelligent, convenient, and personalized education services, and at the same time bringing more fair and convenient learning opportunities to students in rural areas, and gradually eliminating this phenomenon. The "digital divide" in education.