laitimes

The big model should be a "strong pill" in the education and training industry

author:Everybody is a product manager
In 2023, large-scale models have sparked a cult in the tech world, and even the education and training fields, which are still slowly recovering, are caught up in this trend. The author of this article analyzes the impact of large models on the education and training industry, let's take a look together.
The big model should be a "strong pill" in the education and training industry

In recent days, the latest news of Baidu mobile phone has come out, and the Xiaodu public account released an article "Xiaodu Qinghe, First Meeting" has further confirmed the previous rumors that Baidu will enter the mobile phone industry. However, unlike the usual mobile phones, Xiaodu Qinghe focuses on learning mobile phones and emphasizes AI tutoring.

In other words, this is a smart education hardware, and it connects to Baidu's hottest "Wen Xin Yiyan".

Big model + education, this seems to be the trend driven by ChatGPT from the beginning, at that time, the explosion of overseas ChatGPT can not be separated from the pursuit of students, in the OpenAI official website announced 6 cases of using GPT-4, two are from the education industry.

In China, well-known manufacturers such as NetEase Youdao, Xueersi, and iFLYTEK have released the application of their own large model results in the field of education. Among them, iFLYTEK released the world's first AI learning machine T20 equipped with a large model, providing advanced functions including composition correction, simulated speaking teacher real-life dialogue, etc., and NetEase Youdao developed an AI speaking teacher based on the "Ziyue" large model.

In 2023, the science and technology circle will collectively rejoice for the big model, and in addition to the wind, it will even pull up the education and training industry that is still slowly recovering.

First, AI makes money from education and training?

Artificial intelligence has no prospects, no "money scene" has always been a mustard in the industry can not be underestimated, with the serious imbalance between the payment of capital and recovery, gradually, vertical scenes are more attractive than the whole field. For example, Huawei's Pangu model emphasizes the application in mining and power scenarios, and Tencent's CV model is implemented in industrial quality inspection and medical imaging.

Automotive, industrial (manufacturing), aerospace satellites, energy, biopharmaceuticals... These are the main areas in which the future big model will go deeper, but no one expected that compared with the above representative industries full of science and technology, it was education and training that first saved the big model from drought and flood.

Taking iFLYTEK as an example, since FY2016, the contribution of smart education business to iFLYTEK's revenue has been directly reflected in the financial report, from 2016 to 2021, the proportion of education products and services revenue to the total revenue was 27.44%, 25.54%, 25.45%, 23.36%, 30.86%, and 32.82%, respectively, with a compound growth rate of more than 44%.

In 2022, artificial intelligence companies on the market basically maintain the "face project", Yuncong Technology's revenue in the third quarter of last year was 65 million yuan, down 80.07% year-on-year, and in the first three quarters, Cloudwalk Technology accumulated a loss of 589 million yuan, which has approached the loss of 632 million yuan in 2021.

SenseTime is no exception, according to incomplete statistics, from 2018 to 2022, the total loss exceeded 40 billion yuan. In contrast, iFLYTEK's rare revenue is precious, and the financial report shows that iFLYTEK's AI learning machine revenue in 2022 will increase by 53%, and the revenue growth of offline stores will reach 106%, and stores and agents will continue to be profitable.

It is reported that in 2015, iFLYTEK began to cooperate with the education department to operate the Mandarin computer test, Chinese and English composition grading, bet on intelligent voice, and open up the information teaching and test market.

Today, although the scenes that artificial intelligence can reach are all in unison, for iFLYTEK, the teaching and examination system still accounts for a large revenue contribution, and even new educational hardware such as learning machines has been added. The data shows that in iFLYTEK's financial report for the first half of 2022, education products and services accounted for 27.49% of revenue, teaching business accounted for 1.54%, and the education field accounted for 29.03%.

Nearly 30% of the proportion, further proves that large models in the education and training market has great potential, especially after the "double reduction", the offline attributes in this field gradually weakened to disappear, hardware facilities to make up for the trend, and headwinds rising, coupled with AI's increasingly powerful interactive functions, making the binding between artificial intelligence and educational hardware tighter.

I don't know if AI has finally found a trace of monetization redemption in educational hardware, or the scale of the high-growth education hardware market year after year has produced some visible application qualitative changes due to the blessing of large models.

iFLYTEK's publicity in the learning machine is inseparable from the AI system, such as AI test preparation, AI leak detection, AI tutoring, AI English, AI stylus... are common words. It is reported that from January to June last year, iFLYTEK AI learning machine in JD.com and Tmall GMV increased by more than 170% year-on-year, sales increased by 100% year-on-year, and became the sales champion during 618.

Perhaps, this is a win-win, although in the eyes of many AI companies, educational hardware is far less promising than industry and automobiles.

Second, the education and training world "sits up in a dying dream"

The advent of ChatGPT is like throwing a stone in the world's science and technology, in just a few months, the domestic business order alone has been severely impacted by the big model, Baidu Wenxin words, Alibaba Tongyi Qianwen, SenseTime Ririxin, SenseNova system, Huawei Cloud Pangu, Zhihu "Zhihaitu AI", 360GPT, Kunlun Wanwei "Tiangong" large model, Jingdong Yanxi... Endless.

What is ridiculous is that it is not only the technology giants who suddenly strive for strength on the big model, but also the education and training companies that once died up because of the "double reduction" have also started the road of research and development. In May, Xueersi plans to develop a large mathematical model, mainly focusing on problem-solving and lecture algorithms in the field of mathematics, for mathematics enthusiasts around the world.

Coincidentally, NetEase Youdao launched "Ziyue", in the field of education, even if there is no self-developed model, it will choose to cooperate with related technologies, such as Duolingo, Quizlet, Khan Academy and other platforms, choose OpenAI. Overseas large models have long been associated with disciplines, and Google has acquired many products focusing on mathematical calculations.

The same is the research and development of large models, the education and training industry wants to rely on the east wind of new technology to come back to life again, I am afraid it is not so simple, technology, funds, follow-up loading and application... Everywhere are the details of survival worth pondering in the education and training industry.

Let's start with the technical issues. It is undeniable that compared with ordinary language models, professional teaching models have higher requirements for language organization ability and logical reasoning ability, and have low error tolerance. MathGPT has put forward three requirements to make up for the shortcomings of the existing general-purpose language model:

First, the topic should be solved correctly; Second, the problem-solving steps should be stable and clear; Third, the problem solving should be interesting and personalized. But at present, these requirements of the education and training industry can not even meet the hottest ChatGPT, Google acquired several major mathematical products, such as Photomath, Microsoft Mathematics, Mathway is still using non-large language model traditional AI technology, plus the database way to solve mathematical problems, education and training enterprises cross-border self-research success probability can be imagined.

The second is the issue of funding, and AI startup projects never end to burning money. The success of ChatGPT is due to Microsoft's generous funding and computing power support, the US market research institute TrendForce estimates that 20,000 GPU chips are required to process ChatGPT's training data, and as OpenAI further expands the commercial application of ChatGPT and other GPT models, its GPU demand will exceed 30,000.

Operating costs are even higher, semiconductor research company SemiAnalysis said that according to the GPT-3 model, ChatGPT's daily operating costs are as high as $700,000, and at $700,000 per day, ChatGPT's annual operating costs are as high as $255.5 million, and many large-model startups, just renting 1,000 GPU cards a year will cost tens of millions to 100 million yuan.

How much money can an education and training enterprise that has not opened for more than a year have to support such a huge expense? Previously, the 2022 fiscal year performance report released by Good Future (the parent company of Xueersi) showed that as of February 28, 2022, Good Future achieved net income of $4.391 billion.

The big model should be a "strong pill" in the education and training industry

$4 billion looks like a good life, but in the face of AI entrepreneurship is just a drop in the bucket, not to mention, education and training companies are still losing money for years, in fiscal 2022, Good Future has an operating loss of $615 million, an increase of 40.2% year-on-year. The rise of large models has indeed made several education and training giants sit up in their dying dreams, but some things can only be viewed from afar, and may not be able to play profanely.

Third, the hardware "gives way", and the content should stand

In 1986, in the era when the per capita salary was not even over 100, a "Chinese learning machine" was a top luxury that countless students did not dare to think of, and later, the little bully learning machine was born, and since then it has completely opened the journey of domestic education hardware. The sudden explosion of educational hardware is precisely these two years.

At the beginning of 2021, Baidu's Xiaodu Technology took the lead in releasing the Xiaodu intelligent learning tablet; In April 2021, Tencent launched a smart educational computer based on Linux system deep customization; During 618 in 2021, Alibaba released more than 40 educational smart hardware products, and launched the first home learning smart screen product, Tmall Genie E1; In July 2021, iFLYTEK released a new high-end flagship product of the AI learning machine T10.

Offline education and training have fallen, and countless companies are waiting to fill this blank market with hardware. This begs the question, is educational hardware highly profitable? Can it lead countless large manufacturers to bend their waist? In fact, traditional educational hardware is not very profitable.

Taking Shulang as an example, this company, which has been deeply engaged in the smart education hardware industry for 23 years, has a gross profit margin of more than 22% in 2022, but its net profit margin is only 0.8%. Next, AI penetrates into hardware facilities, which should change the inherent market status quo, and educational hardware is no longer simply based on equipment to win, but instead focuses on content.

Hardware "gives way", content should stand, so that whether it is for enterprises or user groups, it is a benign change.

In terms of enterprises, one of the key reasons why large models are sought after everywhere in the education and training industry is that they are eager to use AI to form a new content ecology, and then compete in the homogeneous hardware ocean. For example, Xiaodu Qinghe learning mobile phone is equipped with Xiaodu AI learning large model, AI learning big model is based on Wen Xin Yiyan, built-in many AI teachers, and launched interactive lectures, homework assistants, speaking teachers, precision learning and other learning functions.

In addition, the large model can also fully expand learning resources, it is reported that Xiaodu Qinghe contains 1000+ textbooks, 30000+ courses of famous teachers of general subjects, and 7000+ high-frequency test center knowledge cards. He Xiaodu, who also began to use large models to play content cards, plans to launch an "AI assistant", which is not only a composition assistant, a speaking assistant, but also a reading assistant and a math assistant.

In terms of users, today's parents and students have always paid more attention to the content and interoperability of learning machines than other technology construction, and as we all know, the student group is comparable to the collection of "Dragon Ball" for the use of learning equipment. Mobile phones, dictionary pens, telephone watches, translators, listening treasures, flashcards... The whole nine yards.

From the current hardware segmentation track, it can be easily seen that users are embarrassed by complex hardware, and powerful content carrying may be able to achieve multi-purpose in one machine. Simplifying equipment is not necessarily a way to simplify learning procedures, and the essence of the teaching and training industry is efficiency in the final analysis.

Of course, there are still difficulties in activating the content of the large model, that is, the correct rate of problem solving. When ChatGPT wrote papers and homework, the physics teacher of the National People's Congress High School did a test, the content was to let ChatGPT solve the 2021 Beijing High School Examination questions, and the results showed that 5 physics questions were done correctly, 10 biology questions were done correctly 6 times, 13 history questions were done correctly 8 questions, and 9 mathematics subjective questions were done correctly 6 questions.

The human brain will make mistakes, and the intellectual brain is far from being 100% correct, which is the most contradictory crux in the landing process of big model + education and training.

Columnist

Tao always reasonable, WeChat public account: Dao always reasonable (ID: daotmt), everyone is a product manager columnist. He used to be an independent writer and an in-depth observer of the Internet and technology circles.

This article was originally published by Everyone is a product manager, and reproduction without permission is prohibited.

The title image is from Unsplash and is based on the CC0 protocol.

The views of this article only represent the author himself, everyone is a product manager, the platform only provides information storage space services.

Read on