Can iFLYTEK take off with a large model?

Original source: Heike Finance

Author: Xu Junhao

Image source: Generated by Unbounded AI‌

Since the sudden emergence of ChatGPT in December 2022, the large model has quickly become a popular project in China. The track is surging and the variables are increasing day by day.

Wang Huiwen, the co-founder of Meituan, who had to retire due to personal health reasons, has not only resigned from the position of non-executive director of Meituan, but also light years away from the company he founded and ignited a big fire in the direction of domestic large-scale entrepreneurship. After much attention, it has also found a solution that looks pretty good-Meituan bought 100% of the shares of Light Years Away at a total price of about 2.065 billion yuan, and the good brother Wang Xing replaced Wang Huiwen as the A trader from light years away.

Looking back over the past six months, after Baidu followed in the footsteps of ChatGPT and took the lead in launching the large-scale model Wenxin Yiyan in China, many leading technology companies have made moves, such as Ali Tongyi Qianwen, Tencent Hunyuan, Huawei Pangu, etc. are listed here. According to the "China Artificial Intelligence Large-scale Model Map Research Report" issued by the New Generation Artificial Intelligence Development Research Center of the Ministry of Science and Technology, as of May 28, 2023, 79 domestic large-scale models with parameters above 1 billion have been released.

Among the endless large models, the entry story of the veteran AI company iFLYTEK is also worth mentioning.

HKUST Xunfei released the Xinghuo Cognition Large Model on May 6, 2023, and released a new v1.5 version one month later. Liu Qingfeng, founder and chairman of HKUST Xunfei, said that by the time of the HKUST Xunfei Global Developers Conference on October 24, the Xinghuo model will surpass ChatGPT in Chinese and reach the level of ChatGPT in English.

These actions and expressions have strongly promoted the rise of HKUST Xunfei’s stock price. In June 2023, the stock price of HKUST Xunfei reached its highest point since its listing at 82 yuan per share, corresponding to a total market value of more than 150 billion yuan. You know, in January 2023, the market value of HKUST Xunfei was less than 80 billion yuan.

However, after the release of the large model of Spark, there were many controversies.

One of the controversies came from a domestic agency called SuperCLUE. In the comprehensive evaluation benchmark list of Chinese general-purpose large-scale models released by the agency on May 9, the Xunhuo large-scale model of HKUST ranked fourth after humans, ChatGPT 4, and ChatGPT 3.5. Netizens found that SuperCLUE was modeled on the international NLP data sets GLUE and SuperGLUE, which are authoritative natural language processing datasets, but it is actually a self-assessment by domestic civil organizations. GLUE and CLUE, like Li Kui and Li Gui, can't be distinguished unless you look carefully. According to public information, the SuperGLUE list generally has more than 20,000 questions, but there are only 100 domestic SuperCLUE questions.

On the day when the list was released, the SuperCLUE official website showed that Cui Yiming, a senior researcher at the Xunfei Joint Laboratory of Harbin Institute of Technology, ranked first among the member consultants of the Chinese benchmark assessment. The next day, May 10, the website deleted the relevant information. In the Chinese large-scale model ranking list in June 2023, the organization ranked humans at 0, the first and second places were still ChatGPT, and the third and fourth places became 360 Zhinao (4.0) and Wenxinyiyan (v2.0.4), Xunfei Xinghuo (v1.5) ranked sixth.

The topic didn't end there.

On May 24, HKUST Xunfei’s stock price plummeted, falling more than 9% at one point during the session. The official response of HKUST Xunfei stated that the stock price drop was caused by a false essay written by a generative AI product, and HKUST Xunfei mentioned in the essay was exposed to be suspected of collecting a large amount of user privacy data and using it for artificial intelligence research as false information , the company's legal department will take legal measures if it is found that there are malicious rumors.

Anyone with a discerning eye can tell at a glance that this is pointing the finger at Baidu Wenxin's words.

Wenxin Yiyan market director Zhang Quanwen said in a circle of friends on the same day that friends and merchants attributed the stock price drop to Wenxin Yiyan's proposition. To come and not to be indecent."

Regardless of whether it is based on the facts or the marketing of Pengci, the domestic large-scale model has just hit the road, and it is hard to say which one is stronger and weaker. However, the distance between the Spark large-scale model and the ChatGPT claimed by iFLYTEK is obvious to all. In the context of the annual revenue growth rate falling from 40.61% in 2021 to 2.77% in 2022, it is still unknown what the big model can bring to the business direction of iFLYTEK.

From Speech to AI

HKUST Xunfei once had a first-mover advantage in the field of domestic intelligent voice.

The company, which has been established for more than 23 years, had sales of over 100 million yuan in 2004, and its cooperation with Lenovo, Huawei and other companies has made it a giant in the field of voice interaction.

In May 2008, HKUST Xunfei was listed on the Shenzhen Stock Exchange. The company's annual revenue was 257 million yuan, a year-on-year increase of 25.14%. The 257 million yuan is composed of three main businesses, which are voice support software with a revenue of 54.58 million yuan, voice industry applications/systems with a revenue of 76.31 million yuan, and information engineering and operation and maintenance services with a revenue of 126 million yuan. It is particularly worth noting that the information engineering and operation and maintenance service business, which accounts for 49.17% of the total revenue, is mainly supported by Anhui, where iFLYTEK is located, which contributes 99.49% of iFLYTEK’s revenue from this business.

Around 2010, when smartphones were on the rise, iFlytek quickly seized the opportunity to launch the iFLYTEK Voice Cloud. In 2011, the first entry-level speech synthesis chip was launched, which also made the iFLYTEK input method enter the field of vision of mobile Internet users. Now Xunfei input method supports speech recognition of 23 dialects including Cantonese, Hakka, Hokkien and 5 minority languages including Tibetan.

In 2014, HKUST Xunfei launched the Xunfei Super Brain Project, which shifted from NLP to AI. For the whole year of 2015, HKUST Xunfei’s revenue was 2.5 billion yuan, of which the highest proportion was still information engineering and operation and maintenance services, with a revenue of 720 million yuan (28.83%); followed by educational application products, with a revenue of 660 million yuan (26.4%); Telecom voice value-added products, revenue 313 million yuan (12.52%); voice support software, revenue 303 million yuan (12.12%); industry application product IFLYTEK-C3, revenue 247 million yuan (9.91%); the rest accounted for Smaller ones include audio and video surveillance, communication data analysis applications, college education and teaching, and other business income. In other words, software and information technology services mainly support iFLYTEK’s revenue.

With the rapid development of the Internet, many major manufacturers have also entered the field of voice interaction, which makes the position of HKUST Xunfei gradually less stable. According to data from the China Speech Industry Alliance, iFlytek accounted for 54.3% of China's smart voice market in 2012, and dropped to 44.2% in 2015. Baidu ranked second with 27.8%, and Apple accounted for 6.9%. company. Baidu, Netease and other big companies have made achievements, which will inevitably have an impact on HKUST Xunfei.

The outside world also has doubts about iFLYTEK's AI products, such as simultaneous interpretation products. In 2018, someone published an article on the Internet, arguing that the translation of iFLYTEK products actually included human translation elements. HKUST Xunfei responded that the product uses the "man-machine coupling translation assistant" mode, that is, the machine provides the results of speech recognition and machine translation to the simultaneous interpreter to assist the simultaneous interpretation work. This is also tantamount to an objective admission that "Xunfei Hearing" is far from reaching the level of simultaneous interpretation, and cannot replace simultaneous interpreters with AI.

Despite the shortcomings, iFLYTEK’s business is still performing well, with an average revenue growth rate of more than 25% for 10 consecutive years. But the more emerging technologies, the faster the internal iterations in the field, including voice interaction and AI. According to the "2023-2028 China Intelligent Voice Industry Market Development Status and Investment Direction Research Report" issued by Huajing Industrial Research Institute, the scale of China's intelligent voice market has increased from about 10 billion yuan in 2017 to 34.1 billion US dollars in 2022. In 2023, it can reach US$38.2 billion; and in 2022, the domestic market share will be higher than that of HKUST Xunfei (44.2%), Baidu (37.8%), Apple (15.4%), Nuance (6.9%) and Xiaoi Robot (3 %).

The voice field is under pressure, and generalized AI is even more difficult.

Research firm iResearch's "China Artificial Intelligence Industry Research Report" shows that the market size of China's AI industry has reached 154.6 billion yuan in 2020, 195.8 billion yuan in 2022, and is expected to reach 612.2 billion yuan in 2027. CAGR is the compound annual growth rate of 25.6%.

The market is positive and the competition is fierce. In the entire AI industry map, including computer vision, big data intelligence, conversational AI, content review, intelligent driving and other fields, iFLYTEK’s technology layer and specific application layer need to face domestic Baidu, Ali, Tencent and other companies. Overseas Google, Microsoft and other big companies are attacking the city.

The financial report shows that HKUST Xunfei’s revenue in 2022 will be 18.82 billion yuan, an increase of only 2.77%; the revenue in the first quarter of 2023 will be 2.887 billion yuan, a year-on-year decrease of 17.64%. HKUST Xunfei officially attributed the performance problems in the first quarter to the special environment in December 2022 and January 2022 and the subsequent Spring Festival holiday. It believes that this is a short-term impact and will not affect long-term business fundamentals. The official also stated that after being included in the US entity list in 2019, it will be under pressure again in October 2022; the fourth quarter of 2022 and the first quarter of 2023 are under adjustment, and the launch of large models is expected to start from the second quarter , revenue and gross profit will grow positively.

**Model color geometry? **

According to the financial report, for the whole year of 2022, the main business of HKUST Xunfei is still software and information technology services, accounting for 98.41% of total revenue; the largest proportion is education products and services, accounting for 32.74% of total revenue; Next is open platform and consumer business (24.66%), followed by smart city (23.63%) and operator-related business (11.14%); the combined proportion of smart cars, smart medical care and smart finance is 6.2%.

The rapid development of AI business requires iFLYTEK to keep up with the technological trend, and the importance of large models is self-evident.

The investment in research and development shows that iFLYTEK attaches great importance to large models. According to the financial report, HKUST Xunfei will invest 2.936 billion yuan in research and development in 2021, which will increase to 3.355 billion yuan in 2022, a year-on-year increase of 14.28%; the number of R&D personnel has increased from 8,367 in 2021 to 9,281 in 2022, a year-on-year increase of 10.92%. The proportion of the total employees has increased to 61.68%.

But in fact, apart from the algorithm, the difficulty and focus of large model training lies in the need for huge data and huge computing power, including hardware devices such as CPU (central processing unit), GPU (graphics processing unit), ASIC (application-specific integrated chip) and support for large A software platform and framework for large-scale parallel computing, in addition to involving a large amount of data storage, transmission, and related human maintenance and upgrades. According to NVIDIA research data, the largest model of ChatGPT 3 needs to use 175 Billions of parameters, and needs 512 V100 graphics cards to train for 7 months or 1024 A100 chips for up to one month. The monthly cost is in Millions of dollars or more.

This means that large factories with strong financial strength have more confidence to burn money, and it is difficult for ordinary enterprises to participate in the competition. For the whole year of 2022, Ali and Tencent’s R&D investment will exceed 50 billion yuan, while Baidu’s R&D investment is 21.4 billion yuan, and its Wenxin Yiyan still has a large gap with ChatGPT, not to mention that the R&D investment in the same period is only about 3 billion yuan HKUST Xunfei.

Even so, many people still have high hopes for iFLYTEK. This is not only because it has a certain first-mover advantage in voice and AI, but also because this company has some national color. The latest financial report, the first quarter of 2023, shows that the state-owned legal person China Mobile Communications Co., Ltd. and the University of Science and Technology of China Asset Management Co., Ltd. rank first and fourth among the top ten shareholders of iFLYTEK, with a shareholding ratio of 10.66% respectively. % and 3.59%, as the founder and chairman Liu Qingfeng ranked second with a 7.24% shareholding ratio.

This is considered to be easier for iFLYTEK to create a closed business loop from To G to To B. This is also mentioned in the financial report. For example, the financial report shows that as of the end of 2022, iFLYTEK’s intelligent marking technology has been officially delivered and applied in the college entrance examinations of 14 provinces and cities; the judicial business field has been connected to 466 sub-platforms, involving 26 provinces and municipalities, with a total of Covering more than 1,800 courts.

However, it is difficult to directly link the development of G-end and B-end with the development and application of large models, and it is impossible to completely dispel market doubts. Shortly after the release of the large model of Spark in May, some netizens found that the latter's answer would include "I was developed by OpenAI" and other content, thus suspecting that the large model of Spark was developed by OpenAI. HKUST Xunfei said that this is due to the high popularity of ChatGPT and the presence of more words such as OpenAI and ChatGPT in the training data, which caused answer errors.

Common inspection criteria for large models include multi-style, multi-task long text generation, multi-level cross-language comprehension, pan-domain open knowledge question and answer, situational thinking logical reasoning, etc. According to the observation of Haike Finance, a considerable number of netizens have tested the large model of Xinghuo. In the face of some logical reasoning questions that are not too complicated from a human perspective and some Chinese and mathematics questions in the college entrance examination, the Spark Big Model failed to give correct answers.

Because the development cost of large models is high, it will take time for application and commercialization to cover the cost. OpenAI announced in March 2023 the open API (Application Programming Interface) at a price of $0.002 per 1000 tokens. It also said that it will open the API of Whisper, an artificial intelligence-driven speech-to-text model, at a price of $0.006 per minute for developers. According to foreign media reports, OpenAI will lose more than 540 million US dollars in 2022, all thanks to Microsoft's blood transfusion behind it.

It is undoubtedly more difficult for the large-scale model of Spark to benchmark ChatGPT to directly bring profits to HKUST Xunfei in the short term, and HKUST Xunfei still needs to partially rely on government subsidies. The financial report shows that from 2018 to 2022, HKUST Xunfei received government subsidies of 276 million yuan, 412 million yuan, 425 million yuan, 438 million yuan, and 473 million yuan respectively.

How to Build Barriers

The benefits of the development of large models to iFLYTEK are directly reflected in the products.

HKUST Xunfei’s revenue growth in the education field has exceeded 30% in the past five years. The large model can promote the development and sales of products in downstream fields such as education, office, and medical care on the basis of synergy at the data and application levels, and expand the influence of smart products. and market share. A research report of Minsheng Securities in May 2023 concluded that with the full support of policies, HKUST Xunfei uses GBC to build a positive closed loop of education, and fully accelerates its penetration into the national market.

HKUST Xunfei has first implemented the large model on the existing AI learning products, such as Xunfei AI learning machine T20, T20Pro and so on. HKUST Xunfei released its 618 battle report on June 19, claiming that sales increased by 125% year-on-year. AI learning machine series, translators, smart recording pens, and smart office books won the sales champions of JD.com and Tmall dual platforms in their respective categories. But the official did not announce the specific sales figures.

As far as educational tablet is concerned, there are many players in this field, which is already a red sea. Among them, there are not only old-fashioned learning machine manufacturers represented by BBK and Shushulang, but also Internet forces such as Baidu and Netease Youdao. According to IDC data, from the second half of 2021 to the first half of 2022, Baidu's small learning machine market share will rank first in the learning machine tablet market, followed by BBK and HKUST Xunfei.

If we compare the market situation of entertainment products, the sales volume of Microsoft Xbox is in the shadow of Sony PS4, and it is far from dominant; however, Microsoft retains core technologies such as human-computer interaction interface and the matching software and hardware ecology such as Microsoft Cloud and Microsoft operating system, so that Xbox have a certain irreplaceability.

In contrast, at present, iFLYTEK’s AI and large-scale model level is difficult to build technical barriers, and it is difficult to stabilize its position in the same low-ranking situation. Even if the market still has demand potential, it is not easy to achieve explosive growth in it. .

HKUST Xunfei also has advantages in some AI subdivisions, but these are far from the main revenue, such as AI medical products, intelligent medical assistants. As mentioned in the financial report, Smart Medical Assistant is the only intelligent auxiliary diagnosis system in the industry that has passed the national licensed physician qualification test; as of the end of 2022, the system has covered 30 provinces and 380 districts and counties across the country and has been applied on a regular basis. Provided more than 530 million AI-assisted diagnosis and treatment suggestions. The annual revenue of this business in 2022 is 467 million yuan, accounting for only 2.48% of iFLYTEK’s total revenue in the same period.

More importantly, in 2018, HKUST Xunfei’s sales expenses were 1.726 billion yuan, exceeding the R&D investment in the same period, and the sales expenses in 2022 have increased to 3.164 billion yuan.

At the same time as the official announcement of AI, iFlytek also carried out diversified exploration. For example, in February 2019, Xunfei Yunchuang, a subsidiary of HKUST Xunfei, invested in South China brain-controlled intelligent technology to get involved in the research and development of brain-computer interfaces; in September 2022, HKUST Xunfei acquired Beijing Global Insurance Brokers from Ping An Group. Co., Ltd., which was renamed Beijing Kexun Insurance Brokers Co., Ltd.; in October 2022, HKUST Xunfei registered the relevant trademarks of Xunfei Selection.

It is not known how much pressure these layouts involving different fields and different levels of investment can disperse for iFLYTEK’s main business and how much new ideas it can provide. What is known is the self-reduction of iFLYTEK’s revenue goals.

At the annual planning conference in February 2021, HKUST Xunfei once proposed a revenue target of 100 billion yuan in the next five years, which represents a compound annual growth rate of 50%. There is obviously still a gap between 40.61% in 2021, and 2.77% in 2022 is getting farther and farther away. In January 2023, the revenue of 100 billion yuan was updated to achieve a goal of more than 50 billion yuan in the next five years. Comparing the two, it has been cut in half.

The cost is as deep as the sea, and the players are as numerous as the cloud, and there may be great opportunities for large models. As of the publication of this article by Heike Finance, there are only 113 days left before iFlytek claims to catch up with ChatGPT. Of course, looking at the changes in revenue targets, the final result of catching up with ChatGPT is not so important. The important thing is whether HKUST Xunfei can install a new engine for strong growth. The market is watching and voting is about to begin.