iFlytek responds to "shelling", AI large model craze disenchants

Article|Digital Intelligence Business Show Cui Si

Editor|Zhang Hongyi

"What is the relationship between the Spark model and OpenAI?"

"I have nothing to do with OpenAI. I am the iFlytek Spark cognitive model. It was independently developed by a team of outstanding artificial intelligence scientists, engineers and linguists from iFlytek."

Soon, the "Business Show" answered the questions input by the Spark Cognitive Large Model. This is inconsistent with a recent rumor about the iFlytek Spark model.

Recently, some netizens discovered that the iFlytek Spark model will appear in some "Q&A" with content such as "I was developed by OpenAI". This led to an article about "The iFlytek Spark model is questioned about the cover-up of OpenAI." ChatGPT'" news spread.

On May 11, iFlytek stated in the latest investor relations activity record that the Spark model "shelling OpenAI's ChatGPT" is neither factual nor logical.

iFlytek specifically stated that if it were a shell of ChatGPT, it would be impossible for the iFlytek Spark model to respond faster than ChatGPT; it would not be possible for the iFlytek Spark model to have better text generation, knowledge question and answer, and mathematical capabilities. The results in other aspects are better than those of ChatGPT.

We have reason to believe that iFlytek at this time needs the story of a large model more than any other company, rather than a "shell" accident. Regardless of the merits of the large model itself, iFlytek's rapid refutation of the rumor shows how important the Spark model is to it. The impact of the large models was unparalleled in saving stock prices, both during the quarter and beyond.

At the end of April this year, iFlytek released its 2022 financial report and 2023 first quarter report. The financial report showed that net profit fell sharply. The market seemed to have lost most of its confidence in it instantly, and the stock price fell all the way. It wasn't until May 6 (Saturday) that iFlytek released the Spark model. After the market opened on May 8, its stock price began to rise by 10%. In the following days, it once showed an upward trend. The market seems to have regained confidence in this company.

Launching large models, and then experiencing the incident of being questioned about "containing OpenAI", iFlytek must also face a common AI topic: large models are gathering together, and it is time to rationally disenchant.

In this era of large-scale models emerging, iFlytek is not a "lonely warrior". Since the release of ChatGPT, the technological arms race has never stopped, and its influence has also intensified in China. Various companies have been "rolling" in the field of AI. Some time ago, large-scale models broke out, and they were evaluated as "quite ten years old". The momentum of the pre-Internet 'Battle of One Hundred Regiments' or even 'War of Thousand Regiments'."

According to incomplete statistics, in just 4 months after the release of ChatGPT, at least more than 30 domestic R&D institutions and companies have launched their own brand of large models and related products after the release of ChatGPT.

With many companies claiming to be “domestic firsts”, ChatGPT-like technology has become very common in the entire technology circle, and the capital market has also begun to fluctuate. The media coverage surrounding the big model was overwhelming, followed by a brief silence. And the next wave is most likely still on its way.

The opportunities brought by large models are self-evident, but in this crazy competition, who can really make it to the end? In the decisive battle for the future, what is the real core competitiveness of each company?

Judging from the historical laws of business development, the market will collapse after the hustle and bustle, and the frenzy of large-scale models cannot last forever. To truly seize the opportunities for industrial transformation brought by AI technology, we must understand the core changes brought about by large-scale models and think calmly.

01 Beyond ChatGPT?

Currently, the Spark Cognitive Large Model is far ahead in the country. Its Chinese performance has surpassed ChatGPT, and its English performance is also close to the leading position. Liu Qingfeng said confidently at the iFlytek Spark cognitive model launch conference.

As the fifth company to officially release large models after Baidu, Alibaba, SenseTime, and Kunlun Wanwei, iFlytek couldn't wait to show off its powerful capabilities as soon as it came on the scene.

Liu Qingfeng presided over the entire press conference, and demonstrated a number of capabilities including text generation, language understanding, knowledge question and answer, logical reasoning, mathematical ability, programming ability, etc. with Liu Cong, dean of iFlytek Research Institute of HKUST.

During the live demonstration session, Liu Cong asked: "Why are you called Xinghuo?" "If a young man quarrels with his girlfriend, should he rather die than surrender or be willing to bend and stretch?" "What did Confucius say at the 2008 Beijing Olympics?" Xinghuo quickly "answered" all kinds of weird questions and gave appropriate answers.

In terms of mathematical ability, Liu Cong also asked: "There are three kinds of flowers in the flower bed, a total of 88 flowers. Among them, the number of rose flowers is 4 times that of chrysanthemums, and the number of peonies is 5 times that of chrysanthemums, 2 less. So, may I ask? How many peonies are there in the flower bed?" Such calculation problems. Xinghuo also quickly gave the answer and listed the relevant steps to solve the problem.

Liu Qingfeng immediately said that mathematical ability represents the intelligence of a large model to a certain extent. The iFlytek Spark model is not only far ahead among domestic systems, but has also surpassed ChatGPT.

In this regard, the "Business Show" also conducted multiple rounds of "digging" tests, such as asking "In what year did Yao Ming win the title of Asian Footballer?" "How many years did Fan Zhiyi serve in the NBA? How much did he win? A goal?" and other questions. Xinghuo accurately identified man-made "traps" and gave appropriate answers.

This shows that Xinghuo has undergone certain training on common sense issues and has certain error correction capabilities.

For the same question, "Business Show" was also tested through ChatGPT-3. It also gave appropriate answers, but it was not as comprehensive as Spark in terms of information richness.

In terms of mathematical ability, the "Business Show" also tested "chickens and rabbits in the same cage. There are 25 chickens and rabbits in total, and there are 74 feet in the cage. How many chickens and how many rabbits are there?" "One A stock rose by 10% today and fell by 10% tomorrow. Will you make a profit or lose?" and other questions, Xinghuo also quickly gave the correct answer.

"Business Show" threw the same mathematical questions to ChatGPT-3. In this regard, the results showed that ChatGPT-3's answers were more logical. The latter first gives a conclusion, and then gives specific examples, and there will also be extended questions to answer questions and solve doubts in conjunction with the original question.

Xinghuo also supports voice questioning and output, that is, asking questions by voice, the answers output by the model can also be converted into voice, and the voice style can be adjusted through continuous dialogue.

"Business Show" asks through voice "Chickens and rabbits share a cage. There are 35 chickens and 94 legs in the cage. How many chickens and rabbits are there in total?" "China has won a tennis Grand Slam professional player Who is it?" and other questions. Xinghuo also accurately recognized the voice content and gave corresponding answers.

However, this operation requires the questioner to speak very slowly and speak in standard Mandarin word for word. Otherwise, Spark may not be able to accurately recognize the corresponding text, or the text recognition may be inaccurate, resulting in incorrect answers.

After multiple rounds of testing, "Business Show" believes that Spark has certain large language model capabilities and can output its own business value in certain specific fields.

At the press conference, iFlytek also released a series of products that combine the Spark model. For example, iFlytek's smart office instinct combines real-time voice transcription with ink screen paper-like writing to form a copy of the meeting minutes. Streamlined meeting minutes; "Spark Large Model Intelligent Cockpit" provides multi-wheel, multi-person, multi-region, and multi-modal voice interaction for thousands of models; generative RPA (Robotic Process Automation) based on large models allows digital employees to Smarter...

But Spark is not without its shortcomings. Liu Qingfeng himself also admitted at the press conference, "There are still many shortcomings in large model technology that need to be overcome. These problems include not updating new knowledge in a timely manner, confusing factual questions and answers, and fabricating history and traditions." Plots in culture, etc.” He then mentioned that the above issues will be significantly improved this year.

It is understood that the Spark Cognitive Large Model will start in December 2022. At that time, iFlytek launched "1 N" large model technology research. Among them, "1" is the base platform for general cognitive intelligent large model algorithm development and efficient training solutions, and "N" refers to the application of cognitive intelligent large model technology in education, medical care, human-computer interaction, office, translation and other industries. .

In less than half a year, the Spark Cognitive Large Model was officially launched. The short development time and the hasty model release directly led to many companies, including iFlytek, being questioned by the outside world.

In such a short period of time, the launch of Spark only marks a new starting point, and it needs to go through multiple rounds of iteration and optimization. Liu Qingfeng announced the development plan of the Spark Cognitive Model at the press conference: On June 9 this year, the Spark Cognitive Model will break through open question and answer, and its multi-round dialogue capabilities and mathematical abilities will be upgraded; on August 15, the Spark Cognitive Model will be launched. The model will break through the code capabilities and multi-modal interaction will be upgraded; on October 24, the Spark Cognitive Large Model general model will directly benchmark ChatGPT, with Chinese capabilities surpassing the latter and English capabilities equivalent to the latter.

Judging from the planning date, iFlytek has even specified the date, which may indicate that it is eager to put its large model capabilities into practice and promote commercialization. The industry believes that this may be related to iFlytek’s poor commercialization capabilities in recent years.

02 There is an urgent need for large models to “boost confidence”

Previously, iFlytek’s excellent performance and profitability were often praised by the industry. However, after 10 consecutive years of growth, the myth basically ended in 2022.

iFlytek's 2022 financial report shows that the company achieved revenue of 18.82 billion yuan, a slight increase of 2.77% year-on-year; gross profit was 7.684 billion yuan, a slight increase of 2% year-on-year; net profit attributable to the parent company was 561 million yuan, a year-on-year decrease of 63.94% ; Deducting non-net profit was 418 million yuan, a year-on-year decrease of 57.31%.

This sentence can be rewritten as: Data from iFlytek that has included government subsidies show that by 2022, its government subsidies are expected to reach 1.1 billion yuan. At the same time, this is also the first time that iFlytek's net profit growth rate has declined year-on-year in the past five years. Data shows that from 2018 to 2022, iFlytek's net profit growth rates were 24.71%, 51.12%, 66.48%, 14.13% and -63.94% respectively.

The capital market was quite disappointed with this report card. After the financial report was released, iFlytek’s share price fell by more than 9%.

iFlytek gave three reasons for such a drastic change in performance.

First of all, it is the impact of the general environment. iFlytek said that "due to the special social and economic objective environment in December last year and January this year, as well as the impact of the Spring Festival holiday, some projects were unable to be promoted smoothly and in a timely manner."

Secondly, after being included in the U.S. Entity List in 2019, it was put under extreme pressure again on October 7, 2022. Due to adjustments to the supply chain and related contract signing, the pace of orders in the current quarter was affected.

The last point is related to the Spark model. iFlytek mentioned that the "1 N Cognitive Intelligent Large Model Special Research Project" launched in December last year had an impact on current profits, which confirmed that iFlytek has invested heavily in large models.

Coming to this year, these three major reasons continue to affect iFlytek’s performance.

Data shows that in the first quarter of 2023, iFlytek achieved revenue of 2.888 billion yuan, a year-on-year decrease of 17.64%; a net loss of 57.895 million yuan, and a net profit of 110 million yuan in the same period last year; a net loss of 338 million yuan attributable to the parent after deductions. Yuan, net profit in the same period last year was 146 million Yuan.

According to Jiemian News, iFlytek President Wu Xiaoru revealed that the serious decline in iFlytek’s net profit in 2022 and the first quarter of 2023 is mainly due to the company’s investment in the expansion of cooperation platforms for sustainable operating businesses such as education and medical care, new product research and development, and About 800 million yuan has been added in the independent controllability and localization adaptation of core technologies.

However, iFlytek does not seem too worried about the decline in performance. At the performance meeting, it was stated that based on the current progress in domestic substitution and business development, it is expected to achieve positive growth in revenue and gross profit starting from the second quarter of this year, and is confident to achieve the goal of high-quality growth throughout the year.

iFlytek’s core businesses include educational products and services, information engineering and open platforms. Among them, educational products are its main source of revenue, accounting for 32.74% of total revenue.

iFlytek has also built a business system for three types of customers: G-side, B-side and C-side: G-side is mainly for prefectures, cities, counties and other regions, covering various schools and users in the region with solutions for teaching students in accordance with their aptitude; B The end-end mainly provides big data precision teaching, English listening and speaking classes, smart homework, etc. for schools; the C-end mainly provides products such as AI learning machines, personalized learning manuals, and after-school service courses for parents.

The financial report shows that the G-side business has been applied in more than 50 cities and districts (counties); the iFlytek after-school service business in the B-side business has covered more than 300 districts and counties and more than 12,000 schools; the C-side business Sales of AI learning machines increased by more than 50% that year, but the specific sales volume and amount were not disclosed.

In fact, its performance is still far from the target performance. According to an announcement in early 2022, iFlytek expects its business in various regions to maintain 50% growth. At the same time, it was mentioned that the revenue of the personalized learning manual business is expected to grow by more than 70% in 2022, the AI ​​learning machine revenue target is to grow by 200%, and the target is to achieve annual revenue of 10 billion yuan during the 14th Five-Year Plan.

In addition to the main business failing to meet expectations, iFlytek's smart city, open platform and consumer business, smart cars, smart medical and other business performances are also relatively average.

The financial report shows that the three major sectors of information engineering, digital government industry applications, and smart political and legal industry applications and the open platform business under Smart City have all shown a year-on-year decline. Although smart cars, smart medical care and smart finance businesses are showing year-on-year growth. For example, smart finance has a year-on-year growth of 19.33%, but their proportion in the overall revenue scale is really pitiful. Smart finance only accounts for 1.25%. Smart cars and smart finance have a year-on-year growth trend. Medical care accounted for 2.47% and 2.48% respectively.

It seems that iFlytek’s main business is not performing well, and its innovative business is far from reaching the stage of large-scale revenue. iFlytek currently needs to make full use of the huge value brought by large-scale models to enhance the commercial competitiveness of various businesses.

But it will take time to verify whether the large model that is currently at the forefront can help iFlytek achieve its goals.

03 Big model craze disenchantment time

For several months this year, technology companies around the world have been in a state of near madness. Large-scale model technology has received widespread attention and application at home and abroad. In particular, domestic technology giants have launched a series of their own large-scale model products.

According to incomplete statistics from "Business Show", companies that have launched large-scale model products include Baidu Wenxinyiyan, Alibaba Tongyi Qianwen, Huawei Pangu, SenseTime Ririxin, Kunlun Wanwei Tiangong and HKUST News Feixinghuo, and large model companies founded by Internet tycoons such as Wang Huiwen and Wang Xiaochuan also quickly received financing.

But can such a large model run smoothly? What is the core competitive value of large models? What other disruptive opportunities can big models bring?

The crazy wave of large models has also reached the disenchantment stage of rational examination.

“There are many large model products emerging now, but the cost of training and debugging large models is very high. Most companies can’t afford it. At the same time, self-research is less economical for small and medium-sized enterprises, and competition in the future is more likely. It happens among giants." Dong Hao, an investment manager at a venture capital institution, told "Business Show".

NVIDIA’s research shows that the largest model of GPT3 requires 175 Billions of parameters and requires 7 months of training using 512 V100 graphics cards, or up to a month using 1024 A100 chips. The monthly training cost for large models is in the order of millions of dollars or more.

Last month, at the Artificial Intelligence Large Model Technology Summit Forum hosted by the China Artificial Intelligence Society, Tian Qi, chief scientist in the field of artificial intelligence of Huawei Cloud, also mentioned in his speech that the single cost of large model development and training is as high as 1,200. Ten thousand U.S. dollars.

The development cost of large models is so high, but the application cost (charge) is very low. OpenAI opened its API (Application Programming Interface) in March this year, allowing third-party developers to integrate ChatGPT into applications and services through the API. Its interface service is priced at $0.002 per 1,000 tokens, which is about 90% cheaper than the GPT 3.5 model.

Economic considerations show that it is difficult to recoup this huge R&D cost, let alone achieve profitability. Therefore, small and medium-sized enterprises cannot afford this business, and only large enterprises have enough funds and resources to invest and compete for future market share.

Perhaps because of this, companies that have released large model products have built large model capabilities into their existing mature products. This approach can improve the artificial intelligence capabilities of existing products and attract more customers to purchase specific products, rather than just charging interface service fees. "An industry insider told "Business Show".

But even so, the competition for domestic large models is extremely fierce. As an office field with relatively mature applications of large models, many companies have launched related products. For example, Baidu’s “Wen Xin Yi Yan” has the ability to make PPT; DingTalk can evoke more than 10 people after being connected to Alibaba’s “Tongyi Qianwen” AI capabilities; ByteDance’s office application Feishu will also launch the AI ​​assistant “MY AI”; Kingsoft Office will also launch the “WPS AI” application, etc.

This time, iFlytek also released office products such as voice recorders, translation pens, and office notebooks at the Spark model launch conference, aiming to capture more users in office scenarios. But whether it can really succeed in digging gold depends on the actual feedback from users.

"Although the competition for large models is mainly between giants, it is difficult for the giants to be the only one. The key lies in data. Players accumulate different data in different fields, so they may build their own in specific fields. Core advantages and unique scenarios." Dong Hao further said.

For example, Alibaba has e-commerce data from Tmall and Taobao, and logistics data from Cainiao. This is unique data that other companies cannot obtain. Similarly, iFlytek has been deeply involved in the education field for many years and has accumulated its own Unique data. These data may be the key to truly widening the gap between enterprises.

In other words, it may be difficult for small and medium-sized companies without continuous financial support to make truly universal large-model products. However, the general-purpose large-model products launched by major manufacturers are already eager to consider commercial realization and lack the patience and patience for research and development. technological breakthrough.

Dong Hao said bluntly, "It all happened too fast. Within a few months, all major models seemed to be mature and ready for commercial use, but in fact there must be many bubbles in it."

This fanatical competition for large models has just begun. While giants and companies are entering the game and fighting each other crazily, many people in the industry are also shouting: It is time to return to the original intention, remain in awe of technology, and continue to explore business. Instead of blindly pursuing speed, scale and efficiency, we can usher in the best era of large models.

(The interlocutors in this article are all pseudonyms)

The above is the detailed content of iFlytek responds to "shelling", AI large model craze disenchants. For more information, please follow other related articles on the PHP Chinese website!