Behind one limit after another, can iFLYTEK's performance support today's high valuation?
Author| Shengma Finance Yang Yifan
Edit| Ouyang Wen
On May 6, iFLYTEK Spark Cognitive Model was born, at the press conference, iFLYTEK Chairman Liu Qingfeng bluntly said, "During the iFLYTEK Global Developer Conference on October 24, Xinghuo is expected to be able to comprehensively benchmark ChatGPT: to surpass ChatGPT in Chinese and reach the same level as it in English."
Shengma Finance learned that since ChatGPT became popular, large models in the technology field that claim to benchmark ChatGPT abound, and there are many companies claiming to be "domestic premieres", but after a period of heat, many large models fell silent. So, what is the quality of the Spark model that iFLYTEK has high hopes for? Can the release of the Spark model lead companies with declining performance to fight a turnaround battle?
The evaluation list ranks more than Wen Xin?
On June 14, the mobile version of the Xinghuo cognitive large model was officially launched, and the IOS and Android versions have also been made public, in this regard, Singma Finance has also conducted internal testing through its mini program.
In the assistant center, the application scenarios of the Spark model are quite rich, including high-frequency user scenarios such as workplace, life, travel, writing, etc., and each scenario is subdivided into a dozen to dozens of different functions. In the test, questions such as "Xiaoming's mother has three sons, of which the eldest son is called Dabao, the second son is called Erbao, and what is the third son called" can be accurately answered. For some common sense questions, such as the classification of amino acids, relatively accurate answers can also be made.
Considering that iFLYTEK's own business is mostly focused on the field of education, Shengma Finance increased the difficulty and selected two basic college entrance examination Chinese and mathematics questions, but at this time, the Spark model could not cope with it easily.
Note: The above math questions do not give accurate answers
After several tests, the ability of Xinghuo cognitive large model is not bad, as an assistant for professionals, but it seems to have a long way to reach the intelligence level of ChatGPT 4.0.
In many evaluations of the Xinghuo cognitive big model, Shengma Finance found that the overall list released by SuperCLUE, a comprehensive evaluation benchmark for Chinese general large model, shows that the current GPT-4 comprehensive ability ranks first, and the Xinghuo cognitive big model ranks fourth, with a total score of 76.67 and 53.58 points, respectively; Baidu's Wen Xin Yiyan is 32.61 points, and the MOSS score of the first domestic conversational large model released by the Natural Language Processing Laboratory of Fudan University in China is 36.52 points.
Judging from the score of the list, the Spark model has far left other domestic models. However, Shengma Finance learned that the initiator of the SuperCLUE evaluation is Chinese GLUE Organization (CLUE), which is modeled after GLUE by domestic non-governmental organizations, and its authority and influence are not judged for the time being. However, the main attribute of its WeChat account is individual, and its public account was only updated on May 12 this year, a week after the release of the Spark model, and the last update was in November last year.
In addition, on the day of the release of the evaluation list, its official website information showed that Cui Yiming, who ranked first among the advisors of the Chinese benchmark evaluation members, was the director of the academic advisory committee and a senior researcher of the iFLYTEK Joint Laboratory (HFL) of Harbin Institute of Technology. On May 10, the official website deleted this consultant's information.
iFLYTEK's follow-up "co-investment" pressure is greater
From the test point of view, the initial version of the Spark large model has certain capabilities, but the large model not only needs to be invested in the early stage of white flowers, but also needs continuous and continuous upgrading in the follow-up to maintain a certain level of computing power.
According to NVIDIA's research, the largest GPT3 model requires the use of 175 billions, which requires 7 months of training with 512 V100 graphics cards, or up to a month with 1024 A100 chips. The monthly cost of training large models is in the order of millions of dollars. Previously, Tian Qi, chief scientist of HUAWEI CLOUD artificial intelligence, also mentioned in his speech that the single cost of large model development and training is as high as 12 million US dollars.
At present, the companies that have launched or will soon launch large-model products on the market include Baidu Wenxin Yiyan, Ali Tongyi Qianwen, Huawei Pangu, SenseTime Rixin, Kunlun Wanwei Tiangong, etc. However, compared with large manufacturers such as Huawei, Tencent and Baidu, iFLYTEK does not have an advantage in data and computing power. With the company's revenue and profit volume, it is not dominant in capital, and the future competitive pressure can be imagined.
According to the financial report, in 2022, iFLYTEK achieved revenue of 18.82 billion yuan, a year-on-year increase of only 2.77%; The net profit attributable to the parent was 561 million yuan, down 63.94% year-on-year, which was the first decline in iFLYTEK's net profit in nearly five years.
Since the beginning of this year, the decline in iFLYTEK's performance has continued, and the data of the first quarterly report shows that iFLYTEK achieved operating income of 2.888 billion yuan, down 17.64% year-on-year; The net profit loss was 57.9 million yuan, down 152.25% year-on-year, and the net loss attributable to the parent after deduction was 338 million yuan, compared with 146 million yuan in the same period last year. Singma Finance understands that since its listing, iFLYTEK has only lost its net profit in the first quarter of 2020.
As for the loss in the first quarter, iFLYTEK once said that "the company launched the "1+N cognitive intelligence large model special research" on December 15, 2022, and will officially release the iFLYTEK Spark large model on May 6, 2023, and the new investment in the above research projects has affected the current profit to a certain extent. In other words, the development of the Spark Grand Model has affected the company's profit revenue to a certain extent.
In addition, from the current capital situation, its asset-liability ratio in the first quarter of 2023 was 48.73%, with a total amount of nearly 15 billion yuan, of which notes payable and accounts payable were as high as 6.3 billion yuan. On the other hand, the company's monetary funds are only 32.19 yuan, and the cash flow in the first quarter is -16.684 billion yuan, and the company's financial pressure is greater.
Therefore, in the future, whether the huge business system established by iFLYTEK around the Spark Cognitive Model can operate in the future is a question mark, and the most urgent thing at present is how much financial resources iFLYTEK still has to support the subsequent R&D and upgrading of the Spark model.
The performance did not rise, and the stock price went first
Although the Spark model is still in the research and development stage and the commercialization process is far away, the management of iFLYTEK knows how to "create momentum" in the capital market.
Singma Finance noted that when iFLYTEK announced the launch of the big model, the company had just handed over an ugly financial report. Affected by this, on April 21, iFLYTEK's share price plunged by 9%. Subsequently, iFLYTEK publicly stated that it was about to launch the cognitive large model "Spark", and iFLYTEK's stock price rose again.
In the days before and after the launch of the mobile version of the Xinghuo Cognitive Model, iFLYTEK's share price rose by nearly 17% again (June 9-14), and the closing price reached 75.13 yuan on June 14.
But on the other hand, the company's major shareholders are quietly reducing their holdings. On March 22, iFLYTEK simultaneously announced the reduction plan of major shareholders and the change of the rights and interests of actual controllers.
According to the announcement, iFLYTEK's major shareholder China University of Science and Technology Asset Management Co., Ltd. plans to reduce its holdings by no more than 23.2308 million shares by collective bidding or block trading from April 21 to October 12, accounting for 1% of the company's total share capital. The actual controller Liu Qingfeng and the major shareholder "China University of Science and Technology" acted in concert, and the change in other voting rights owned by the actual controller Liu Qingfeng led to a decrease of 24.3142 million shares.
In fact, as early as February 8, 2021, iFLYTEK proposed at the annual planning meeting of the enterprise that it should achieve the goal of "one billion users, achieve 100 billion revenue, and drive a trillion industrial ecology" in the next five years, while iFLYTEK's revenue in 2020 was only 13 billion yuan.
As soon as this news came out, iFLYTEK's stock price first pulled two limit boards. Similarly, at the annual meeting in February 2022, Liu Qingfeng announced the launch of the "iFLYTEK Super Brain 2030 Plan", according to which family companion robots will enter thousands of households in 2030. Not surprisingly, iFLYTEK's stock price rebounded again.
However, when it comes to the company's financial reports, the company is still a long way from achieving the above goals. Behind one limit after another, whether iFLYTEK's real performance level can support today's high valuation needs to be viewed rationally.
END
Follow us to read more
Shengma Finance sincerely original, unauthorized reproduction is prohibited