Ma Huateng believes that AI is an opportunity similar to the industrial revolution that invented electricity that has not been encountered in hundreds of years. But Liu Qingfeng is different, and Xinghuo iFLYTEK will surpass ChatGPT this year.
©️ Understand the original of Cai Di · Author|Yifan
The history of iFLYTEK's growth is also a history of controversy.
In 2008, its revenue was only 258 million yuan, and its market value at the beginning of its listing was only 3.2 billion yuan.
In 2022, its revenue was 18.8 billion yuan, and now its market value has soared to 170 billion yuan.
In just 14 years, its revenue has increased by about 73 times, and its market value has increased by about 50 times.
In the long run, this company, which has been jumping up and down sharply, has brought investors two completely different evaluations: high-growth high-tech companies and pompous general technology companies.
The company's chairman and founder, Liu Qingfeng, is legendary and controversial.
The founder seems to like chasing the wind and talking nonclessly.
His latest crazy remark was that on October 24 this year, iFLYTEK's Spark cognitive big model will surpass ChatGPT in Chinese capabilities and reach the level of ChatGPT in English.
More than 50 large model companies in China, many of them brush the sense of existence, many companies have threatened to surpass ChatGPT, not Liu Qingfeng alone. Baidu Li Yanhong and Wang Xiaochuan pinched each other and also earned enough eyeballs.
Beyond ChatGPT and specific standards, no one can tell, at present, this is just a "war of words".
It doesn't matter whether you can surpass ChatGPT, the important thing is to brush up the sense of existence, iFLYTEK's market value has more than doubled, the market value of 360 has tripled, the red-clothed sect owner is busy with intensive large model press conferences and does not forget to divorce, and the wealth brought by large models is also a once-in-a-century event.
Compared with the story of "transcendence", Emperor Zhicai pays more attention to the e-commerce 618 sales ranking of iFLYTEK AI learning machine, after all, Liu Qingfeng shouted "beyond" to cheer, and the learning machine is better sold.
1 | Base color
OpenAI sharpened its sword in eight years, and ChatGPT became famous in one fell swoop, and its valuation soared to $29 billion.
Microsoft, which invested in ChatGPT, is connecting all its software to ChatGPT, adding trillions of dollars to its market value.
Nvidia, which supplies chips for large models, has a trillion-dollar market capitalization of more than 190 percent this year, and its founder said it was AI's iPhone moment.
Musk, the original organizer of OpenAI, seems to be panicking about ChatGPT, publicly signing a petition to slow down the development of GPT-4.
Capital and the industry have reached a consensus that the big model is real technology, rather than the blockchain, meta-universe, VR and many other outlets that were previously flocked to and abandoned by Chinese technology companies.
In April, more than 60 large models emerged in China, seemingly overtaking ChatGPT overnight.
Tencent Pony, which has always been low-key and late-handed, is also laying off employees on the left hand and increasing AI with the right hand. However, unlike dozens of large model companies in the previous one, he is not in a hurry to release test products.
Unlike Liu Qingfeng, who will overtake ChatGPT this year, Pony is much calmer: "At first, we thought that this was an opportunity for the Internet to not meet in a decade, but the more we thought about it, the more we felt that this was an opportunity similar to the industrial revolution that invented electricity in hundreds of years, so we felt that (AI) was very important."
Time does not wait for anyone, Liu Qingfeng wants to compete for this year.
Liu Qingfeng made a high-profile declaration: "On October 24, the Spark cognitive big model will surpass ChatGPT in Chinese ability and reach the level of ChatGPT in English."
This also means that iFLYTEK Spark will overtake ChatGPT in 130 days.
This is not Liu Qingfeng's first goal. He once said that from 2021 to 2025, it is necessary to achieve the goal of "one billion users, achieve hundreds of billions of revenues, and drive a trillion industrial ecology".
In 2021, iFLYTEK's revenue was 18.314 billion yuan, and in 2022, it was 18.820 billion yuan, a year-on-year increase of only 2.77%. To achieve this goal, it means that iFLYTEK's revenue will increase fivefold by 2025, with a compound growth rate of more than 70%.
Why is 2022 falling short of expectations?
iFLYTEK explained that due to the impact of the epidemic, bidding was delayed, especially in the fourth quarter, when more than 20 projects and more than 3 billion yuan of contracts were postponed.
Liu Qingfeng was still optimistic at the results conference: "If it were not for the epidemic, we might have increased by more than 560%."
Such performance goals may really rely on large models to achieve.
From the perspective of net profit, iFLYTEK is not ideal.
Its decade-long performance myth has been shattered, and it even suffered a net profit loss in the first quarter of this year.
In 2022, the company's net profit was 561 million yuan, a year-on-year decrease of 63.94%. What is shocking is that this data is the performance data of iFLYTEK after entering government subsidies. In 2022 alone, iFLYTEK's government subsidies amounted to RMB1.1 billion.
In 2019, 2020 and 2021, iFLYTEK's government subsidies included in the current profit and loss were 412 million, 426 million and 439 million, accounting for 50.24%, 31.23% and 28.21% of the total profit, respectively.
In fact, this is the first time in recent years that iFLYTEK's net profit growth rate has declined. In the past five years, iFLYTEK's net profit growth rate has been 24.71%, 51.12%, 66.48%, 14.13% and -63.94%, respectively.
Last year's lackluster performance seems to have spilled over into this year.
According to the financial report, in the first quarter of 2023, iFLYTEK achieved operating income of 2.888 billion yuan, a year-on-year decline, and a net profit loss of 58 million yuan, compared with the profit of 110 million yuan in the same period last year.
Although iFLYTEK said that the dismal performance in the first quarter of this year was mainly due to the increase in R&D funds, it is difficult to hide the current situation that its performance has fallen into a dilemma and its revenue is weak.
Distrustful people began to doubt iFLYTEK's hematopoietic ability again.
The 100 billion revenue plan is in vain, relying on subsidies to make a profit. In 2023, the new story of iFLYTEK needs a large model, and it will surpass ChatGPT in about a year, which inevitably raises questions.
2 | Controversy
iFLYTEK's growth has always been accompanied by turmoil and doubts.
iFLYTEK started out through speech recognition and speech synthesis, then moved into the AI track and gradually grew into the first AI stock.
By dismantling its revenue composition, it can be found that education products and services, information engineering and open platform are the three most important revenue items of iFLYTEK.
Among them, "education products and services" are mainly sales of teaching hardware products, accounting for 32.74% of the total revenue.
In the education track, iFLYTEK's route is relatively successful: first obtain recognition on the G-side, then promote it to major schools (B-side), accumulate data advantages in subdivided industries, and then expand business to the C-end through schools, such as selling hardware products and services.
iFLYTEK's "Technology Open Platform" landing application story focuses on education and comprehensively covers "tracks" such as smart cities, smart hardware, smart cars, smart healthcare, and smart services, constituting a "platform + track" story.
The emergence of large models has undoubtedly filled its platform with unlimited imagination.
On May 31 this year, iFLYTEK, Huawei, Baidu, and the Institute of Automation of the Chinese Academy of Sciences were awarded by the China Academy of Information and Communications Technology as the deputy leader of the large model working group, and will work with the Academy of Information and Communications Technology to promote the development of the domestic large model industry.
This shows that iFLYTEK's large model research and development has been recognized by the government.
However, founder Liu Qingfeng's non-stop posture seems to make more investors more confident in iFLYTEK, but on the other hand, some people are more skeptical.
On May 24, two fictional "small essays" by AI flashed the stock price of iFLYTEK, causing it to fall by more than 9% at one point. These two essays argue that iFLYTEK is involved in illegally obtaining user data.
Although iFLYTEK denied it and said it was responsible, it still had a negative impact on its reputation and stock price, which attracted a lot of attention from the market.
The head of marketing at Wenxin Yiyan, a subsidiary of rival Baidu, said: "It's not easy for China to develop its own large-model technology, so it's not easy to focus on business, and ask friends to solve their own problems, don't touch others at every turn."
Now, iFLYTEK released iFLYTEK Spark, pointing to the large model track, just when people rekindled their confidence, iFLYTEK Spark was exposed to the "shell" OpenAI's ChatGPT, and the market was in an uproar.
Although iFLYTEK quickly refuted the rumor and said that the Spark model "ChatGPT of OpenAI in the shell" is neither factual nor logical. The reply said: "If it is a shell ChatGPT, it is impossible for the iFLYTEK Spark model to respond faster than ChatGPT; There will be no situation where the results of the iFLYTEK Spark model are better than ChatGPT in terms of text generation and mathematical ability."
However, this did not completely calm such voices, and there were many doubts in the market, and some people even ridiculed "I think not, because (Spark) is too different".
3|Competition
Since the beginning of this year, a tense atmosphere has enveloped tech companies around the world. Domestic and foreign technology giants, around the big model, launched a life and death battle.
The battle for big models is a war that burns money.
Microsoft has injected more than $13 billion into OpenAI, which means that ChatGPT's research and development expenses will be billions of dollars. Last year, OpenAI posted a net loss of $544.5 million, and it spent $400 million on computing power and data alone.
According to a16z estimates, a company's annual AI computing expenditure is only more than $50 million, which is enough scale effect to support its own mass purchase of GPUs.
According to LatePost, after the Spring Festival this year, China's major Internet companies with cloud computing businesses have placed large orders with NVIDIA. Byte has ordered more than $1 billion in GPUs from Nvidia this year, and another large company has also ordered at least more than 1 billion yuan.
However, iFLYTEK's annual report data shows that its R&D expenses in 2022 will be 3.111 billion, and this 3.1 billion yuan is not only for the research and development of iFLYTEK.
iFLYTEK's R&D investment in 2022 wants to catch up with the leading ChatGPT, which has made many investors question.
The big model competes around three elements: computing power (chip), data, and model.
From the chip point of view, the chip most likely to be used by the Xunfei Spark model is the Cambrian chip. China's most representative local AI chip, Cambrian chip has always had a good performance, in 2016 launched the world's first commercial terminal intelligent processor 1A, which can be integrated into the terminal SoS chip, can process 16 billion virtual nerves per second, peak computing capacity of 2 trillion times per second.
TrendForce data shows that based on the processing performance of the NVIDIA A100 graphics card, training ChatGPT will require the use of 30,000 A100 (unit price of more than $10,000). From the perspective of the scale of GPT parameters of each generation, GPT-1/2/3 has 1.2/15/175 billion training parameters, respectively.
According to OpenAI, GPT-4 will use more computing resources than GPT-3. This means that if iFLYTEK Xinghuo wants to surpass, it will need more powerful computing power support.
Unlike iFLYTEK Spark, Baidu and Ali's chips are independently developed, "Wen Xin Yiyan" uses Kunlun chips, and Ali uses Damo Academy independent research and development chips. Among them, the Kunlun chip can respond to 1 trillion deep semantic matching and logic every day.
From the data level, the tracks in the data sets of the three giants are different. Specifically, iFLYTEK's data is mainly concentrated in the field of education, while Alibaba's data is concentrated in the field of e-commerce, and iFLYTEK chose to land the application in the education track.
In fact, the complexity and scale of a large model directly depend on the amount of data required for model training. As a large company that can participate in the big model strangling battle, whether it is iFLYTEK, Baidu, Alibaba, it has quite rich data accumulation, how to make good use of its own data advantages and form a large model with its own characteristics is the key.
From the perspective of the model, the Spark model focuses on multi-task, multi-style long text generation, multi-level cross-language and cross-language understanding, pan-domain open-ended knowledge question and answer, situational logical thinking reasoning and multi-question analytical mathematical ability, multi-functional multilingual code ability, and multi-modal input and expression ability.
Wenxin's ability comes from ERNIE-ViLG, a cross-modal large model of Wenxin, which is a large language model independently developed by Baidu; Tong Yi Qianwen uses deep learning techniques, including recurrent neural networks (RNNs) and long-short-term memory networks (LSTMs).
It is worth mentioning that iFLYTEK has accumulated good strength in the past, relying on iFLYTEK has rich experience and strong talent reserve in Transformer deep neural network algorithm.
In 2022, iFLYTEK won 16 championships in the key core technology areas of artificial intelligence, including 13 international competitions in cognitive intelligence such as OpenBookQA, QASC, and ReClor.
On June 9, iFLYTEK launched the Xinghuo App and iFLYTEK Spark Mini Program that can conduct multiple rounds of AI conversations, and updated the Xinghuo cognitive large model and launched V1.5.
According to Liu Qingfeng's plan, there is still half a year before iFLYTEK surpasses ChatGPT. And more people expect that iFLYTEK and all China's big models are getting better and better, but don't fall into a "war of words".
After all, Chinese people need to embrace this unprecedented era of AI revolution more calmly and ideally.
Note: The data is based on public disclosure and does not constitute any investment advice, investment is risky, and caution is required when entering the market.
—END—