"The big model is an opportunity for medium-sized companies that have both more resources than small companies and the determination to be all-in."
Zhao Dongshan, reporter of "China Entrepreneur"
Editor|Li Wei
Image source|Interviewee
After the Spring Festival in 2023, Li Tao, founder and CEO of APUS, put forward a clear requirement in the company - all employees must be in artificial intelligence. The requirements are so stringent that even administrators must consider how AI can improve their jobs. Li Tao even proposed to link the degree of employees' use of artificial intelligence to salary increases.
In the past 10 years, APUS has excelled in mobile phone systems and software. In the face of such a request from the boss, some employees are naturally confused, they do not understand what Li Tao is going to do, just like when Li Tao first met Sam Altman, the founder of OpenAI, he himself did not expect to have such a grand picture in front of him.
In 2019, Shen Nanpeng led a group of entrepreneurs and trainees of Sansi Institute to study in Silicon Valley, in this trip, Li Tao, Wang Xiaochuan, Wang Xing, Zhang Yiming and other emerging entrepreneurs are listed, the important itinerary of the study tour is to visit many Silicon Valley bigwigs, such as Yahoo founder Yang Zhiyuan, Zoom founder Yuan Zheng, etc., one of the stops, they met Sam Altman.
The case Sam Altman showed you at that time was the application of artificial intelligence in game scenarios. "At the time, everyone thought it was cool, but they didn't expect it to be so big." Li Tao recalled to "China Entrepreneur".
Time to November 2022, the advent of ChatGPT made Li Tao feel "cold in the back of his neck", and "glimpse of the world" was his feeling at that time.
Li Tao was surprised to find that ChatGPT subverted the previous way of information interaction, as if communicating with a real person, you can continue to ask, it uses a dialogue robot to constantly give you answers, it even knows what you are thinking. Moreover, ChatGPT has an extremely low threshold and can be used by many people. All this gave Li Tao a strong impact.
From that moment on, Li Tao deeply realized that a new technological change had arrived, a new carnival was about to begin, and as one of the first groups to realize this trend, Li Tao also began his own big model exploration and preparation. On April 18, 2023, APUS released the self-developed 100 billion parameter scale multimodal artificial intelligence model "AiLMe", which was recently opened for public testing.
The dawn appeared
Before excitedly threw himself into the big model, Li Tao was in a great pain. In his own words, "There is no way out." ”
In the past few years, in addition to the objective environment such as the epidemic and changes in the global situation, Li Tao clearly felt the arrival of the end of the Internet era, he watched the people around him speculate on Bitcoin, do cryptocurrencies, and then join Web3, meta-universe, and switch to wave after wave, in Li Tao's view, "the essence is that everyone can't find a way out."
In fact, just a few years ago, APUS had just completed a transformation.
From its establishment in 2014 to around 2019, APUS has been a company mainly engaged in mobile phone system and tool application software development, and its main markets are in emerging regions and countries such as Southeast Asia. After 2019, Li Tao became more and more aware that globalization is the fate of this generation of Chinese companies. In addition to Southeast Asia and other markets, there are many developed country markets to be tapped, such as the United States, Europe, Japan, South Korea and so on.
However, in the face of developed country markets, it is inevitable that the original tools and systems can no longer be used, otherwise there will be no competitive advantage in these markets. To this end, in the past few years, APUS's organizational capacity building and team genes have changed, in addition to the original tool products, Li Tao led the team to make large-scale attempts in the direction of content, transactional, and social products, and emerged small explosions such as PickU and CutCut.
However, those small achievements cannot hide Li Tao's sentimentality about the exhaustion of mobile Internet dividends. The emergence of ChatGPT made Li Tao feel that he saw a light.
In the process of making content products, Li Tao has come into contact with a large number of user needs, just taking the image field where his product CutCut is located as an example, users will have automatic retouching, aging, getting younger, gender transformation and other needs, and in order to meet these needs of users, it is necessary to introduce new artificial intelligence technology. Taking this opportunity, APUS completed the first step of the accumulation of artificial intelligence capabilities.
After experiencing ChatGPT, the ease of use and affordability it showed made Li Tao once again realize the judgment and connection role played by artificial intelligence technology capabilities.
In December 2022, Li Tao convened a large model team within the company and quickly set up a project, shifting all the focus to artificial intelligence and large models. Li Tao also made three conclusions: first, the era of general artificial intelligence is coming, and the operating system of the artificial intelligence era is a big model; Second, APUS must have a large model; Third, you have to do it yourself.
In March this year, APUS was specially spun off to establish the Artificial Intelligence Research Institute, which was led by Zhang Xu, the chief expert of the Institute. Zhang Xu graduated from Tsinghua University with a bachelor's, master's and doctoral degree, and is also one of the main forces of this big model.
Zhang Xu told China Entrepreneur that AiLMe adopts the current mainstream Transformer architecture in the industry, and at the same time designs a set of "plug-in" architecture, which can learn and use "tools" independently, and these "tools" are plug-in connected to the AiLMe platform. In the generation of text, images, video, and audio, AiLMe follows the current mainstream Diffusion Model in the industry and inherits the ecology of the open source community.
Turn around sharply
On April 18, 2023, when APUS released the AiLMe, it was somewhat surprising, because it was the second big model launched by a Chinese company after Baidu.
Li Tao told China Entrepreneur: "There are nothing more than two factors behind it, one is the problem of consciousness and determination, and the other is the problem of resources and ability. APUS is not because it is fast, but because it is early. "Just like Baidu launched the Wenxin model in 2019, Li Tao gathered the company's efforts to build the big model in December 2022.
Li Tao's awareness and determination about the big model, in addition to seeing the seeds planted by Sam Altman in 2019, is also because of his experience in the past 20 years.
Before founding APUS, Li Tao worked as a microwave communication at Datang Telecom, and then started a business with Zhou Hongyi to do 3721, which was later acquired by Yahoo of China; He then joined Qihoo 360 as an executive and successfully launched popular products such as Mobile Security Guard and Mobile Assistant.
Li Tao clearly knows that a new era of technology means huge dividends for operating systems. He has personally experienced the PC Internet era with windows + IE as the operating system, and the mobile Internet era with Apple + Android as the operating system, and he can't bear to miss the upcoming era of general artificial intelligence with large models as the operating system.
What's more, during the period when the mobile Internet went overseas, Li Tao had already proved his judgment once with speed. APUS was founded in 2014, and Li Tao was one of the first entrepreneurs to bring overseas to China's venture capital circle. At that time, based on Launcher desktop system, cleaning security, acceleration and other tool products, APUS achieved the growth of "more than 10 million in 1 month and 100 million in 6 months".
When it comes to resources and capabilities, although APUS is not as strong as the funds of large manufacturers, Li Tao believes that "the AI big model is an opportunity for medium-sized companies, because it has both more resources than small companies and the determination to all in." ”
Li Tao revealed that at present, APUS's annual revenue is more than 1 billion yuan, and the company also has more than 1 billion yuan in cash on its books. In addition, APUS has more than 2.4 billion users worldwide, covering more than 200 countries and regions.
In addition, another factor that APUS moves fast enough is the use of an open R&D model. In addition to the APUS Artificial Intelligence Research Institute, APUS has also formed open source project teams with university laboratories such as the Chinese Academy of Sciences, Tsinghua University, Nanyang Technological University, and The Hong Kong Polytechnic University, so that the construction of large models can be quickly implemented.
High-quality data is more important than computing power
Li Tao divided the factors that affect the quality of large models into four categories: algorithms, computing power, data, and scenarios. In his opinion, if the factors that determine the effect of large models are sorted in order of importance, they should be: training data quality> training data size> parameter scale.
On the technical roadmap, APUS uses large parameters and big data to train medium-scale models.
"AiLMe is a model with 100 billion parameters, but the data behind it is actually 2.2 trillion tokens, they are originally pursuing the so-called large parameters, APUS is slightly different in training, and our scenario dependence will be higher." Li Tao told China Entrepreneur.
In order to solve the problem of computing power, APUS has built two intelligent computing centers in Zhengzhou and Singapore, in addition to the self-built GPU server computing power, APUS also rents and purchases with service providers such as Tencent, Alibaba, and Amazon, and synergizes with third-party computing power.
In terms of data, there are four data sources for the APUS big model: first, public industry data; Second, public data on the Internet; Third, datasets dedicated to training artificial intelligence; Fourth, real user demand data, that is, millions of people use the product every day to generate data and interaction, this part of the weight in model training is also very high.
In Li Tao's view, half of the shock of ChatGPT is attributed to the OpenAI team, and the other half is due to global users, because it is users around the world who open the imagination space of ChatGPT and open up the thinking mode.
Compared with everyone's anxiety about computing power, Li Tao pays more attention to high-quality data.
"This is usually the real needs and problems of real users, we have all been desensitized, it will not involve data privacy security, more attention to the problems and needs of users, as for who you are is not important." Li Tao explained, for example, APUS's "Intelligent Q&A Master" product generates hundreds of thousands of user questions every day, which can be used to train data and models.
For specific application scenarios, APUS distills four vertical models of text, image, video and audio from the AiLMe large model, and develops AI products such as intelligent question answering master, stick figure, ink dyeing, star night, KJV Bible Now, and PicPik directly facing users.
In Li Tao's view, the biggest benefits of application scenarios are two, first, providing high-quality data; Second, provide real-time data.
In terms of business model, Li Tao adopts a short-term to C and medium and long-term to B model. He believes that TOC is easy to form a closed loop, users pay spontaneously, more flexible and more elastic. But in the medium and long term, it must be to B, because the scenarios that a company can cover must be limited, and more industry partners are needed to build together. In the future, it will be everywhere, all services will use it, and the entire market will be more than $10 trillion.
At the same time, Li Tao also predicted that in the future, many industry services will be supported by distilling refined models on the basis of large models, rather than using large models to support them globally, because the cost is too high, and a large model will distill several empirical models to support different services.
From the Spring Festival to the present, APUS internal employees from some people puzzled, some people did not believe, to slowly everyone to try, everything is quietly changing. Today, almost all employees of APUS have the awareness of adopting AI tools, and Li Tao also put forward the promise of "if employees use AI to increase efficiency by 30% by 10%, efficiency by 50% by 20%, and efficiency by 50% by 30%".
In this context, APUS is divided into three groups of people: a group of people focusing on making larger models; A group of people focus on using large models to C-end products; The other group is employees in functional departments who can adopt any AI product. The product selection of internal employees is also forming an invisible pressure, such as Li Tao may ask: "You see that colleagues do not use your products, do you have to work hard?" ”
In the past 20 years, Li Tao has personally experienced the ups and downs of 360, and also witnessed the rise of new giants such as ByteDance and Meituan, and he intuitively judged that the general AI big model will become the mainstream in the next 20 years, "This is a new technological revolution, since I have seen it, I must run ahead of others." ”
Past experience tells Li Tao that the early stage of any technological change is a period of huge dividends, and it is also the time window that creates the most value for society. "For those who do business, always capture the trend of this era." Li Tao said.