Shulou(Shulou.com)11/24 Report--

Recently, Li Tao, chairman and CEO of APUS, attended the third China (Ningbo) Software Summit and delivered a speech. With the theme of "AI to the Future of New Wisdom", with years of experience in the Internet industry and the analysis and research on the development of artificial intelligence as the anchor, he deeply analyzes the significance of AI to the development of human society, the prospect of industry and the pain point of current industry transformation, and puts forward some innovative viewpoints such as the "six elements" of AI model value creation.

The speech is based on industry insight, with profound ideas, sharp views and vivid cases, which brings new thinking for better application of artificial intelligence and leading industrial change. At the same time, Li Tao expounded APUS's AI strategy and the enabling ability of the AI model in his speech, which brings inspiration for the industrial ecological prosperity and the value creation of the large model.

First, the AI model leads the fourth industrial revolution, the Internet era of low-profit transition to artificial intelligence industry opportunities, sharing the dividend of early technological innovation

It has become the consensus of the industry that artificial intelligence, represented by the large model, is leading the fourth industrial revolution. Looking back on the first three industrial revolutions, steam-electricity-Internet, all have a profound impact on human civilization, especially the golden decade of the development of the Internet, information transmission and perception have been greatly improved.

Human behavior processing mode is closely composed of four links: perception, judgment, decision-making and execution. For a long time in the past, although the Internet has made profound changes in information transmission and perception, the production efficiency in judgment and decision-making has not been greatly improved.

With the emergence of artificial intelligence represented by large models, new industrial opportunities are coming. The ability of AI in judgment and decision-making will bring about a hundred times or even ten thousand times the efficiency evolution, which will determine the industrial opportunities in the next 30-50 years.

From 1999 to today, I have been on the front line of the Internet, like you, feeling the rise and fall of the Internet industry, but also personally feeling the difficulties that developers, programmers and other peers are facing today. It is undeniable that the whole Internet has entered the end of rapid development, and the Internet is in the era of small profits.

But what is exhilarating is that we have also seen a new and more dazzling era, the "AI era", in which human society turns to artificial intelligence comprehensively, rapidly and systematically, and will enjoy unprecedented technological dividends.

Second, the AI big model brings about industrial change. As an operating system, the big model builds the underlying capabilities and services, and uses AI to reduce costs, increase efficiency and create scene value.

In the traditional Internet era, traffic dividend is the entrance, Windows and iOS are operating systems, Internet services are based on PaaS and SaaS layers, and commercialization is dominated by App applications, platform advertising and transformation; in the AI era, AI will restructure all industries, including the Internet.

In the future, artificial intelligence will completely replace the Internet portal represented by search engines, and the big model will serve as the "operating system" of the AI era, building basic capabilities and services for the Internet, industrial Internet, agricultural Internet, Internet of things, meta-universe, etc. At the commercial level, AI is used to reduce costs and increase efficiency to realize the value of the scene.

Objectively speaking, the current development of the agricultural Internet all over the world lags behind relatively. The essential reason lies in the low efficiency, that is, the low efficiency of "decision-making and judgment" that I just mentioned. Without AI empowerment, the agricultural Internet is very difficult to really develop.

The amount of data in the Internet of things is far larger than the amount of Internet data actively built by human beings. In the face of more massive data in the future, it is difficult to achieve the Internet of things without the assistance and efficiency improvement of AI.

In addition, many people think that the popularity of the concept of "meta-universe" has declined, why it has not really developed because of its loud thunder and little rain. It is also because there is no artificial intelligence empowerment.

The construction of every scene in meta-universe may take half a month, a month or even longer. Digital primordial and digital twins can not show the maximum efficiency in meta-space, while the emergence of artificial intelligence represented by large models, it will really make the meta-universe become a reality more quickly.

Third, industrial change is taking place, the industrial elements of IT are facing restructuring, and practitioners need to rethink the choice of career path.

I started writing programs in foreign exchange when I was in college, and today I want to share a new career perspective with IT practitioners. Many people may not be willing to accept it, but more than 90% of programmers should pay more attention to transformation and make up their mind to transform themselves from a traditional, machine-oriented programmer who writes code in machine language to a "AI engineer" facing big models.

In the past, we have mastered from assembly to Python, Java and other language paths are about to close, these machine-oriented programming languages need programmers, product managers, project managers to achieve, write the code and then let the machine understand, and finally form the product.

AI big model will largely replace these traditional professional roles, natural language can easily control the big model, we only need to give the product and module service framework. This leads to a corresponding change in our behavior patterns. In the future, users' simple requirements can be directly put forward to the AI large model, that is, the ability to call the model can be met, while some complex requirements will be realized by AI engineers using natural language to drive the large model.

Fourth, the five-layer pyramid of the AI model structure makes the industrial innovation ideas fall to the ground with high quality.

Large model training needs a lot of computing power, data and scene support. According to the development status and future trend of AI, APUS draws the "five-layer pyramid" of AI large model architecture.

The general large model is trained on the basis of massive data and computing power, and then the vertical application capabilities of text, image, audio and video are refined, and then the general large model is used as the base. Open source technology and service capabilities to e-commerce, manufacturing, education and other fields, and cooperate with thousands of industries to build application scenarios to form an industry-adapted application layer model. Based on this, AI applications are derived to meet the needs of B-end, C-end and G-end users.

What is worth paying attention to is that the quality of training data > the scale of training data > the scale of parameters, the massive data is the "material basis" to make the large model smarter, while the high-quality data with aligned values is the "spiritual guarantee" for the application of the large model. Only by taking the training data seriously can we create a reliable and available AI model.

5. Six elements of AI model to realize value creation: robust computing power, global knowledge base, high-quality data, continuous evolution algorithm, value alignment, value creation.

As we all know, robust computing power is the underlying support for large models to meet the needs of complex scenarios, agile training and high-speed parallelism; as for global knowledge base and high-quality data, we must face up to the fact that in domestic mainstream model training, Chinese datasets account for only 3% of the global corpus and are mixed with network junk data, which has become a yoke for the development and application of domestic large models.

Especially under the influence of uncertain factors, chips such as A100, A800, H100, H800 and RTX 4090 have become the death of the industry, while domestic computing power still needs time to break through. The must-answer question for domestic large model manufacturers is how to maintain the evolution of AI under the limited computing power. The continuous evolution of the algorithm is a key to break the door of dilemma.

As far as APUS itself is concerned, we have recently focused on the incremental pre-training of APUS large models and the input and output of long text and long data. At present, APUS large models already support the input and output of 30,000 words of text length, while the demand for computing power has been reduced by 80%. In other words, the continuous evolution algorithm can reduce the demand for computing power of large models.

I have repeatedly talked about "values alignment". In the land of China, any foreign model should give way to the local model. Why?

The core reason is values alignment, the big model without values alignment can not be safely applied in the land of China. Therefore, in order to achieve a breakthrough in domestic models, we must consider building a red corpus to complete the alignment of values.

Finally, the value creation of artificial intelligence will eventually be combined with the application scenario. APUS sincerely hopes to have in-depth exchanges with developers, programmers, customers, partners and other industries, so that we can really see your usage scenarios, and use artificial intelligence to help you use and improve large models, so that AI models can really drive industrial development and produce value and effectiveness.

VI. Four differentiation advantages derive the AI strategy of APUS: customize the big AI model for China, integrate the application of the big model with value creation, and build AI ecology.

Before All in AI, APUS was an Internet company that has been doing global services for a long time. When APUS transformed into artificial intelligence, we formulated an AI strategy of customizing artificial intelligence models for China, actively building artificial intelligence ecology, and integrating large model applications and value creation.

Behind the strategic setting, where is the strength and ability of APUS? I disassemble and elaborate on four differentiation capabilities:

In the past nine years, APUS has provided more than 200 mobile applications to more than 200 countries and regions around the world, providing services in 25 languages for 2.4 billion users, which has brought a steady stream of global high-quality data feed to the training of APUS models. At the same time, APUS has established two major computing centers in Zhengzhou and Singapore to provide computing support for APUS models and overseas products at the same time.

In addition, APUS also works with Ali, Tencent, and other manufacturers to provide a variety of computing power combinations and flexible expansion solutions to help enterprises and developers achieve rapid R & D landing, significantly reduce costs, and enable flexible industrial collaboration to achieve value landing.

Based on the understanding of the future development of China's artificial intelligence industry, APUS began to accumulate a red corpus a long time ago to align knowledge base, law and other values, so that the output of the APUS model can conform to the mainstream values and avoid position errors and other values deviations in the call of model capabilities.

For the industrial architecture of the APUS model, I think it will be more intuitive and convincing when combined with the scene of the APUS model. Since APUS released the APUS model in April, on the one hand, it has launched a number of AI products such as simple brush painting.

On the other hand, the AI upgrade of the mobile App being used by 2.4 billion users is not only to provide better services for users around the world, but also to continuously bring fresh user feedback and high-quality data sets to the APUS model, making the model "ingredients" more fresh and richer, and making the APUS model smarter and easier to use.

APUS also uses the APUS model as the base to distill four refined models: the text model "Yi Sparrow 8", the image model "Yi Sparrow 3", the video model "Yi Sparrow 4" and the audio model "Yi Sparrow 6" to flexibly meet the application of vertical scenes. For developers and industries, APUS has opened up API and refined modeling capabilities.

For B and G applications, APUS model has been deeply applied in hospital, bus manufacturing and process management, long-term governance of network information industry and other multiple fields through APUS model, APUS network information model, APUS e-commerce model, APUS manufacturing model and APUS education model, which can be used for reference for industrial upgrading.

Taking the current hot e-commerce business development as an example, APUS e-commerce model can support the whole chain to support e-commerce projects to quickly adapt to changes under the AI tide and improve business efficiency.

In intelligent selection, data is used to help merchants screen best-selling products more accurately; in intelligent display, with only one product map, large models can generate three-dimensional displays of different models and scenes, eliminating the need for models. find scene shooting materials and other complicated links and costs; in intelligent delivery, use data and programs to carry out advertisements more intelligently, automatically optimize delivery strategies and execute them quickly and effectively In intelligent customer service, APUS intelligent customer service has played a role in the government government service hotline in first-tier cities, and has become an efficient tool for the government to carry out social supervision. This ability to access e-commerce enterprises can obviously help enterprises to improve service efficiency and optimize customer experience.

In addition, APUS also provides capacity output beyond the AI model, through industrial intermodal transport and investment empowerment, actively build a more dynamic ecological cycle. If the developer has a good product idea, it can be achieved by calling the large model ability; if the product suffers from the lack of promotion ability, APUS can do joint operation together to help developers to create a landing and make the product stand out.

Li Tao pointed out that APUS can also provide financial support to innovative enterprises with market prospects and business models, so that AI applications can have fertile ground for growth and AI industry opportunities to achieve value benefits!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.