news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

"As a technology, big models need to be truly used by the industry to solve practical problems in the industry in order to realize their true value." Guo Wei, vice president of Intel's Marketing Group and general manager of Intel China's Network and Edge and Channel Data Center Division, has pointed out many times in exchanges with Titanium Media APP.

It has been almost two years since ChatGPT was launched. In the past two years, AI big models have penetrated into all walks of life at an unprecedented speed. Especially since this year, both big model manufacturers and client companies have shifted their focus from the size of model parameters to how big models can solve industry pain points, trying to find scenarios where one or more big models can be implemented in their own industries.

From general use to industry, scenarios are the key to the implementation of large models

If the competition among major manufacturers' big models can be divided into the first and second halves, in the first half, major manufacturers launched general big models one after another and demonstrated their strength by continuously improving the parameters of the models; in the second half, when AI big models attracted more attention from industry-side companies, more and more small-parameter models emerged, and industry application scenarios naturally became the focus of attention of Party A companies and Party B big model suppliers. The implementation of applications on the terminal side has become the key to whether the big models can truly be "monetized" on the industry side at this stage.

In this regard, Guo Wei told Titanium Media APP that Intel believes that as AI big model technology gradually penetrates the industry, the model will gradually change from a general big model to an industry-specific big model. "In this process, the size of the model will definitely change," Guo Wei pointed out. "At the same time, the industry model will combine more industry-specific knowledge and focus more on one or several specific scenarios."

Coincidentally, Cao Peng, Chairman of JD Group's Technology Committee and President of JD Cloud Business Unit, also publicly stated at the JD Cloud Summit recently that "general big models are built up with computing power, while industry big models are driven by business." There are already more than 350,000 JD's own delivery staff, more than 200,000 merchants, more than 30,000 doctors, more than 20,000 procurement and sales operations, and more than 10,000 R&D personnel using big model-driven applications in more than 100 AI scenarios at JD.

The 2024 Global Digital Economy White Paper shows that the number of large AI models in the world has reached 1,328, of which China accounts for 36%. On the other hand, IDC data shows that in 2022, China's generative AI accounted for 4.6% of the total investment in the AI ​​market. With the rapid development of generative AI technology, the proportion of generative AI investment will reach 33.0% in 2027, with an investment scale of more than US$13 billion and a five-year compound annual growth rate (CAGR) of 86.2%.

Obviously, just like the previous era of cloud computing, in terms of technology alone, China may still have some gaps from the world-class level. But when we focus on the scenarios, China has rich application scenarios, and these scenarios are the key to AI big models truly "entering the homes of ordinary people."

From the cloud to the edge, how can large models be used better?

As big models gradually move from large parameters of general-purpose big models to small parameters of industry-specific models, they are also showing a trend of moving from the cloud to the edge. IDC data shows that by 2026, 80% of global companies will use generative AI, and 50% of global edge deployments will include AI.

In this regard, Sachin Katti, senior vice president and general manager of Intel's Network and Edge Group, once said that the future of artificial intelligence will depend on an open ecosystem, and the application of artificial intelligence is shifting from data centers to edge computing.

Not only that, Chen Wei, Intel's vice president and general manager of the Network and Edge Business Unit in China, also told Titanium Media APP that judging from this year's development trend, as far as Intel's customers are concerned, more customers are exploring solutions based on edge big models this year. "This year is completely different. This year, it can be said that more than half of our customers are exploring solutions based on edge big models, and there are also many actual cases." Chen Wei further pointed out.

As big models move to the edge, companies need to consider many factors when deploying them, such as latency requirements, practicability, adjustable optimization of micro data, and information security requirements.

However, the deployment of big models on the edge is still in its early stages of development, and there are many deployment models available on the market. "Through Intel's observations, we see that many customers are building hardware and software solutions and optimizing performance based on general big models. This is a rapidly developing model." Chen Wei said this when faced with the question raised by Titanium Media APP on how to better apply big models on the edge.

When talking about the trend and prospects of the implementation of China's big models in the industry, Guo Wei told Titanium Media APP that since this year, more and more industry users have implemented a number of big industry models. "Last year, there may have been more storytelling. This year, we have seen more and more cases of implementation in the industry, and they are all scenarios that can effectively solve the pain points of the industry." Guo Wei said.

From the current application point of view, the model parameters on the end side are much smaller than those of the general large model. At present, the model parameters on the end side are mostly between 7B and 10B. At this time, model compression becomes a very important and difficult task. The current mainstream mode is to inject industry-specific knowledge into it while compressing, which is commonly known as the technical solution of progressive pruning.

In this regard, the relevant technical director of JD Cloud told Titanium Media APP that taking the unmanned logistics vehicle scenario as an example, through quantitative reduction and neural network search, the number of parameters can be compressed by about two times without reducing the performance of the model itself, and the latency can be reduced by about half.

At the same time, Guo Wei shared his analysis on the prospects of China's development of a large industry model with Titanium Media APP. He pointed out that China may develop faster in three aspects.

First, the development of the model itself in industry applications. As mentioned above, "in terms of technology alone, China may still be a little behind the world's first-class level, but China has a wealth of application scenarios." In terms of industry implementation, especially in terms of integration with applications, Guo Wei believes that China can move faster. "A notable feature of the Chinese market is the rapid development of applications, which can quickly explore effective ways to solve industry pain points." Guo Wei emphasized.

Secondly, the improvement of model reasoning ability. When large models are truly applied in the industry, it is not enough to rely solely on model training. To truly solve the pain points of the industry, the improvement of model reasoning ability is particularly important. In Guo Wei's view, model reasoning ability will also be significantly improved this year.

Third, in Guo Wei's view, in terms of the implementation of big models in the industry, it is necessary to coordinate the distribution of computing power between the cloud, the terminal, and the edge. "If it is just some standard applications of vertical big models, the big models may be mainly deployed in the cloud. But due to the needs of industry implementation, it will inevitably prompt AI computing power to be distributed to the edge and the terminal." Guo Wei told Titanium Media APP.

Many scenarios have been implemented

As Guo Wei said, since this year, many application scenarios of large industry models have been deployed on the terminal side.

In the field of education, seewo has jointly launched the "AI+Education" solution based on Intel's technology. With the help of Intel® Core™ Ultra processors, seewo can run large AI teaching models locally without relying on cloud computing resources, thereby improving system response speed and stability and ensuring a smooth experience during the teaching process.

In the retail industry, Tousle's Day has created a smart store solution through edge computing technology. This solution can use the existing audio and video acquisition equipment in the store, and through data integration and optimization management platform, create a set of exclusive digital store models for different stores according to their needs and characteristics, providing different functions from product display, store service, store hygiene, employee norms, to crowd insight and customer flow insight, which enhances Tousle's Day's differentiated management capabilities for store operations, improves management efficiency, reduces the cost of manual store inspections, and strengthens the implementation of employee norms, accelerating the intelligent development of the baking industry.

In the field of healthcare, JD Health launched the "Jingyi Qianxun" medical model based on the JD Yanxi model, which integrates a large number of clinical practice guidelines, medical literature and expert knowledge, and can quickly complete the migration and learning of various scenarios in the field of healthcare. Based on the "Jingyi Qianxun" model, a series of solutions including cloud clinics, post-diagnosis follow-up, expert joint consultation, clinical research, doctor IP brand incubation, and "smart doctor assistant" tools were launched for practicing doctors, which not only optimized the efficiency and quality of doctors' online diagnosis and treatment, but also improved the safety of practice.

In terms of security, many security vendors including 360, Qi'anxin, Tencent Cloud, and Sangfor have launched big model products for the security industry based on big model capabilities, or have implanted AI big model capabilities into existing security products. In the face of the increasingly severe security challenges brought by AI in the AI ​​era, they are striving to achieve the goal of "defeating AI with AI."

In the office, many manufacturers have proposed the concept of AIPC and focused on office AI assistants. In AI Chatbot scenarios, large models can be quickly deployed on Intel Core Ultra AI PCs. Users can smoothly use the powerful text creation, programming, mathematical calculation and logical reasoning capabilities of large models to enjoy a convenient and secure localized intelligent interactive experience.

In terms of document processing, users can use AI PC to efficiently process emails and spreadsheet data, automatically generate work documents, improve work efficiency, quickly draft contract documents, and use AI to intelligently analyze litigation cases to provide professional document writing capabilities.

The above are just a few of the most popular scenarios for AI big models. In addition, industry big models have been widely used in many fields such as finance, manufacturing, and exploration. The era of industries using the capabilities of AI big models to empower their businesses has begun.

Data is still an unavoidable topic

As we all know, data is the "nutrient" for the rapid development of big models. Whether it is a general big model or an industry big model, data is a topic that cannot be ignored if you want to realize real value and reduce "illusions".

As far as the industry is concerned, industry big models require a large amount of industry-specific data. This puts higher demands on the data governance capabilities of enterprises that want to use industry big models to reduce costs and increase efficiency, as well as to transform from digitalization to digital intelligence. From the current industry application point of view, the original data of enterprises is not enough to train a complete enterprise-level industry big model. The main solution is to use synthetic data. Next, on the one hand, enterprises need to improve the quality of their own data. On the other hand, how to select excellent synthetic data has also become a thorny problem that enterprises need to face.

On the other hand, industry big models are different from general big models in that most industry big models need to be deployed on the end or edge. Gartner predicts that by 2025, more than 50% of the data managed by enterprises will be created and processed outside the data center or cloud.

In other words, by 2025, the amount of data on the edge will exceed 50%. In this process, as the amount of data on the edge grows, the bandwidth of data transmission is also a major challenge that enterprises need to face. "Although China is the world's leading country in the construction of transmission bandwidth for the entire infrastructure, when a large amount of data is generated at the edge, it may still cause a network storm. We still need to further optimize network management and data transmission strategies." Guo Wei told Titanium Media APP.

Looking to the future, Sachin Katti told Titanium Media APP that while AI drives data to improve quality, these data will in turn support AI to become better, ultimately achieving a virtuous circle.

Obviously, big models have entered a period of rapid development in the industry. In this process, companies must follow the principle of "data-driven, scenario-oriented" to better implement big models in the industry.(This article was first published on Titanium Media APP, author: Zhang Shenyu, editor: Gai Hongda)