Opportunities and challenges faced by domestic large models

Source: Economic Daily

Recently, six departments jointly issued the Action Plan for the High-quality Development of Computing Power Infrastructure, which proposes to create a number of new computing power businesses, new models and new formats, and deepen the application of computing power to empower the industry. In recent years, as a representative of the application of computing power, generative artificial intelligence technology has continued to develop, and a variety of large-model products have competed to appear. Massive application scenarios have become a high-quality "test field" for the development of artificial intelligence in mainland China, and a large number of new technologies, new applications and new breakthroughs have accelerated their growth and released the potential of computing power applications.

At present, what is the development trend and application effect of domestic large models? How to apply in more scenarios to better promote the development of the real economy? What are the challenges to development?

Since the beginning of this year, artificial intelligence has become a hot field in the global technology industry, and hundreds of companies and institutions have successively released products related to large language models, and artificial intelligence applications have opened up a new situation in the field of large models.

In this global competition, the continent is at the forefront of the trend. The development of domestic large models accelerated, and in just half a year, they walked out of the laboratory and were tested for the public. iFLYTEK's "Spark", Tencent's "Mixed Yuan", Baidu's "Wen Xin Yiyan"... The data shows that at least 130 companies in the mainland research large-model products, of which 78 are general-purpose large-model companies.

Not only big models, but also new artificial intelligence technologies are accelerating into thousands of industries. Provide personalized recommendations for consumers, predict the weather, guide mine production, help programmers "write code", help scientists "engage in scientific research"... The rich application scenarios vividly reflect the artificial intelligence to empower industrial development and benefit life.

What are the opportunities and challenges for generative artificial intelligence (AIGC)? How to make the big model develop healthily and better serve us? In the past few days, the reporter has visited many industry companies and experts to find answers.

Applications accelerate implementation

On August 31, the first batch of large models that passed the "Interim Measures for the Management of Generative Artificial Intelligence Services", Baidu Wenxin Yiyan, Baichuan Intelligence, and SenseChat were successively announced to open to the public. At the same time, more large models of enterprises are also rapidly deploying and launching. The mass listing of domestic large models means that the "100-model war" is moving from the "birth" of the previous stage to a new stage of "use".

After opening, the big model is even more popular.

On August 31, the first day of Wenxin's opening, Baidu's official platform data showed that in just 24 hours, Wenxin Yiyan replied to more than 33.42 million questions from netizens, with an average of more than 23,000 questions in one minute. On the same day, the Wen Xin Yiyan APP appeared in many app store hot lists.

On September 5, the Xunfei Xinghuo cognitive big model was opened to the whole people, and the number of users exceeded 1 million within 14 hours of launch.

On September 7, the Tencent Mixed Yuan model was officially unveiled. Jiang Jie, vice president of Tencent Group, said that Tencent will fully embrace the big model. At the same time, it was also announced that it will open up the capabilities of the hybrid model through Tencent Cloud to serve thousands of industries with self-developed technology. The Tencent Hybrid Model will serve as the foundation of Tencent Cloud MaaS services, which can be used by enterprise users not only directly, but also as a base model to build exclusive applications for different industry scenarios.

In addition to openness, there is also open source. On September 7, Ant Group officially opened up the graph learning system Ant Graph Learning (AGL), which is the industry's first general industrial graph learning system.

Large models are openly used and tested by the public, which brings convenience to people's work and life in many aspects, such as helping to learn knowledge, modify articles, and generate solutions.

What is the future trend of big models? Liu Cong, vice president of iFLYTEK and president of the research institute, said that the big model after opening up may change the work and life of ordinary people. In the next few years, general cognitive intelligence technology will continue to develop rapidly, bringing major model innovation and industrial changes, such as changing the mode of information distribution and acquisition, revolutionizing the content production model, completing tasks with natural interaction, realizing expert virtual assistants, subverting traditional manual programming methods, and becoming a scientific research work accelerator.

"The value of a big model is in the application. Only by empowering the development of intelligent economy and intelligent society in diversified practical application scenarios can we find industrial value and achieve the big model itself. Deng Zhidong, director and professor of the Visual Intelligence Research Center of the Institute of Artificial Intelligence of Tsinghua University, said that models and algorithms, data and knowledge, chips and computing power, scenarios and real industrial application requirements are the core strength of the development of the digital ecosystem and the key factors of big model competition.

After opening up, under the "feeding" of massive data, large models will become more and more "smart", how can mature AI empower hundreds of industries and transform excitement into productivity to promote development?

Liu Yuanchun, president of Shanghai University of Finance and Economics, said that for general artificial intelligence, the long-term value of large models will be realized through industry applications, and application scenarios are the key. As the world's second largest economy, Continental has the advantages of ultra-large-scale market and digital resources, and has made great achievements in digital technology and application, providing deeper and richer landing scenarios for large models and an environment that can "continuous training and lifelong learning". At the same time, this also requires the large model to be more practical and help the development of the real economy.

"The general large model has great capabilities, but it does not solve the specific problems of many enterprises. Building your own exclusive model based on the large model of the industry may be a better option for enterprises. Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of Cloud and Smart Industry Business Group, said, "With the development of big language models, industry and society will also move from digitalization and networking to intelligence. The fundamental goal of artificial intelligence development is to land in the industry and serve people. Enterprises that can truly solve user needs and are closer to scenarios and data will have a future of big models. ”

AI empowers the industry

"AI's potential benefits to the global economy will reach $25 trillion, which is one of the most important tracks for enterprises today, but this era has just begun." Daniel Ni, chairman of McKinsey China and global senior managing partner, said that the technology storm of generative AI is expected to open a new round of technological and industrial changes related to the next 8 to 10 years. Generative AI will boost productivity and provide new jobs.

Yutong Zhang, managing partner of GSR Ventures, predicts that AIGC will unlock huge value potential in key corporate functions, with the four most affected business functions being customer operations, marketing and sales, software engineering, and product development. In the gaming industry, for example, AIGC will optimize internal efficiencies and subsequently reshape the development process and lead gameplay innovation.

At the Tencent Global Digital Ecosystem Conference, Jiang Jie demonstrated the application capabilities and innovative exploration of Tencent Meeting, Tencent Docs, Tencent Advertising and other products and businesses after accessing the Tencent Hybrid Model. He painted a future work scene - after hours of online meetings with multiple people, artificial intelligence technology can quickly and accurately extract the views of all parties and intelligently generate meeting minutes through simple natural language instructions; We will see more and more advertising creatives generated by large model technology, text, images, videos to achieve natural integration...

"The application of large model technology is translating into real productivity. Within Tencent, Tencent's hybrid model has deeply supported more than 50 businesses. Jiang Jie said.

However, although the mainland model and its industry scale have advantages, the current application mainly focuses on leisure scenarios with high fault tolerance rate and simple tasks, and the application in more valuable serious scenes, work scenes, and professional scenes needs to be further expanded.

Up to now, SenseTime has established in-depth cooperation with more than 500 customers in vertical industries such as finance, healthcare, automotive, real estate, energy, media, and industrial manufacturing, providing customers with large-model AI technologies and services by providing a variety of flexible API interfaces and services, and realizing various generative AI applications with low threshold, low cost and high efficiency. "In the past six months, large models and generative AI have been arguably the world's most high-profile technological breakthroughs, and SenseTime has ushered in a critical period of development. We want to bring stronger big model capabilities to the industry and help users produce disruptive products. Xu Li, Executive Chairman and CEO of SenseTime, said.

Tang Daosheng believes that "with large model generation technology as the core, artificial intelligence is becoming the key driving force for the next round of digital development, and it has also brought new ideas to solve industrial pain points." Big models need to be based on industry scenarios and integrated with enterprise data to unlock maximum value."

For artificial intelligence technology, the time of launching the product is not the most important, the key is to do a solid job in the underlying algorithms, computing power and data. Fortunately, on the track of large models, the mainland has more than 100 domestic large models competing, reflecting the importance that science and technology enterprises attach to scientific and technological innovation. This is the core password of the domestic large model to strengthen itself.

Jiang Jie introduced that Tencent's hybrid model is trained from scratch, and the company has mastered the full-link self-research technology from model algorithms, machine learning frameworks, and artificial intelligence infrastructure. "Since 2021, Tencent has successively launched a number of tens of millions and hundreds of millions of parameter models, continuously optimizing the underlying algorithm development of large models in practical applications and improving engineering capabilities."

On September 8, at the Bund Conference held in Shanghai, Ant Group officially released the financial big model and open-sourced the generative AI programming platform CodeFuse. He Zhengyu, chief technology officer of Ant Group and president of the platform technology business group, said that Ant Model takes the pure self-developed technical route, takes full-stack layout and long-term development as the principle, and aims to create industrial value. To this end, Ant has invested heavily in the underlying infrastructure of large models, and has built a Vancard AI cluster, leading the industry in training efficiency, and providing strong support for the industrial application of large models.

"In the current international environment, full-link self-research of large models is one of the best technical paths for the development of general artificial intelligence in the mainland." Liu Yuanchun said that with the help of full-link self-research, we should continue to accumulate talents and technologies, gradually form a systematic industrial chain, talent chain, technology chain and innovation chain, and finally form an endogenous dynamic, walk out a Chinese path for the development of general artificial intelligence, and help digital technology innovation make breakthroughs.

Liu Cong believes that after the centralized launch of large models, there will be a situation in the future that "the leftovers of general large models are king, and vertical large models bloom in a hundred flowers". Leading enterprises in subdivided industries with data and understanding scenarios should cooperate with platforms that can provide general cognitive intelligence large models, and take advantage of their advantages of safety and controllability, training optimization and self-iteration capabilities to polish their products for long-term sustainability.

"If the cognitive big model wants to achieve in-depth application in the industry, it needs to meet three key elements: first, it is safe and controllable, including content security and algorithm security; The second is scenario-driven, which should produce real application value in visible and tangible scenes, and use statistical data to prove application results; The third is to have an exclusive model that can protect users' proprietary data, establish a private cloud platform, and ensure customer data security. Liu Cong said.

Liu Dian, associate researcher at the Institute of China Studies of Fudan University, said that the development of artificial intelligence is not achieved overnight, and it is necessary to continue to strengthen investment in artificial intelligence, adhere to independent research and development, and be self-reliant.

Scale effects, large markets, and rapid data feedback can motivate enterprises to continue to invest and iterate efficiently, giving artificial intelligence more use, thereby affecting economic development and industrial structure.

Not long ago, the China Academy of Information and Communications Technology, the Fifth Research Institute of Electronics of the Ministry of Industry and Information Technology, Huawei and other units jointly established the "Large Model Industry Working Group" to jointly promote the application of China's large model and industrial incubation. Huawei, together with Facewall Intelligence, Zhipu AI, iFLYTEK, and Cloudwalk Technology, released the Ascend AI large-model training and promotion integrated solution, building a large-model super factory and accelerating the application of large-model in various industries. 26 industry leaders, research institutes, and Huawei will jointly innovate the application of basic and industry large models based on Ascend AI.

Safety is long-term

The openness of large models and the in-depth application of AIGC not only put forward higher requirements for data, algorithms, and computing power, but also pose more challenges to security, privacy, and ethics. Only under the premise of ensuring data security and privacy protection, and improving the ethics and security of artificial intelligence can AI technology truly release the application value.

"We prefer that AI be a human-friendly aid that assists humans in some challenging tasks." Zhu Jun, deputy dean of the Institute of Artificial Intelligence at Tsinghua University, believes that "but it does not necessarily need to align with humans in form and behavior." ”

Artificial intelligence is a new field of human development. Scientists have long feared that AI is out of control. In this regard, Qiao Yu, a professor at the Shanghai Artificial Intelligence Laboratory, said: "We not only need the participation of the artificial intelligence community, but also need to bring in social science scholars to jointly establish a large model framework of artificial intelligence to ensure that it is in line with human values." ”

He Jifeng, an academician of the Chinese Academy of Sciences, reminded that the ability to generate large models diversifies privacy leakage methods and makes privacy protection more difficult.

The East China Branch of the China Academy of Information and Communications Technology, OPPO and Xi'an Jiaotong University recently released the white paper "Research on Trustworthy AI Security and Privacy Technology". Lin Chenhao, a distinguished researcher at the School of Cyberspace Security of Xi'an Jiaotong University, said that at present, the establishment of trusted AI related standards in the mainland is still in the initial stage, and AI technology is still in the initial stage of industrialization, and it is recommended to accelerate the establishment of trusted AI standards and regulatory systems, pay attention to the cultivation of trusted AI technology talents, promote the solution of trusted AI hot and difficult problems, build a trusted AI security level assessment and defense system, and build a trusted AI ecology for the whole industry.

Despite the confusion and challenges, industry practitioners remain optimistic about the future of big models.

Duan Runyao, director of the Institute of Hundred Metric Computing, said that the combination of quantum computing and artificial intelligence can efficiently and quickly solve problems that cannot be solved at present, "What we need to do is to identify really key application scenarios to make quantum computing worthwhile."

"With the opening up of big models, everyone can have 'superpower' AI assistants, freeing them from tedious labor and doing more valuable and creative things." Liu Cong believes that in the future intelligent world, AI will achieve comprehensive resource inclusion. Jing Xiandong, chairman and CEO of Ant Group, said that the big model is not just a new technology, but a new world. (Economic Daily reporter Li Zhiguo)