news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Source: Xianning News Network

On July 19, the 2024 China Unicom Partner Conference was grandly opened in Shanghai. At the main forum of the conference, Jian Qin, general manager of China Unicom Group, released China Unicom's artificial intelligence innovation achievement - Yuanjing 2.0; at the sub-forum on artificial intelligence innovation and development, Zhu Changbo, president of China Unicom Digital Technology Co., Ltd. and director of China Unicom Artificial Intelligence Innovation Center, delivered a keynote speech entitled "Unicom Yuanjing: A big model that better understands the industry, an intelligent engine for industrial upgrading", interpreting the capabilities of Yuanjing 2.0.

Director Zhu Changbo introduced that since the release of MetaView 1.0 at the 2024 Barcelona World Mobile Communications Conference, China Unicom Artificial Intelligence Innovation Center has launched MetaView 2.0 after half a year of technical research and application practice, achieving four capability upgrades: base capability upgrade, MaaS platform upgrade, security capability upgrade and industry application upgrade. These four major upgrades have given MetaView 2.0 a stronger engine, forming the distinctive features of "easier to customize, better understand the industry, and more trustworthy". At the forum, Director Zhu Changbo also released three major basic models: the 204 billion Yuan Jing multimodal large model, the Yuan Jing Wenshengtu large model, and the Yuan Jing voice large model; the Yuan Jing MaaS platform-RAG (search enhancement) and the Yuan Jing MaaS platform-intelligent body core components; as well as 35+ industry large models and hundreds of excellent cases.

Director Zhu Changbo emphasized that in the intelligent world, the model base capability is like the foundation, which determines how far and how high the model application can go. Yuanjing 2.0 attaches great importance to the core technology research of the basic large model. First, it has achieved a leap from the scale of model parameters from billions to hundreds of billions and trillions, and the trillion-parameter MoE large model has been basically trained; second, it explores the evolution from loosely coupled multimodality to native multimodality, so as to achieve a silky effect like GPT-4o in multimodal dialogue.

The three basic models released by Yuanjing 2.0 include: 204 billion parameter Yuanjing multimodal large model, innovative design uses composite visual encoding module, improves the precision of model perception, solves the shortcomings in accurate counting, spatial perception, precise reasoning, etc., and enables Yuanjing large model to play a key role in more deterministic scenarios; Yuanjing literary image large model innovatively uses long sentence encoding module and cascade diffusion architecture, which can realize highly controllable Chinese literary image, and achieve breakthroughs in Chinese long sentence understanding, local strong controllable modification and Chinese text generation capabilities; Yuanjing speech large model, with one sentence cloning human voice, human-like high naturalness and paralinguistic expression, multi-language and multi-dialect mixed generation capabilities, can realize single model multi-task speech generation. In order to better demonstrate the capabilities of Yuanjing large model, Director Zhu Changbo demonstrated the special functions and cases on site.

The breakthrough in the basic model capability provides China Unicom Yuanjing with a powerful model foundation engine. But this is far from enough. In order to provide customers with "easier to customize" model capabilities, the Yuanjing MaaS platform has upgraded its core components and general components. In his speech, Director Zhu Changbo focused on introducing the core capabilities of the Yuanjing MaaS platform-RAG and the Yuanjing MaaS platform-Intelligent Agent, which correspond to the ability of large models to use plug-in knowledge bases and plug-in tools, respectively. Both components have obtained the highest level certification of the China Academy of Information and Communications Technology.

Director Zhu Changbo introduced that the use of RAG in a large model is like the first time humans use search engines, with problems such as "not being able to search, not being able to see the whole, and not being able to find the right answer". In order to solve the problem of "not being able to search" due to the low matching degree between short questions and long text blocks, Yuanjing 2.0 innovatively uses the cascade segmentation method to cut out text blocks of different lengths, achieving a 5% increase in recall rate; in order to solve the problem of "not being able to see the whole answer" when tables, especially long tables, are cut in the middle when building a knowledge base, Yuanjing 2.0 creatively uses adaptive table splitting and integration, automatically completes information such as table headers and titles, and improves the accuracy of table questions and answers by 20 percentage points; for the problem of "not being able to find the right answer" for string queries such as license plate numbers and fault codes, Yuanjing 2.0 uses a multi-way retrieval fusion method to increase the answer accuracy by nearly 20 percentage points.

Yuanjing MaaS platform-Intelligent Agent is committed to solving the two difficult problems of letting the big model know "when to use tools and what tools to use". First, it clarifies the capability boundary of the big model, so that the big model knows when to answer questions by itself and when to use tools; second, it enhances the precise intention recognition capability of the big model, so that the big model can call the appropriate tools to answer questions; third, the Yuanjing MaaS platform not only provides more than 100 tools for users to choose from, but also supports users to DIY their own tools and customize personalized intelligent agents. Director Zhu Changbo said that China Unicom Yuanjing also sincerely hopes that ecological partners will connect their excellent plug-ins and tools to the Yuanjing MaaS platform, so that the capabilities of Yuanjing intelligent agents can be infinitely expanded.

In addition, Director Zhu Changbo also gave an in-depth introduction to the innovative practices of Yuanjing 2.0 in the field of AI security. First, it demonstrated China Unicom's understanding of AI's inherent security and took the lead in publishing the AI ​​inherent security industry white paper; second, it opened up the open source security test data set that covers the most risk types; third, it formed a leading domestic end-to-end model service security tool chain. Director Zhu Changbo said that relying on the improvement of the above security capabilities, Yuanjing's big model has passed the double filing of the Central Cyberspace Administration of China and obtained the highest level of security evaluation certification of the China Software Evaluation Center's big model; and completed the localization adaptation of Ascend, making the full stack capabilities more autonomous and controllable. Through these efforts, Yuanjing 2.0 has made a qualitative improvement in security capabilities.

"The key to big models is application," Director Zhu Changbo emphasized, "Only by truly going deep into the industry and the scenarios can the value of big models be reflected and a closed business loop be formed." Relying on the upgrade of basic models, MaaS platform and security capabilities, Yuanjing 2.0 is reshaping industrial applications and helping government affairs, enterprises, individuals and families and other thousands of industries to achieve "artificial intelligence +".

Finally, Director Zhu Changbo announced at the meeting that China Unicom Yuanjing has achieved remarkable results. At present, it has formed 35+ industry big models and 100+ benchmark applications, and has helped improve quality and efficiency in many scenarios. Hundreds of excellent cases implemented by China Unicom and its partners were also released on site. For example, all seats of Liaoning 12345 provincial platform have been fully connected to the government hotline big model, which has shortened the work order filling time by 80%, improved the completeness of work order records by 30%, and improved the accuracy of content recommendation by 35%. Many of the above 35 industry big models and 100 excellent cases have also been selected for important international and domestic awards, such as the medical big model selected as the ITU AI4G excellent case, the urban governance big model and the port big model selected as the excellent cases of the World Artificial Intelligence Conference, and the cultural and creative big model selected as the excellent case of the Global Digital Economy Conference. The rich application practice has won the reputation and distinctive characteristics of the Yuanjing big model as "a big model that understands the industry better, and an intelligent engine for industrial upgrading".

Looking to the future, Director Zhu Changbo said that the development of the artificial intelligence industry cannot be separated from the strong support and joint efforts of the vast number of partners. Standing at the new starting point of China Unicom's 30th anniversary, Unicom Yuanjing will work with the vast number of partners to move forward and build a new height of artificial intelligence.