JD.com launched a large model of Yanxi, taking the lead in laying out industrial applications
On July 13, the large model of JD Yanxi was launched at the 2023 JD Global Technology Explorer Conference and JD Cloud Summit. Compared with the general-purpose large-scale model, this large-scale model derived from the industry and service industry integrates 70% of general-purpose data and 30% of the original data of the digital intelligence supply chain, and has "higher industrial attributes, stronger generalization capabilities, and more security. With the advantage of "guarantee", it is committed to in-depth knowledge-intensive and task-based industrial scenarios such as retail, logistics, finance, health, and government affairs to solve real industrial problems.
In order to consolidate industrial applications, JD.com has launched a complete set of tools for large models, including the infrastructure supporting the development of large models - Yanxi AI development computing platform, vector database, hybrid multi-cloud operating system cloud ship, high-performance storage platform Yunhai, software Hardware integrated virtualization engine Jinggang and other core products.
Relying on JD Yanxi's large-scale model and JD's advantages in the industry and supply chain, JD.com has also released two major service platforms, including the newly upgraded Uplus DaaS and Yanxi intelligent service platform, as well as retail, financial, urban, and health services. , and logistics industry solutions in the five major areas.
Xu Ran, CEO of Jingdong Group, said: "Jingdong's large-scale model technology evolution follows Jingdong's technological pursuit: cost, efficiency, experience, credibility, inclusiveness, and breakthrough. Cost, efficiency and experience are inherited from Jingdong's business philosophy. , is also the first principle of retail. Credibility, inclusiveness and breakthrough are the promises of technology to serve the industry and society.”
JD.com believes that JD.com’s connotation of technology pursuit, in the field of AI technology represented by large models, can enable industry partners to use the achievements of technological innovation with more confidence and low threshold, help technology go out of the walls of laboratories and companies, and truly promote the industry. The improvement of production efficiency will play greater value in industrial scenarios.
Big model that understands the industry best
Create value with industry data and Know-how
"The real realization of the value of the large model must be in the industrial application." Xu Ran said that the value of the large model = algorithm × computing power × data × industrial thickness² "The first three indicators are important, but the technology is implemented in the industrial scene Application and creation of actual value are the key. When the industrial efficiency and the expansion of industrial boundaries are qualitatively improved, the large model will have more important practical value and significance, which will be no less than another industrial revolution.”
"Cutting into the large-scale model from the industrial side is like climbing the technical Mount Everest from the north slope. Although the road is more difficult, there are more magnificent scenery." Xu Ran said.
The industrial scene is an area that JD.com is very good at. As a new type of entity enterprise, JD.com’s own development comes from experience in industrial scenarios, and it has a rich industrial layout including retail, logistics, technology, health, industry, production and development. This has accumulated tens of billions of high-quality interactive data for JD.com every year, and the database for training Yanxi's large model consists of 70% general data and 30% supply chain original data.
The profound industrial data accumulation and the Know-How of thousands of industries have become the core advantages of JD.com's layout of large models. The trained large models naturally have "industrial genes". In addition, the Yanxi large model also has the ability of industry generalization, and can guarantee the safe use of customers on and off the cloud.
"The digital intelligence supply chain trains the big model of the industry, and the big model relies on the supply chain to go deep into the industry," said Cao Peng, chairman of the technical committee of JD Group and president of JD Cloud Business Unit.
For the application practice of large models, JD.com also has a clear "three-step" plan. At present, JD Cloud has built a general-purpose large model based on internal practices; by the end of this year, JD.com will use high-complexity scenarios on a large scale to iteratively produce solid industrial services; it is expected that in early 2024, the large-scale model capabilities will be externalized Serious business scenes are open. At present, Jingdong has reached the second step and has achieved rich practical results internally.
"In a sense, our big model is equivalent to taking general education and four years of professional undergraduate education," said Dr. He Xiaodong, Dean of JD Discovery Research Institute and President of JD Technology's Intelligent Service and Product Department. The difficulty of large-scale models is not technological catch-up, but industrial breakthroughs. The industrial practice of Jingdong Yanxi large model is a part that Jingdong pays great attention to.
Build a large model "training camp" and "ammunition depot" to get through every step of technology implementation
Jingdong has been laying out large-scale models since a long time ago. In 2021, JD.com will launch the billion-level model K-PLUG, and in 2022 it will launch the tens-billion-level model Vega. This conference will showcase the 100-billion-level model.
It was the early layout that allowed JD.com to form a complete tool chain of a large model: Yanxi AI development computing platform, vector database, and the newly upgraded infrastructure Yunjian, Yunhai and Jinggang.
JD.com's Yanxi AI development computing platform can provide customized solutions for customers' large-scale model development and industry application development. It not only has an industry knowledge base, which has accumulated JD’s own know-how in retail, logistics, health, finance and other industries, but also has more than 100 training and reasoning optimization tools, which can quickly convert general models into professional models.
At the conference site, JD.com demonstrated the operation of transforming a general-purpose model into a health industry model. Usually, it takes more than 10 scientists a week for customers to complete this process, from data preparation, model training to model deployment. However, using Yanxi AI to develop a computing platform only requires 1-2 algorithm personnel, and it can be completed in a few minutes. Through the platform model to accelerate the optimization of tools, it can also save 90% of the inference cost.
The Jingyi Qianxun large model released by JD Health is based on the general large model of JD Yanxi, which can quickly complete the migration and learning of various scenarios in the medical and health field, so as to realize the comprehensive AI-based deployment of products and solutions. It can provide a solid technical foundation for telemedicine services.
In addition, JD.com also showcased the AI marketing operation platform in the financial field. Through simple dialogues, users can generate marketing activities in one stop, covering operational strategy formulation, marketing personnel scheduling and arrangement, building activity pages, batch generation of marketing copywriting and materials, digital delivery, etc. In the past, this set of processes required five types of functional personnel: product, R&D, algorithm, design, and analyst, but now it has been reduced to one person; in the past, the process required 2,000 human-computer interactions, but it has also been reduced to less than 50. With the support of large-scale model capabilities, the production efficiency of the marketing plan has increased a hundredfold.
In the field of e-commerce, JD Cloud's AIGC content marketing platform is also promising. Starting from a product map, it can understand the characteristics of the product, and quickly win the main product map, marketing poster map, and store detail map needed for e-commerce operations to meet the needs of merchants for quick store opening and marketing. With the blessing of artificial intelligence, the production cost of each set of pictures can be reduced by 90%, and the production cycle is also shortened from 7 days to half a day.
Taking the wealth scene as an example, the Yanxi Youfang module has been launched in JD.com’s institutional fund agency sales system to provide institutional customers with more convenient fund screening and investment research and analysis functions.
If the Yanxi AI development and computing platform is a "training camp" for large models, the vector database is the "ammunition library" for large models and a bridge for the data used by large models. JD.com has been developing vector databases since 2019. After being honed in e-commerce promotion scenarios, the vector database Vearch has been able to support tens of billions of high-performance retrievals, with latency reduced to milliseconds and availability of 99.99%. It has served more than 100 large and medium-sized enterprises. Business users. JD.com uses the vector database for large model pre-training, reducing the cost of reasoning by 80%.
JD.com has also made full preparations for the computing resources required for the large model. In Chongqing Sun Island, JD.com’s first nationwide supercomputing center, the Tianqin α cluster, is running day and night. The inference speed has increased by 6.2 times compared with the past, and the inference cost has been reduced by 90%, which has become a powerful computing power behind technological innovation.
The three core products of JD Cloud also ushered in an upgrade this time: the newly upgraded hybrid multi-cloud operating system Yunjian, which reduces the resource cost of GPU/NPU by 50%, and provides the highest business continuity commitment in the industry. The high-performance storage platform Yunhai improves the reading and writing performance per second of a single cluster by 5 times, reaching tens of millions; the latency is reduced by 50%, reaching the level of 100 microseconds; and the cost of use is reduced by 30%. Jinggang, a virtualization engine integrating software and hardware, offloads virtualization to the smart network card, reduces virtualization loss to 0, improves network performance by up to 4 times, and supports 100Gb VPC network, and the performance of a single cloud hard disk instance exceeds one million IOPS.
Upgrading multi-scenario solutions Technological innovations benefit the industry
With technological progress, industrial applications have evolved from process-driven and data-driven to knowledge-driven. JD.com has also upgraded two platforms that can accumulate high-quality industry data: the brand-new Uplus DaaS and Yanxi Intelligent Services to support the new data needs of industrial applications.
The newly upgraded Uplus DaaS has three advantages: more dimensions, including not only the enterprise's own data, but also the safe intercommunication and integration of enterprise, JD.com and ecological data; easier to use, greatly improving the interactive experience, lowering the threshold of use, and using A simple description can generate an answer; it is simpler, and the atomic capabilities of the data development platform, data service platform, and data visualization platform are connected in series to simplify the operation between the original complex systems.
After the Yanxi intelligent service upgrade, it also has richer interactive channels, covering from intelligent customer service to live broadcast room, from mobile terminal to offline intelligent terminal; more vivid interactive effects, supporting voice, video, text and other methods; more professional The interactive content generated in the professional field is more controllable and usable. At present, it has served serious scenarios such as the 12345 hotline, bank digital employees, and expert diagnosis scenarios.
To realize the industrial value of technology, the supply chain is an important starting point. JD Cloud has gone deep into more than 20 digital scenarios such as omni-channel one-pack goods, supply chain financial technology, industrial collaboration, data middle platform, and digital warehouse, and has served 100 cities, 2,500+ large enterprises, 944 financial institutions and over 2.5 million Small and medium enterprises. The digital practice of these thousands of industries has created the cloud that understands the industry best.
"Intelligence takes place first in digitally-leading industries," said Gao Liqiang, vice president of JD Group and president of JD Technology Products and Solutions. In multiple scenarios such as retail, finance, cities, and logistics, JD.com has released brand-new intelligent solutions to upgrade service capabilities with the help of large-scale model capabilities.
At this conference, JD Logistics released Jinghui 3.0, a one-stop digital intelligent supply chain data management platform. With the blessing of large models, its sales forecast, inventory, supply and replenishment plan are more expressive. The interactive supply chain control tower can help users quickly locate and solve supply chain problems, reducing costs and increasing efficiency.
In the past, the development of artificial intelligence has reached the verge of application explosion several times, but many of them have not created much value. One of the important reasons is that the technology has not formed a solid application in the industry. As a new entity enterprise, JD.com has always been rooted in and serving the real economy. In the process of industrial application of large models, it has drawn a path of industrial value creation with solid support by relying on its solid industrial accumulation.
At the conference site, JD.com announced to join hands with retail and financial industry partners to explore supply chain optimization, customer experience, marketing cost reduction and efficiency enhancement in retail scenarios, and explore financial customer service, robo-advisors, and product marketing in financial scenarios to jointly build Innovative applications of large models in the retail and financial industries.
As Xu Ran said: "The journey of JD.com to develop large models is destined to go hand in hand with industry partners. JD.com will work closely with industry partners and continue to contribute to the high-quality development of the industry through technological innovation." $JD.com (USJD)$