This year marks a significant turning point for the large model industry, with increasing client expectations and demands for solutions across diverse business scenarios. In this context, intelligent agents are experiencing explosive growth—as more clients seek to develop applications that evolve towards the next generation of agents. To put it simply, if large models are akin to the brain, intelligent agents act as the limbs that perform tasks. They can dissect complex client needs, activate workflows, and serve as genuine business assistants. Their accessibility allows a broader range of individuals to engage with them. Currently, most AI-native applications can be developed using intelligent agents, and all major large model companies and ecosystem partners are focusing on these agents, indicating that the industry has entered a pivotal phase centered around "mandatory intelligent agents."
At the JD Cloud Summit on July 30, 2024, held in Shanghai, JD officially launched the Yanxi Intelligent Agent Platform, a one-stop development environment for intelligent agents. JD Cloud identifies intelligent agents, digital personas, and embodied intelligence as the primary mediums for interaction between large models and end users, positioning intelligent agents as vital drivers for enterprise AI-native applications. The Yanxi platform includes a full-stack suite of products designed to accelerate the deployment of large models across various scenarios. "General large models rely on computational power, while enterprise models emerge from actual business pursuits," noted Cao Peng, Chair of JD's Technology Committee and President of JD Cloud Business Division. He emphasized that to fully leverage the potential of large models within industries, a robust product and tool ecosystem is essential.
Insights reveal that the full-stack products for large models were incubated within JD’s extensive supply chain network. Internally, over 100 applications of large models are active, supporting more than 600,000 employees and 200,000 merchants. JD representatives expressed surprise at the frontline enthusiasm for intelligent agents, identifying them as one of the most effective platforms for deploying large models this year. The Yanxi platform, developed since October last year and opened to select ecosystem partners this spring, has already seen employees create over 3,300 intelligent agents in just a few months—exceeding expectations. The platform has rapidly integrated over a thousand workflows and knowledge bases, proving incredibly popular, likely due to the extensive number of business personnel and the interconnected nature of operations.
Interestingly, JD has invested heavily in AI algorithm teams for its core retail supply chain over the past few years. However, the transformation brought about by large models has led to disruptive changes for groups previously unaffected by AI. Frontline staff, functional teams, and product managers are now developing agents tailored to their workflows. For instance, a long video editing agent was created to streamline the process of editing training videos, significantly reducing the time needed from hours to just minutes. Another intelligent assistant aids delivery personnel in route planning and operational prompts. The JD Intelligent Agent Market features numerous active agents, including one for telemarketing quality checks, which has replaced manual reviews with automated processes, achieving millions of interactions.
The Yanxi platform supports no-code development, enabling staff without algorithm backgrounds to build intelligent agents through a user-friendly interface. This approach facilitates rapid realization of AI ideas and fosters collaborative organizational change. Where coordination previously required meetings and planning, departments can now spontaneously register their tools and APIs on the intelligent agent platform, linking foundational capabilities for more seamless cooperation.
Having refined internally, the Yanxi platform is now available to external users. It offers both public cloud and privatized deployment options, and unlike many platforms aimed at individual developers, the Yanxi platform targets industry-specific needs. Its development garnered extensive feedback from JD’s retail, healthcare, and logistics sectors, leading to a wealth of industry solutions. Preconfigured templates and plugins allow clients to create customized intelligent agents in under a minute. To optimize usage, JD representatives recommend focusing on two key areas: continually discovering high-demand applications to establish benchmarks, and enhancing platform operations to support thousands of user applications by categorizing intelligent agents for easier access.
As businesses harness the capabilities of intelligent agents, the AI-native applications emerging from them are reshaping the enterprise software market. Some agents have entirely replaced existing SaaS software, while others are being integrated into SaaS systems. The Yanxi platform also simplifies application deployment via web and API integrations, facilitating internal communication channels like corporate WeChat and collaborative office tools. Although intelligent agents are still in their infancy, their ongoing development hinges on the creative input of enterprise users, driving the evolution of AI agents.
Behind the Intelligent Agent: The Large Model Development Pipeline
The Yanxi Intelligent Agent Platform currently integrates with dozens of large models, powered by the Yanxi AI Development and Computation Platform, which has recently undergone a comprehensive upgrade to version 2.0. This platform emphasizes key capabilities vital for the deployment of large models. The first capability is model compression and scaling, allowing enterprises to tailor large models based on client needs. "Even with significant investment, general large models often need fine-tuning for specific scenarios," explained Cao Peng, underlining the necessity for speedy response and cost-effective inference.
The second key capability addresses data preparation, requiring robust toolchains to handle the vast amounts of multimodal data unique to each industry. Incomplete process data represents a significant barrier to effective large model deployment. As a JD health representative pointed out, understanding expert reasoning is essential to overcoming challenges like hallucination issues. Current strategies involve leveraging Retrieval-Augmented Generation (RAG) technologies for automated inference paths, complemented by expert insights to enhance data quality.
The third crucial capability involves model evaluation, encompassing both general and domain-specific assessments. JD has established benchmarks for evaluating general large models, while also focusing on specific sectors like healthcare and retail to ensure comprehensive capabilities. Automated and manual evaluation methods are employed to maintain measurement accuracy in a continuously developing industry.
Accelerating Large Model Deployment through Enhanced Infrastructure
As large models and intelligent agents gain traction in various industries, the foundational infrastructure—comprising computing, networking, and storage—must evolve to meet new challenges. One key aspect is creating an open infrastructure platform that supports multi-cloud, multi-core, and multi-active frameworks to accommodate diverse models and complex business requirements. Current demand for multimodal large models necessitates tenfold or even hundredfold increases in computing capacity. Companies globally face similar challenges, leveraging heterogeneous computing to enhance model training efficiency and yield a better cost-performance ratio.
Additionally, storage requirements for model training peak during specific intervals, requiring the capacity to handle terabytes of data in mere seconds— a stark contrast to traditional methods that allow these tasks to be spaced out over months. Enhanced storage systems capable of delivering higher throughput, input/output operations per second (IOPS), bandwidth, and lower latency are essential for optimizing model performance. As Cao Peng noted, the efficiency of storage systems can influence the duration of training periods significantly.
At the JD Cloud Summit, the advancements in JD Cloud’s AI infrastructure were highlighted, particularly the next-generation storage product, Yunhai, which has been refined through over a decade of handling JD’s complex scenarios, making it adept at supporting large datasets for model training. The newly launched version 3.0 of Yunhai promises enhanced performance attributes crucial for large model deployment.
Moreover, JD has established the Cloud Ship AI Computing Cloud Platform for efficient resource management of diverse CPU, GPU, and domestically produced AI acceleration chips. This platform facilitates unified scheduling of distributed computing resources across regions, ensuring cost-effectiveness. The innovative vGPU pooling solution enhances AI computation efficiency, ultimately alleviating common concerns about computational resource availability.
In response to the pricing competition within the large model sector, Cao Peng highlighted JD Cloud’s commitment to leveraging technology upgrades to reduce costs, positioning itself competitively against mainstream providers. The JD Cloud framework aims to integrate infrastructures, model services, and intelligent agent applications, fostering the deployment of large models across industries and transforming business outcomes.
Ranking
