JD.com handed in a great performance

Author | Liu Baodan

Editor | Zhou Zhiyu

AI large models are accelerating their landing in various industries. With its industrial advantages, JD.com is trying to explore its own path for the landing of large models.

Recently, at the 2024 World Artificial Intelligence Conference, JD.com showcased its layout in the field of large models, including the digital human live broadcast model, as well as the landing applications of large models in health, retail, and other scenarios, summarizing the phased achievements of the Yanxi large model launched a year ago.

First is the digital human live broadcast, which is the most representative application of JD.com's large model.

During this year's 618 shopping festival, JD.com began to use digital human live broadcasts on a large scale, achieving industrial applications at scale. Currently, Yanxi digital humans have started live broadcasts in over 5,000 brand rooms, with a total GMV of nearly 10 billion.

Internally at JD.com, digital human live broadcasts are defined as a disruptive interactive innovation. He Xiaodong, Director of JD.com Research Institute and Head of JD.com Technology Artificial Intelligence Business, stated, "During JD.com's 618 festival, the CEO digital human has gradually been accepted by a large number of users, combining disruptive technology with product features that possess disruptive characteristics."

In addition, JD.com's large models have also been deeply applied in content production, marketing, and customer acquisition.

In terms of content production, over 90,000 JD.com merchants use large models to produce store marketing materials at zero cost, generating marketing materials such as product detail images in seconds, achieving significant cost reduction and efficiency improvement.

In terms of marketing and customer acquisition, merchants can use AI outbound calls for user interaction. Compared to regular voice outbound calls, the conversion rate has increased by over 70%, and GMV has increased by over 290%. In addition, video outbound calls can also combine Yanxi digital human short video technology to automatically generate video content, improving the efficiency of content production and user reach.

Another important application of JD.com's large models is JD Health. According to Wang Guoxin, Head of JD Health's Intelligent Algorithm Department, over the past year, JD.com's core work on the model side includes teaching the model what it doesn't know, integrating the model effectively with offline digital systems, and extensively applying personal data in addition to utilizing its world knowledge.

JD Health has three main landing practices: intelligent services before diagnosis, creating a doctor copilot based on technology to reduce the workload and misdiagnosis rate of doctors, and providing personal digital assistants to assist patients in the treatment process by reminding them to take medication.

Currently, JD Health has built intelligent consultation, intelligent diagnosis, intelligent Q&A, digital therapy, and other products based on the model, and provides solutions for patients, doctors, hospitals, and enterprises.

JD.com has already extensively used the Yanxi large model in scenarios such as digital human live broadcasts, e-commerce, and health. However, overall, JD.com's exploration of large model applications is still in its early stages.

JD.com holds high expectations for the application prospects of large models. According to He Xiaodong, every technological era needs truly disruptive products to define it, such as smartphones in the mobile era, where various computations are associated with PCs "In the AI era, we hope to see truly disruptive applications and products that define this era."

JD.com is working towards this direction, with the company's advantage lying not only in its technical expertise on the model side, but also in its scene advantages and industrial chain data.

He Xiaodong stated that JD.com's large-scale model layout revolves around building its own basic capabilities such as foundational technology and cloud platforms; and delving into applications such as customer service, marketing, digital humans, and other series of products, hoping to explore high-value disruptive applications through this.

JD.com believes that the disruptive killer applications of large-scale models cannot arise from the optimization of general large-scale models, but must be a combination of technological progress and industrial accumulation. Essentially, JD.com's large-scale model possesses industrial genes, which sets it apart from other models.

He Xiaodong mentioned that besides 70% of the data being public data, JD.com also has 30% of training base data coming from the industry, which has been accumulated over the years in the retail, logistics, finance, health, and other sectors during JD.com's business operations. "With this data, our large-scale model can have a deeper understanding in the industry and truly meet industry demands."

However, JD.com also faces many challenges, as almost all major tech companies are heavily investing in the large-scale model track, making this round of technological competition unprecedentedly fierce. Moreover, technology is still in a rapid iterative process, and there are many uncertainties regarding future commercial monetization.

Artificial intelligence is considered the core of the next generation information industry after the Internet and cloud computing, and will largely determine the future landscape of internet applications. It is certain that JD.com has already taken big steps towards the AI era