JD's hidden move has been dormant for five years

Miao Zhengqing

Miao Zhengqing

Tiger Sniff Official Team

follow with interest

Product | Tiger Sniff Commercial Consumer Group

Author | Miao Zhengqing

Title Image | Visual China

Article Summary

JD.com is focusing on the B-end market in the AI field, launching products such as Yanxi Big Model, Digital Human, and Intelligent Agent Platform, and integrating DeepSeek open-source models to reduce costs. In the face of industry competition, JD.com optimizes its models through technologies such as reinforcement learning and synthetic data, solves the illusion problem, and lays out a three-step strategy (language big model → multimodal → embodied intelligence) to explore the commercialization path of AI in customer service, marketing, and other scenarios.

•    Technological breakthrough: DeepSeek's deep reasoning and open-source model reshape the industry landscape, driving AI into the L2 stage

•    Strategic Focus: JD focuses on deploying B-end AI tools, covering 800000 merchants, and focusing on customer service, marketing, and digital human scenarios

•    Open Source Impact: DeepSeek breaks down closed source barriers and forces companies to explore differentiated business models

•    Three step path: language model → multimodal → embodied intelligence, JD anchors the evolution direction of AGI technology

•    Application breakthrough: JD focuses on solving the DeepSeek illusion problem and improving the reliability of serious applications in B-end scenarios

•    Future prediction: Synthetic data and reinforcement learning become key, AI will deeply integrate physical world task interaction

Faced with the impact of DeepSeek, Internet giants are readjusting their AI strategies.

In early March, Tencent Yuanbao fully embraced DeepSeek and leveraged DeepSeek to achieve a comeback in popularity. Almost simultaneously, ByteDance's Volcano Engine and Feishu, as well as Alibaba's International Station, Alibaba Cloud, and DingTalk, announced their integration into DeepSeek.

DeepSeek, a major Internet company, has become increasingly fierce, and behind this is the key game of each major company against AI.

In this wave of DeepSeek integration frenzy, JD.com is also among them. In early February, JD Cloud officially launched the DeepSeek-R1 and DeepSeek-V3 models, and took the lead in launching the DeepSeek all-in-one machine in the industry.

Unlike Tencent and ByteDance's efforts in the AI to C market through Yuanbao and Doubao, in 2024, JD.com will focus on developing B-end products such as the Yanxi big model, digital humans, intelligent agent platforms, and intelligent coding assistant JoyCoder on the AI side. As of the end of 2024, 800000 merchants on the JD platform have already used JD AI tools.

However, facing JD.com is also a more intense 2025: with ByteDance and Alibaba investing in the AI to B field, as well as AI unicorn companies such as Baichuan and Zhipu further shifting towards the B-end market, competition around AI to B will further intensify.

What underlying logic has DeepSeek changed in the AI community? What are the key opportunities in the B market? Is JD.com ready?

After JD Cloud announced its integration with DeepSeek, Tiger Sniff immediately communicated with He Xiaodong, President of JD Technology's Artificial Intelligence Business Unit and Dean of JD Exploration Research Institute, asking him to share a series of thoughts on the impact of DeepSeek, competition in the AItoB industry, and trends in AI technology. It is worth noting that the Chinese Association of Artificial Intelligence recently released the announcement of "Wu Wenjun Artificial Intelligence Science and Technology Award" in 2024, and the artificial intelligence team of JD Science and Technology won the special prize of Wu Wenjun Artificial Intelligence Science and Technology Award, the highest prize of China's intelligent science and technology, by virtue of the project of "key technologies and industrial applications of multimodal interactive digital people", which is also the only special prize of this year.

Attached is a transcript of the communication, with some deletions and modifications:

DeepSeek reshapes the business landscape

Tiger Sniff: What was the biggest impact of DeepSeek's wave of enthusiasm at the beginning of the year on you?

He Xiaodong: I think there are two interesting points, DeepSeek's deep reasoning technology and its open-source R1 model. These two points have profoundly reshaped the landscape of the AI industry.

Technically speaking, OpenAI has divided its API into five levels, namely L1 to L5. The first level focuses on understanding human-computer dialogue language, while the second level involves deep reasoning. If we look at it according to this classification, OpenAI's O1 and DeepSeek's R1 mean that we have officially moved from L1 to L2 stage.

In fact, by the end of 2023, there were rumors in the Silicon Valley tech community that OpenAI was developing a reinforcement learning algorithm project called Q-Star in preparation for GPT-5. At that time, we only heard the name, but as soon as we heard the name, we could think that Q-ability is the classic reinforcement learning. We associate it with a method similar to AlphaGo, which combines reinforcement learning with search for deep reasoning. At that time, we thought that Go was ultimately a fixed or closed domain, and if we could achieve open domain, that is, open intelligence on large models, there would actually be many technical challenges.

But it can be seen that by the end of 2023, everyone has already seen this technological direction, and it will start to be developed in July 2024. In April 2024, when I attended an AI forum at Tsinghua University, I expressed a viewpoint that there may be two directions in the future. One is deep reinforcement learning, which uses models to train models, because relying solely on human training models is already approaching the top of GPT 4. The second direction is "synthetic data", because ordinary data is almost used up, and the next step may require more efficient synthetic data to do this.

It can be said that technological progress is indeed quite fast. Although many people have seen this technological direction, not many have actually achieved it. But for the future, everyone should have more confidence, especially after the open source of models and papers.

Another impact or influence is that open source itself has reshaped the business landscape. For example, OpenAI was the first to be impacted, and its business model was greatly affected. But ultimately, open source is a very good thing. It reduces the cost of models, opens up space for applications, and allows companies of all sizes to use AI capabilities to build applications at a lower cost. This will bring about a diverse ecosystem.

JD.com is also considering how to take advantage of this opportunity, including quickly uploading DeepSeek to JD Cloud to provide this capability to our large and small customers. We also offer the ability to privately deploy DeepSeek, including industry-specific solutions.

Tiger Sniff: As you mentioned earlier, everyone has been discussing or seeing some trends since the end of 2023, but why is there only one company that has emerged?

He Xiaodong: In the first half of 2024, many people still focus on the "big" aspect, such as pursuing larger models, more cards, and achieving trillion or 10 trillion scale. People only see the so-called size scale and do not truly realize the importance of reasoning.

In the second half of 2024, everyone will see the importance of reasoning. DeepSeek is a very pure technology company, and based on this purity, they can better see the depth of the technology itself, so they will invest more resolutely in reasoning and other aspects.

Tiger Sniff: Another hot topic sparked by DeepSeek is open source. It seems that a year ago, even in the fourth quarter of 2024, the big modeling circle's judgment on open source and closed source is not as strong as it is today. Why is DeepSeek resolutely taking this path?

He Xiaodong: Going back to a year ago, only OpenAI's technology was leading and could stand out from others. It was not open source, so people would think that the significance of open source was not that great. Actually, over a year ago, Meta's Llama model was also open sourced, but its performance was indeed slightly inferior to OpenAI. And, people may think that closed source may be a viable business model.

Why has DeepSeek had such a significant impact on open source? I think it's because it challenged OpenAI's moat with a high-quality model, and people will think that the open source ecosystem can be built. So often, the influence of key companies and individuals can change people's fundamental views on certain issues.

Tiger Sniff: What impact does the open source imagination space have on what JD.com is doing?

He Xiaodong: The AI technology competition will continue, which means further investment is needed. We can't stop and we won't stop. From JD's practice of expanding models, we have seen many opportunities in the application layer, especially the impact of DeepSeek on the surrounding ecosystem. A large number of customers have realized that a high-quality DeepSeek model can be deployed at a relatively low cost. So you can see that JD Cloud and other clouds are all being deployed. As a result, it also brought some opportunities.

In terms of technology and application, everyone will have greater actions. Technically, we will absorb the techniques from these latest papers, such as further strengthening training and utilizing techniques including distillation. The top models on the market are currently trying this comprehensive distillation to improve efficiency.

In terms of application, we are embracing DeepSeek itself and quickly integrating it into our entire AI product line, providing end-to-end services directly to customers. For example, we have been trying to use it to write live streaming scripts and marketing copy. But at the same time, some problems were also discovered, such as high hallucinations. In serious usage scenarios, there are still many areas that need to be fine tuned, such as using DeepSeek to generate equity copy. Sometimes, there may be some non-existent equity, which is unacceptable for merchants. For the entire B-end, the biggest challenge currently faced by the open-source DeepSeek in adaptation is the illusion, and we are trying various ways to reduce it.

The AI to B market is not just about technology

Tiger Sniff: What do you think is the competition situation in the AI to B market this year?

He Xiaodong: Actually, both AI to C and AI to B are what we want to do. AI to B can actually be divided into two levels, one called Model Service and the other called Software Service, with the former being a more fundamental service. This Model Service includes models, API interfaces, as well as DeepSeek cloud deployment mentioned earlier, and the development of higher-level intelligent applications. We are currently providing such capabilities, and in the near future we will provide a new generation of capabilities, such as API calls.

Software Service， It is an end-to-end product, such as some deep closed-loop AI products, including intelligent customer service, intelligent marketing, and digital human live streaming.

Through these two types of AItoB products, we can form a commercial closed loop, and I think everyone has the opportunity.

Tiger Sniff: It seems that JD's approach to AI has not changed much from before?

He Xiaodong: Yes, JD's strategy is quite firm. DeepSeek open source, for JD.com, can make some of the commercial models we previously wanted to do more efficient or cost-effective. However, the open source of DeepSeek will not fundamentally change JD's approach.

Tiger Sniff: Do you choose more MaaS or SaaS?

He Xiaodong: Both sides are advancing in parallel. The opportunities for MaaS are indeed different after DeepSeek is open sourced. Before the open source era, we often built MaaS on top of the cloud and used our own or industry's open source models. However, at that time, the industry's open source models generally had poorer performance, so our own model fees were generally higher. But the significant impact brought by the open source of DeepSeek is that our own paid model needs to differentiate from the free open source DeepSeek in order to close the business model loop.

So now the situation has changed, one is to compete on who has lower deployment costs, including lower costs or higher parallel reasoning techniques; The second approach is to make further adjustments based on DeepSeek, or to do reinforcement learning or distillation, to change its illusion problem, that is, to cultivate in specific industries and help enterprises build their own large models, so that the value of DeepSeek can be better presented. Both of these models are currently available to MaaS and are part of JD's technological accumulation that can be commercialized.

In the SaaS part, the base model is only a part of it, and overall efficiency, scenarios, and end-to-end experience are all core issues. This will lead to differentiation at the application layer, which will bring about new business models.

We are currently focusing more on AI customer service, AI marketing, and digital people in this area. Taking digital humans as an example, they are no longer just a multimodal combination of language models and other large models. It is a comprehensive product level competition, and businesses value AI driven final sales results. Just like in the Internet era, when the Internet has become an infrastructure, people are not competing for network speed but for different applications. At that time, JD was able to win not only due to technological factors, but also a series of factors such as overall terminal experience, operational methodology, live streaming traffic, and gameplay.

JD launches AI in three steps

Tiger Sniff: Did you have specific goals when you were doing AI?

He Xiaodong: As early as around 2020, we proposed a plan, and now looking back, we are actually following this plan. We talked about achieving a 'one platform three-stage rocket' at that time. A platform means that we need a basic large-scale model platform, and at that time, we had already seen the scale effect value of the model, so this was a must do. On this platform, we have planned a three-stage rocket: from AI intelligent customer service, to AI marketing, and then intelligent interactive media, referring to products such as digital humans.

AI customer service is actually a more concrete product, and we have been doing it well since 2020. At that time, we also started doing AI marketing, mainly to upgrade marketing with AI. We also hope that AI can bring some disruptive innovations to marketing, such as using AI voice for marketing promotion, relationship maintenance, and even as assistants and guides. In this field, we further see the value of intelligent interaction, including extending to new human-machine interfaces such as digital humans today. These directions were planned by us at that time, and we are still following them today.

Intelligent customer service is actually a relatively mature track, and there is still huge space for AI marketing in the future. Look at several foreign giants, many of their AI commercialization includes marketing. Intelligent interaction includes AIGC, which is not just about digital humans, but also AI generated graphics, AI generated audio and video. These technologies can support the emergence of many new apps and formats, which means it can redo social, content, search, and e-commerce. This imaginative space is enormous.

In fact, it is very similar to the Internet in those days. When the Internet first appeared, people thought that selling gateways was the most profitable. Later, the early business model was just sending emails. But if you look back, you will find that these are just the first dishes. Later, so many giant companies and trillion level companies hardly existed in the earliest Internet era 20 or 30 years ago.

Tiger Sniff: What is the proportion of JD's investment in AI in recent years?

He Xiaodong: We have been steadfastly investing in AI. The inspiration DeepSeek gave us is that our entire plan has accelerated, which is actually a dynamic competition. DeepSeek has made the competition around AI even more intense, and of course, there are more opportunities.

Tiger Sniff: Have you ever been indecisive in 2023?

He Xiaodong: In 2023, I was conflicted about how to allocate resources in terms of investment scale, priority, and time and space.

Tiger Sniff: Is there any correlation between your investment in AI and your commercialization?

He Xiaodong: Our larger projects are invested through special projects. JD Exploration Research Institute has many AI directions dedicated to long-term research, and the group has long-term confidence in AI.

Tiger Sniff: When you were planning JD's AI strategy, did you refer to any models from domestic and foreign companies?

He Xiaodong: In the early stage, we mainly conducted research and judgment based on the development trend of technology, and accumulated and broken through core technologies through the trend of technology. This responsibility was undertaken by students from JD Exploration Research Institute and Basic Algorithms.

In the second stage, we aim to establish a capability platform so that we can implement and generate product applications. The third stage is to strengthen the ecology, and we need to develop the technology ecology, product ecology, and service ecology. From what we are doing now, we are in the third stage. But we also see that having APIs alone is not enough for a platform. You need benchmark, demonstrative, and flagship products. On the one hand, these products themselves can bring profits, and on the other hand, these products can show benchmark cases to ecological partners, which can attract everyone to work together on ecology.

Tiger Sniff: What are your core predictions about the future of AI at this moment?

He Xiaodong: Firstly, synthetic data should become a trend, similar to our synthetic materials, becoming the foundation of the future. What we use now is mainly popular data, which is hundreds of billions or trillions of data on the Internet. What we care about is the intelligence in the data rather than the data itself. So when training models, we actually want to extract the intelligence from these data. But the intelligence that mass data can provide is approaching its limit. If we want to further improve the intelligence level of the model, we cannot rely on the intelligence in mass data, but need data with higher intelligence density. That's also why many models now use math Olympiad questions for training, and then distill and perform reinforcement learning through other models.

The synthesis of data and reinforcement learning, as well as the confrontation between models, are the key to the development of models in the future. On a deeper level, it is necessary to further enhance the intelligence level of these models, allowing them to enter the real world and take on tasks, interact, and engage in confrontation in the real world, in order to continue improving their intelligence level.

Tiger Sniff: What preparations has JD Exploration Research Institute made in these areas?

He Xiaodong: JD Exploration Research Institute has a three-step layout - from language modeling to multimodal intelligence, and then to embodied intelligence. This is actually the process of the entire AI moving towards AGI (General Artificial Intelligence) or ASI (Super Artificial Intelligence). That is to say, after expanding from language and cognitive intelligence to multimodal intelligence, the next step must be physical world intelligence.

Title: JD's hidden move, dormant for five years

Article link: https://www.huxiu.com/article/4171750.html

Read the original article: JD's hidden move has been dormant for five years

Miao Zhengqing

Miao Zhengqing

Likes Wang Wei and Bai Juyi, admires Ji Zha

Tiger Sniff Team

242 articles have been published in Tiger Sniff

If you have any objections or complaints about this article, please contact tougao@huxiu.com

People who are changing and want to change the world are all on the Tiger Sniff app

京东这步暗棋，蛰伏了五年