To help brand growth, should we bet on big models or small ones?

Original link:

Large models with hundreds of billions of parameters have profoundly changed the consumption habits of ordinary people – AI customer service gradually replaces manual labor, providing more professional answers to various product inquiries, and is online 24 hours a day, tirelessly; people search in the search box of e-commerce platforms Enter product keywords, and what will pop up is no longer a link to the advertising bidding ranking, but a product specially customized for you based on your personal consumption habits, making consumption more rational.

But this is far from enough. All walks of life hope to use large models to bring more and more profound changes to themselves.

At the end of 2022, several members of the JD Group’s technical committee held a meeting, and they reached a consensus that large models are important enough. The next core question is: What value can JD use large models to create for the industry?

Everyone held their own opinions. Some insisted on making large models “broad”, while others demanded that large models be made “fine” in vertical industries. The final conclusion of the meeting was that’s large models and small models must be put into operation. Ultimately, The purpose is to “put it on the ground”.

Different from ChatGPT, Baidu’s Wenxinyiyan, Alibaba’s Tongyi Qianwen, iFlytek’s Spark Model and other general large models that focus on the basic layer, it is also different from large industrial models that focus on a certain vertical field. Cloud has chosen a relatively unique third path: making the large model “small”, using 70% general data and 30% supply chain native data, and making it more oriented towards intelligent application scenarios.

“If the big model is a game console, no matter how powerful the function is, it will definitely not work without games.” Chen Feng, head of the digital marketing product department of JD Technology Solution Center, told LatePost that what JD Cloud is doing is to ” “Game” is done well, the big model is made “small”, and implementation scenarios that have practical effects on the industry are found to solve specific “small” problems.

Large models bring more new possibilities for brand growth. With the help of Yanxi large model, JD Cloud can help brands use AIGC to quickly produce product titles and detailed pictures, which takes less than one-tenth of the time before; it can also help merchants design virtual live broadcast digital people with exclusive images and generate live broadcasts with one click Words. At present, JD Cloud Yanxi’s multi-modal digital people have settled in more than 4,000 brand live broadcast rooms, driving more than 800 million yuan in GMV; brands can also use the intelligent outbound calls provided by JD Cloud to customize celebrity voices to call users.

As brand management becomes more and more intelligent, live sales, interactive data, customer service interaction data, advertising traffic data, etc. will become data assets for merchants. Exploring this asset well will bring more business growth to the brand. This is the greatest significance of JD Cloud’s implementation of large models into business scenarios today.

Large model + DaaS goes deep into retail application scenarios

“In the past, decision-making in the retail industry often relied on experience, but that is no longer possible now,” said Shang Zhihu, general manager of Yili Group’s Digital Technology Center.

Businesses are increasingly scrutinizing various budgets, and overall advertising budget growth is gradually slowing down. According to QuestMobile, China’s Internet advertising market grew by 54.6% year-on-year in the first quarter of 2021, but in the first quarter of this year, the year-on-year growth rate was only 2.3%.

Limited budgets must be spent to achieve the same or even better results, and the DaaS industry has emerged to help quantitative analysis. The core of DaaS is to use data to solve growth problems, allowing enterprises to have data to apply, rely on data to think, and use data to make decisions.

At present, domestic DaaS products are launched by large Internet companies such as Alibaba and JD Cloud Retail Full Scenario Solutions, as well as vertical manufacturers such as Datablau, Shushu Technology, and Shence Data.

The data sources of large Internet companies are their own retail platforms and merchants. The advantage is that they have rich retail data accumulation and can provide merchants with marketing and growth suggestions that are more in line with the platform itself. built its own logistics 16 years ago and established it six years ago. By opening its capabilities to the whole society, it can also provide more help to businesses in the supply chain dimension.

However, the problems faced by large platforms are also obvious. They have too much and too complicated data. Blindly building a data platform will only waste this data.

The logic of JD Cloud is based on actual applications, sorting out and constructing massive data, and mining the real value of the data.

After being connected to the Yanxi large model, JD Cloud’s Yunding DaaS can more efficiently help merchants refine data into “knowledge” and generate effective information. It is like a sieve that filters out the value of data and truly solves the problems faced by the retail side in specific scenarios.

JD Cloud has launched DaaS products for brand growth since the beginning of this year, and has fully opened up the carrier of JD’s business, data and algorithm capabilities on a safe basis. This is also consistent with JD Group’s “open” strategy: JD Retail’s 2023 The four must-win battles are sinking the market, supply chain middle platform construction, open ecological construction and intra-city business.

Mu Xiaohai, vice president of JD Technology, said that due to data security requirements, although JD Cloud cannot directly provide merchants with data such as users’ purchase history, browsing records and search records on the platform, it can provide industry trends, etc. after privacy calculations. data.

For example, set up a picture scoring model to score the categories sold by this merchant based on the color and composition of the product pictures, and give merchants modification suggestions. It’s not that the more exquisite the picture, the higher the score. If the merchant sells low-priced products for the sinking market, then the model will give a more down-to-earth picture. If you use the traditional manual labeling method, it will be difficult to give detailed suggestions for each merchant.

After using the Yanxi large model, merchants can complete full-cycle management from training models to application service deployment. In the past, a team of more than 10 scientists was required to work, but now only 1-2 algorithmic personnel are needed. The entire process from data preparation, model training to deployment can be completed through the platform. The training efficiency is increased by 2 times, the inference efficiency is improved by 6.2 times, and the cost Save nearly 90%.

On the other hand, merchants’ data in each industry can also feed back large models and output standardized products to serve more merchants.

For example, JD Cloud once helped Wal-Mart generate product posters. Wal-Mart input prompt words to decide what kind of light and what kind of color combination to use. After learning the large model, it can be output to the entire fresh food industry. “Customers are always our best teachers.” Mu Xiaohai said.

Judging from Google’s current practices, the marketing and growth logic of future merchants may change dramatically due to large-scale models: Google has directly inserted advertisements into its conversational AI products; and generates advertising copy based on keywords searched by users. For example, if you search for skin-soothing products, the product copy recommended to you will highlight its skin-soothing effect; in addition, AI can also help merchants generate landing pages for product advertisements – from when users see the product link to clicking on the link. , are ads tailored to their individual needs.

For merchants, large models will completely change the core links of product selection, marketing, customer service, and logistics in the future. When the core links are all intelligent, the demand for more valuable data analysis will be stronger.

Meeting these needs relies on more professional industry knowledge. Just having data and analysis is not enough. The JD Cloud DaaS data platform supported by the Yanxi large model and JD scenes can help the retail industry realize the rapid implementation of intelligent applications. Brands can quickly and easily adjust high-quality data in multiple scenarios to improve marketing efficiency, service efficiency, and supply chain efficiency. .

Make large models smaller to improve efficiency in core e-commerce links

Previously, large Internet companies built data analysis platforms, often in the form of a middle platform. After collecting the needs of multiple businesses, they output relatively standardized and unified solutions. The advantage is that it can save the technical manpower of the entire group, but the disadvantage is that it is difficult to Meet the targeted data analysis needs of each business.

JD Cloud’s idea is to no longer “be greedy for the big and seek perfection”, but to delve into specific business scenarios and provide targeted solutions to merchants’ retail needs to help brand growth.

For merchants, the core appeal is to improve marketing efficiency and bring growth to performance, so this link has become an important focus of JD Cloud’s large model + DaaS.

When an e-commerce merchant sells a product, it often goes through this process – using good content to attract people to come in and take a look, using good customer service to answer customer questions, using good logistics to send and deliver the goods in a timely manner. We can also provide good after-sales service when problems arise.

In the content aspect, JD Cloud DaaS affects the three core elements of marketing by accessing the Yanxi large model: audience channels, content materials, and data.

In terms of determining the audience, in the past, merchants’ advertising mainly used labels to delineate groups of people, and operators judged how much resources to invest in certain groups of people, and increased or decreased the amount of investment at different points in time. Nowadays, with the help of data accumulated by merchants, large models can help merchants first conduct a test on a small group, then analyze the test results and quickly give suggestions for increasing or decreasing the amount of the product.


AIGC content marketing platform

In terms of the generation of marketing materials, JD Cloud AIGC content marketing platform can help merchants quickly generate product main images, marketing poster images, and business detail images required for operations, to meet their needs for rapid store opening and marketing, and make the production of each set of images easier. The cost was reduced by 90%, and the production cycle was shortened from 7 days to 0.5 days.

Yanxi large model + DaaS can also provide more accurate data analysis and management for merchant advertising. Mu Xiaohai said that the large model can combine industry data to make suggestions for merchants’ product production design; after the product is formed, the large model will determine the functions that should be emphasized in the marketing copy for the product based on the people for whom the product is targeted; for digital products that are iterated on a regular basis , Yanxi’s large model can analyze the comments on the previous generation product and the feedback often received by customer service, and provide suggestions for the update of the next generation product.


Jing Xiaozhi

In terms of customer service, JD Cloud has upgraded its intelligent customer service product JingXiaoZhi. Previously, Miniso still used a lot of manual customer service, but after joining Jingxiaozhi, it saved a lot of labor costs. The response accuracy of Jingxiaozhi’s online customer service robot exceeds 97%, and the independent reception rate exceeds 70%, reducing service costs by 40%. The response accuracy of the voice response robot exceeds 93%, and it independently handles 46.1% of customer problems.

In addition, JD Cloud’s smart outbound calls can also help merchants reach users and provide special services. During the period when Yili’s Jindian Organic Milk sponsored the variety show “Sister Riding the Wind and Waves”, Yili chose the highly recognizable voice of Cyndi Wang to call users, and sold 60,000 bottles of the new product that day, with an input-output ratio as high as 1:4.

JD Cloud also combines smart marketing with smart services to realize 24-hour live broadcast by AI virtual anchors through Yanxi multi-modal digital people. In cooperation with Lenovo Group, JD Cloud customized a virtual anchor image based on Lenovo’s brand IP image. Such a virtual anchor has a built-in knowledge base for all categories of the retail industry, and 90% of the questions raised by users about Lenovo computers can be effectively answered. Answer, comments in the live broadcast can also be automatically replied.


Yanxi multi-modal digital human

The virtual anchor can also automatically recommend products based on user behavior, and live broadcast skills can be generated with one click, eliminating the need to spend half a day preparing. At present, the daily transaction amount of virtual anchors can reach up to 2.3 times that of real anchors, and the average hourly transaction amount accounts for 45% of that of real anchors, but the cost is less than one-tenth of that.

Just half a year ago, many companies claimed to be China’s OpenAI, but now not many people have mentioned this goal. The new narrative is: industry big models and “big models empower thousands of industries.”

The reasons behind this are: the parameters of general-purpose large models often reach hundreds of billions, and the GPU computing power consumed to run them is high and the cost is expensive, which is unaffordable for ordinary companies; in specific scenarios, general-purpose large models are inaccurate; and the parameters of large industry models are small. , the cost is lower, and after targeted training, it can answer questions in vertical fields more accurately, and only by helping vertical fields solve problems will the possibility of commercialization be greater.

There are still only a few manufacturers that have actually found implementation scenarios for large models and promoted them for external commercial use. The biggest difficulty is that the quality of data used to train industrial models is often not high, and industrial data is often non-public and scattered in the upper, middle and lower reaches of the industry. Problems such as unstable data collection and fragmentation may occur in actual usage scenarios, which will cause the professional depth, service accuracy, and iteration speed of large industrial models to fail to meet the requirements of actual use.

He Xiaodong, president of JD Discovery Research Institute and president of JD Technology Intelligent Services and Products Department, said that industrial data is divided into static data and dynamic data. Static data is relatively stable and the acquisition path is relatively clear. The disadvantage is that the data does not change immediately and there is a lag.

Dynamic data is data generated every moment in different industrial scenarios. This part of the data is “live” scenario data, which is not easy to obtain, but it is one of the necessary elements for large industrial models.

Dynamic and rich industrial data are exactly the advantages that has accumulated over the years.

Digital supply chain, JD Cloud’s unique advantages

“Currently there are very few teams that have capabilities in the four dimensions of product, engineering, algorithm, and data.” Chen Feng believes that from the data dimension, does have “mines at home”, but this is far from enough. Good technology is needed for alchemy. . began to build its own cloud business unit in 2014, and began commercial use two years later. At the end of 2017, established an artificial intelligence research institute. In the second year after this strategy was set, JD Cloud created the Yanxi artificial intelligence application platform and a series of industrial solutions.

Compared with other DaaS vendors, JD Cloud also has a unique advantage, which is that it can use JD’s accumulation in the supply chain to cover the last and most important link for e-commerce merchants – logistics.

Costco and Walmart, two retailers recognized by the industry as having extremely strong supply chain capabilities, have inventory cycles of 31.3 days and 42.5 days respectively. The number of SKUs they manage is only 4,000 and 50,000.’s 30-day level is already close to the limit of the retail industry.

This set of digital supply chain capabilities accumulated by will become more efficient after being superimposed on the Yanda model, and will also better help merchants achieve the circulation of the same product through multiple channels.

It is able to mobilize tens of millions of goods on, relying on the automation capabilities of JD Logistics, including intelligent picking robots “Sirian Wolf” and “Dilang”, intelligent robotic arms and cross-belt ultra-high-speed sorting systems, and large models can help This automated system increases efficiency.

For example, during a major promotion, the efficiency of the warehouse decreases. In the past, it was necessary to manually survey the warehouse to analyze which link had the blocking point. Now you only need to submit the warehousing data to the big model analysis, and the latter will tell you that the efficiency blocking point is The number of Dilang shelves is not enough, so it is recommended to add another row of shelves.

In the cooperation between JD Cloud and Yili, Yili has integrated warehouses from multiple channels such as, Douyin, and Kuaishou. In the past, only Douyin warehouse could send Douyin goods. Now when channels are sold out, Douyin warehouse can also transfer goods. This reduces Yili’s inventory costs, transportation and distribution costs, and shortens delivery time.

In addition to logistics, JD Cloud has also accumulated reverse customization capabilities.

Taking water purifier products as an example, Jingdong Tokyo Manufacturing Co., Ltd. and Zhejiang Aibote will begin cooperating to produce under-kitchen water purifiers in 2022. Jingdong Made in Tokyo is deeply involved in product R&D and design from three dimensions: consumer trend insights, category and price positioning, and product design details. It uses the C2M model to reversely guide factories in product function definition and design R&D, so as to produce products that better meet consumer needs. The two parties have successively launched reverse osmosis kitchen water purifier products with different fluxes such as 600G, 800G, 1000G, 1200G, etc. This year’s 618, and Aiport jointly launched a full line of water purifier products that doubled in total sales and once again achieved great results. .

JD Cloud’s reverse customization can be roughly divided into several steps: First, rely on data on the JD platform to help merchants gain insight into demand. For example, JD found in the strong category of digital electronics that many users have demand for “game notebooks”, so JD and Lenovo has cooperated to launch the Savior Blade series and the HP Shadow Elf series of game notebooks, with current sales exceeding 1 billion yuan.

After discovering the demand, will also conduct simulation trials on the platform. Brands can sell simulated products on the platform and collect consumer feedback on new products through functions such as questionnaires and trial product placement.

After determining that consumers have demand for this type of product, will also work with brands on research and development. For example, the traditional display product supply chain needs to pass through panel factories, foundries, brands, and then to consumers. New products usually cost nearly It takes 18 months to go on the market; directly feeds user needs to upstream foundries, and some products can be delivered to users in less than 6 months.

Finally, JD Cloud will also help brand owners conduct cross-analysis of JD’s six-dimensional data and launch jointly with them. Take the Head & Shoulders silicone-free small green bottle series products as an example. During the product launch period, it has gone through planting, evaluation, water storage, and explosion. , consolidated the whole process, and promoted by marketing resources such as new product launch, super coupon day, super new product day, etc., achieved sales of over 10 million within 4 months of launch.

When large models improve the granularity and accuracy of data analysis, the growth paths for brands become richer and more certain, which is particularly valuable in today’s environment of compressing costs and reducing investment.

For traditional brand marketing, products, channels, and advertising creativity were once the key influencing factors. As big models gradually transform the advertising and retail industries, how to make good use of big models and use data analysis tools will also become a brand One of the key influencing factors of marketing. In this process, JD Cloud has taken the first step of trying.

This article is reproduced from:
This site is only for collection, and the copyright belongs to the original author.