The Era of Large Models Has Arrived, the Cutting-edge Thinking & Practice of Belle Fashion

[Editor’s note] Digital service providers have always been an indispensable and important part of the e-commerce ecosystem, but they have been ignored because they are too familiar. It seems that we should not forget that in every iteration and evolution of e-commerce, digitalization has acted as a technology pioneer and a builder of application scenarios. We believe that the value of digitalization is indispensable in the future of business driven by big data, artificial intelligence, and algorithms. In order to show the appearance and capabilities of e-commerce in more detail, Ebang Power has specially launched digital practical cases. Through the case dismantling of each issue, we will find the most practical industry-level solutions for e-commerce practitioners. And in every innovative application, find a glimmer of the future.

Before the advent of the era of large models, even if most enterprises realized the assetization of data elements, it was difficult to have real data intelligence because of the lack of advanced data asset production tools.

Belle fashion’s cutting-edge thinking and practice

Few people know that Belle Fashion Group, which has nearly 10,000 stores in more than 300 cities in China, once had a professional product operation team of a certain scale to follow up a series of business operations such as “ordering, replenishment, and transfer” of goods.

The team’s job is to sort through the sales, inventory and other information of all Belle Fashion’s shoes across the country, so as to quantify the life cycle of each pair of shoes – which ones sell well, which ones are not selling, which city prefers which shoes…… Relying on these vast amounts of data, we provide stores with the basis and business suggestions for the operation of goods.

This is an extremely important task. Belle Fashion has about 20 brands, and for such a large fashion footwear and apparel group, inventory optimization and supply chain management are the lifeblood of the brand’s long-term development.

Bringing intelligence to data is a complex, time-consuming and labor-intensive project, and in the past, the work was all about “modeling”: the superposition of layers of indicators, complex calculations, and the unified definition of data logic and dimensions…… In order to spell out the purchase, sale and inventory trajectory of a pair of shoes, and to make business decisions by manually sorting out information and knowledge. Although it is machine-assisted, it is not an exaggeration to say that this is a labor-intensive type of work.

This work will be gradually handed over to AI in the future.

Starting from 2021, Belle Fashion began to cooperate with Deepexi Technology, a data intelligence infrastructure provider, and a data application system called “Beauty Shadow” was born. At the end of 2022, with the explosion of generative AI led by ChatGPT, the data real-time lakehouse platform that has been built for many years will show more powerful productivity with the blessing of AI large models.

The emergence of large models makes the Deepexi Technology team excited, the reason is simple, in the past, after the enterprise launched the data platform, it also needed a certain amount of manpower to sort out the documents about “Know How”, and then develop various tables and models according to the documents to support business decisions. Now, after being trained on a large model, these “Know Hows” have the ability to understand business reasoning and automatically generate code, and from the perspective of system verification, “completely run through!”

The self-generation ability of this large model is the most popular “AI Agent” today.

The so-called agent can be understood as an “artificial brain” that does not need to be driven by humans and can flexibly use tools according to the context. Take an example of a game scenario: Stanford University has launched a “virtual AI town”, in which there are more than a dozen agents with different identities, players can talk freely with these agents, and the AI will build a copy of the game according to the player’s behavior.

In the To B scenario, AI Agent will bring huge business value.

For enterprises, the large model is the “brain”, the Agent is the “hand”, taking the “replenishment” operation in the operation of goods as an example, when the clerk enters the suggested replenishment instruction, the Agent can automatically generate the SQL query code, first identify the supply target, comprehensively calculate the out-of-stock, etc.;

It’s a bit hard to understand, it doesn’t matter, for the user, just type a simple text command in the search box: “What is the restock suggestion for XX products?” and then hand it over to the AI.

On November 28, at the WISE2023 King of Business Conference held by 36Kr, Zhao Jiehui, chairman and CEO of Deepexi Technology, said: “For enterprises, it is a very correct direction to focus on the construction of data platforms and realize intelligent development based on data. Although over the past decade, many have encountered great challenges and difficulties in this direction. However, as the large-scale model begins to really land and integrate with many industries this year, this matter may usher in a new starting point. ”

In order to allow more enterprises to have large model capabilities, Deepexi Technology released the enterprise large model Deepexi in September this year, according to different industry indicators and knowledge systems, to build corresponding production field data models, supply chain operation management field models, financial management field models, dual carbon field models, etc., according to different industry indicators and knowledge systems, so that the data assets of enterprises can play a real intelligent value. At the same time, it is equipped with the Fast5000E cluster computing power of the Deepexi Technology training and pushing all-in-one machine and the FastData data capability of the real-time intelligent lakehouse platform to realize the application of artificial intelligence in the industry in a complete and agile manner.

Large model, the dawn of China’s To B market

Before the advent of the era of large models, most enterprises could not have true data intelligence.

In the past ten years, most “intelligent” data platforms only provide screening of core indicators, such as how many pairs of XX shoes were sold last week and how much inventory is left;

It is precisely because they have not achieved true “intelligence” that many “big data products” on the market are useless within enterprises, which also makes CIOs/CTOs who focus on digitalization feel anxious.

“Many digital leaders of large groups have spent a lot of money on data platform tools, and finally found that there is no in-depth business value at all except for large screens and reports. If you want to form a business model based on data, you need to invest a lot of data development engineers to develop, which will put a lot of pressure on the gross profit of the project”, Zhao Jiehui said in one sentence for the stubborn problems in the process of enterprise digitalization.

In 2018, he founded Deepexi Technology, based on his long-term in-depth observation in the front line of enterprise services, Zhao Jiehui divided China’s enterprise services into three stages:

The first is the era of standard tools. With the wide application of big data basic technology represented by Hadoop, giant enterprises have been able to carry out ultra-large-scale offline data processing, and “big data intelligence” has shown its embryonic form.

But tools don’t generate value.

This brings us to the next stage: the era of heavy service. The biggest difficulty of enterprise services lies in the thousands of customer scenarios, and it is difficult for a single product to meet all customer needs, which requires a large number of pre-sales and engineers to join in the role of pre-sales, who have to go to the front line of the customer’s business to find out the actual characteristics of the customer’s supply chain, inventory scheduling, product sales, etc., in order to play a role in the customer’s real business scenarios.

The problem still exists, from the sale of tools to “tools + services”, customers do have a stronger sense of gain, but for technology providers, service means cost – a large number of engineers enter the site, high man/day deployment costs, and the implementation cycle of annual at every turn, so that the original digital project with smart labels has become a manpower-intensive type of work, and it has also dragged down the gross profit of countless enterprise service companies.

This has also made China’s enterprise service industry now enter a strange circle: only sell standard products, customers do not recognize, do heavy service, manufacturers lose money.

Looking at the enterprise service market in the past two years, whether it is the financing scale of the primary market or the listing situation of enterprise service companies in the secondary market, it can only be described as a “trough”.

The emergence of the large model has become a ray of light in the darkness of China’s enterprise service market.

True data intelligence: unified data platform + AI agent + computing power

Before describing the huge changes brought about by large models, let’s first consider the question, why can’t the data platforms of the past really play a valuable role?

If large models and data applications are high-rise buildings, the data assets of enterprises are the foundation, and if you want the building to be stable, the foundation must be solid.

Once upon a time, the reality of an enterprise’s data assets was not as good as it could be. Especially for large enterprises such as power grids, the internal systems are complex, and if you want to calculate electricity bills, hundreds of systems have their own statistical calibers, which requires enterprises to first unify internal indicators and form a consistent data platform.

Second, before the advent of large models, the data platform’s ability to understand was very limited, which meant that the machine was not “smart” enough. As mentioned above, in order for the data platform to have a more agile response and decision-making ability, the machine needs to understand what a “hot product” is and give specific replenishment suggestions – this natural language understanding ability has been an obstacle to data intelligence for a long time.

Now, with the emergence of intelligence brought by large models, machines have become smarter, and the widespread use of open source large models such as Llama has lowered the threshold for developers to use large models, the above problems have been solved one by one, and the curse that once bound China’s enterprise service market is gradually being unraveled.

“The real foothold of China’s enterprise service market is a unified data platform plus AI Agent and computing power, the data platform allows enterprise services to have a basic data base, AI Agent can make the value of data business agile, and computing power provides a solid foundation for large models. Zhao Jiehui pointed out a clear development path.

The wave of digital transformation that has lasted for many years has also significantly improved the data governance capabilities of enterprises. Taking FastData, a real-time intelligent lakehouse platform provided by Deepexi Technology, as an example, it includes modules such as enterprise multi-business data integration, real-time computing and storage, development and governance integration platform, data analysis and label management, etc., which can solve all the needs of enterprise data analysis in one stop and provide a solid “foundation” for digital delivery of enterprises.

At the AI Agent level, large models also bring new opportunities for data intelligence, with Bill Gates writing in early November: “Agent will not only change the way people interact with computers, but will also disrupt the software industry, triggering the biggest revolution in the computer field since we went from typing commands to clicking icons.” ”

In order to integrate large models into products as soon as possible, from the end of 2022, Deepexi Technology began to develop AI Agent products, and after more than half a year of polishing, it has launched the “FastAGI agent platform” product, which can realize functions such as Text-to-SQL (that is, automatically generate query instructions through text, such as querying popular models), Text to API (that is, automatically pull ERP and other interfaces through text instructions to complete business actions such as “replenishment”) and other functions.

To put it simply, Deepexi Technology’s AI Agent product is to use the current technological breakthrough of large language models to achieve interactive and accurate retrieval of data/indicators, generate rich data charts, and automatically call system APIs to complete business needs;

Compared with traditional IT data products, Zhao Jiehui told 36Kr that the unified data platform (FastData) + AI Agent tool (FastAGI) of Deepexi Technology has great revolutionary significance.

First, large models enable data platforms to be deployed faster. Zhao Jiehui talked about the data platform case of the cultural and tourism management department of a city in China, and after the data platform of such customers was built in the past, it was necessary to sort out the business logic, build SQL, and debug parameters…… It takes at least 3-5 months to achieve initial results, and now with the data tool of large model fusion, you only need to sort out the customer’s business logic a little, and you can land the product in 2-3 weeks, “it will be easier to talk to the customer”.

Secondly, the model of data platform + AI agent tool allows the once labor-intensive big data work to be “reduced”, and at the same time, the ease of use of large-scale model products can also allow digital products to be distributed to stores, so that front-line store managers can also enjoy the dividends of digitalization.

Not only that, through the interaction with the data of the machine, the store managers of Belle Fashion can also let the large model of Belle Fashion learn the thinking mode and behavior logic of the store manager, so as to iterate more intelligently and grow into a competent “digital store manager assistant”.

With the gradual strengthening of the function of large models, the computing power support behind them is particularly important.

Not long ago, the two sides reached a project cooperation on the large-scale model training and push all-in-one machine Fast5000E to jointly carry out industrial innovation practice, becoming one of the first projects in the industry to achieve industry landing. Fast5000E is an AI server jointly released by Deepexi Technology and Huawei Ascend, equipped with Deepexi Technology Deepexi, which can provide enterprises with integrated solutions for training and inference computing power for large models, and is an important deployment of Deepexi Technology at the computing power level.

Break the obsession and face up to the value of service

According to the conventional Chinese enterprise service idea: when a To B manufacturer builds a set of standardized products, whether it is a data platform or an AI agent tool, an inertial approach is to expand the sales of standard products and reduce customized services, so as to increase project revenue and profits.

But the blind pursuit of standardization and reduction of services is the correct way for enterprise services to land?

The answer is no.

“In the domestic enterprise service market, there is always an obsession, that is, we must do standardization, and we must let go of this obsession. In an exclusive interview with 36Kr, Zhao Jiehui said in a firm tone, “Service is essential!”

What needs to be pointed out in particular is that the “service” that Zhao Jiehui is talking about here is no longer a labor-intensive service that consumes high labor costs, but a more agile, more intelligent, and more customer-friendly service.

In other words, the current situation that the enterprise service market needs to face up to is not whether to customize, but how to make customization more agile.

How to break the curse of “service and gross profit” in China’s To B field? The answer is still a large model.

First of all, with the support of large models, which used to cost a lot of money, a large number of engineers were required to do data classification, annotation, logical combing, and indicator definition, etc., and the large model can operate on its own with simple training. The rapid launch of the cultural tourism data platform of a city mentioned above is a true portrayal of the agile landing of large models.

The improvement of service efficiency can directly correspond to the reduction of labor costs. According to 36Kr, Deepexi Technology’s contracts and revenue this year have doubled compared with last year, and the gross profit has increased by 15% compared with last year.

In addition, with the “volume” of large model parameters, hundreds of billions of large models have emerged in the industry, but for a single enterprise, the larger the better, but the more suitable the scenario and the more efficient.

“It’s not that the bigger the model, the better, because the bigger the model means the cost of computing power is high, but we build a small but precise domain model for enterprise business, which is not only low cost, but also more efficient. Zhao Jiehui said.

This is also the reason why Deepexi Technology has built a corporate model. According to the needs of enterprise business scenarios, Deepexi, a large enterprise model released by Deepexi, includes the data model of the production domain, the domain model of supply chain operation management, the domain model of financial management, and the model of the dual carbon field.

The product form of the “enterprise model” is also more in line with the needs of enterprise customers in the digital transformation: starting from a single field, quickly running results, summarizing experience, and then extending to other links, which not only ensures that the effect of the large model is easier to see, but also the overall digital solution is more cost-effective.

Large-scale replicability is also another feature of Deepexi Technology’s large-scale model products.

Production, Supply Chain, Financial Management, Carbon Peaking and Carbon Neutrality …… As the name suggests, domain corpora are similar between firms. Take the “replenishment suggestion” scenario in the field model of supply chain operation management as an example, the “shoe king” Belle Fashion can be used, and other retail brand manufacturers can also use it. In other words, Deepexi, the enterprise model of Deepexi, can be adapted to different types of enterprise customers with only simple debugging.

After the standardization of products has been polished, service is still important, but the focus of service has shifted to the understanding of enterprise business.

If the enterprise model is more in line with the rules of business operation, it is necessary for industry experts to “feed” the machine the appropriate prompt based on an in-depth understanding of the business flow, so as to train the enterprise model into an ideal state.

Based on the deep insight into the industry and rich experience in digital transformation, Deepexi Technology has created a business value innovation service DIC with data intelligence as the core, and the DIC team is composed of experts in various industries to help enterprises improve the construction of data infrastructure, and precipitate industry data based on the experience of enabling the digital transformation of various industries, and further build an industry knowledge base.

“The logic of enterprise services is starting to simplify. Now enterprise digitalization is to do global intelligent construction, not only a tool, but a data platform + AI. Zhao Jiehui summed up 36 Kr.

In the business world, each industry has its own development curve at different stages, and each cycle will also give birth to corresponding consensus. Nowadays, we are standing at the moment of the outbreak of AI, everything is revolutionizing from the bottom, and practitioners in the field of enterprise services also need new guidance and new collisions. The enterprise model is bringing a new dawn to China’s To B market.