From craze to rationality, large-scale models ushered in the "turning point" of the industry

Image source: Generated by Unbounded AI

Everyone talks about the big model, which is a true portrayal of the technology world in the first half of this year.

Judging from the excitement of the market, large-scale models have become a track for various technology manufacturers to rush into. Whether it is Internet giants, technology companies, or even research institutions, they have all joined this large-scale model melee. Large-scale models have become "Battleground".

Just when the domestic large-scale model market is in full swing, ChatGPT, which has driven the popularity of large-scale models in one fell swoop, has experienced a decline in visits. According to the latest data from SimilarWeb, a third-party monitoring agency, in May this year, ChatGPT began to show a slowdown in growth, with a growth rate of only 2.8% in that month, while the first four months of 2023 were 131.6%, 62.5%, 55.8%, 12.6%. This is the first time that ChatGPT has experienced negative traffic growth since its release on November 30, 2022.

This phenomenon may reflect an important industry trend. All parties are changing from the initial technical enthusiasm for large models to calm thinking about commercialization. And the landing of large-scale models is also a topic that all large-scale model companies need to seriously consider.

** "JD.com believes that the value of a large model = algorithm × computing power × data × square of industry thickness." said Xu Ran, CEO of JD.com. JD.com not only pursues the advancement of technology, but also pays special attention to the thickness of the industry - it values how many industrial scenarios the technology can be practically applied and can create various values for the society. **

On July 13, JD.com launched a large model of Yanxi. According to JD.com, this is a new-generation model with parameters reaching hundreds of billions of dollars, and it will be deeply used in retail, finance, logistics, health, industry and other industrial scenarios in the future.

When the industrial efficiency and the expansion of the industrial boundaries are qualitatively improved, the large model will have more important practical value and significance. In the hustle and bustle of large models, we should return to rationality and seriously consider the true value of large models. What kind of big model does the industry need? How should the large model be put into commercial use to reduce costs and increase efficiency for the industry?

Competing for the big model: the consensus from general to industry

Overnight, domestic large-scale models "emerged" one after another.

According to the "China Artificial Intelligence Large-scale Model Map Research Report" released by the China Institute of Science and Technology Information under the Ministry of Science and Technology, as of May 28, at least 79 basic large-scale models with a scale of more than 1 billion parameters in China have been released. The number of large models developed in my country has ranked second in the world.

However, behind the soaring number of R&D, how to realize the commercial value of large models is worth exploring.

At present, more general-purpose large models are coming out. This type of large-scale model has powerful natural language understanding, language generation, and speech recognition capabilities, and performs well in scenarios with strong general attributes such as chatting and entertainment. However, these scenarios are still difficult to achieve large-scale commercialization.

In subdivided industrial scenarios, the answer accuracy of the general large model is low. This is because the industry itself has a small sample size, uneven data distribution, and changing application scenarios, which make the large model unable to self-optimize and improve. The accuracy is naturally not high.

From the perspective of commercialization, enterprises may not need an "all-round" general-purpose large-scale model, but more need an industrial large-scale model that targets subdivided field scenarios and solves practical problems.

Wu Hequan, an academician of the Chinese Academy of Engineering, once pointed out that "Chat-like large models have triggered a new round of upsurge, but dialogue, poetry writing, and painting are by no means all of large models. We need to think deeply about the application direction of large models, and we must actually put large models into In the fields of urban development, financial technology, biomedicine, industrial manufacturing, and scientific research, professional enterprises and organizations are also needed to accelerate their implementation in the real industry, bringing real value to the industry's immediate needs, and truly serving the society on a large scale. "

Nowadays, a consensus is being formed in the large-scale model market that large-scale models that cannot be put into commercial use are just "entertainment tools", and only large-scale models that go deep into the industry and solve practical problems have value. Thinking deeply about the value of large models, a formula proposed by JD.com is more in line with the trend of the large model market, namely: the value of large models = algorithm × computing power × data × square of industry thickness.

** From the value formula of the large model, we can see that the industrial large model and the general large model are not in an opposite relationship. The industrial large model is based on the general large model and is trained. The content is more in line with the needs of industrial vertical scenarios and more targeted. **

In terms of the evolution route of the large model, it is not surprising that JD.com is forward-looking. Since its establishment, Jingdong has been rooted in the industry, and naturally pays more attention to the value of large models in the industry.

"Jingdong regards large models and other technological innovations. In addition to pursuing the advanced nature of technology, it also pays special attention to the thickness of the industry - how many industrial scenarios the technology can be practically applied to truly create value for the society." said Xu Ran, CEO of Jingdong.

**In fact, the development of large models at this stage is moving from "universal" to "industrial". **

The latest research report released by Minsheng Securities mentioned that after the intensive release period of large models from February to March, the product development period from April to May and the policy direction were gradually clarified, the products and applications of large models will start in June. It is expected to usher in a centralized release. The new wave of releases is based on large-scale model application products, and large-scale upgrades have begun to go online, preparing for entering thousands of households.

Recently, most of the latest large-scale models released by domestic technology companies are aimed at vertical industries. Overseas, various companies have successively released large-scale industry models of different scales, with the purpose of applying them to the industrial field.

From the actual trends of major manufacturers at home and abroad, it is not difficult to see that the industrial model will be more able to help industrial partners to complete digital transformation, reduce costs and increase efficiency, and create greater value for the industry and society.

Large industrial model, it is difficult to lay a solid foundation for the industry

The difficulty of constructing industrial large-scale models is much higher than that of general-purpose large-scale models.

If the general large-scale model tests the computing power and algorithm accumulation of the enterprise, then the industrial large-scale model tests the enterprise's access to and understanding of business scenarios, as well as the accumulation and application of industrial data.

An industry consensus is that in the training of industrial large models, the most difficult thing to obtain is industrial data. Industrial data is often in the hands of enterprises. Due to data security and other considerations, few enterprises are willing to disclose private data. However, these industrial data often directly or indirectly affect the technical iteration speed, model accuracy and business professionalism of the industrial large model.

"Industrial data can also be divided into static data and dynamic data. Static data is relatively stable, will not change immediately, and the acquisition path is relatively clear. Dynamic data is the data generated every moment in different industrial scenarios. This part of the data is 'Living' scene data. It is not easy to obtain, but it is one of the necessary elements of the industrial model." He Xiaodong, President of JD Research Institute and President of JD Technology Intelligent Service and Product Department, emphasized.

However, the training of industrial large models cannot only use industrial data, but still needs to use a large amount of general data to provide commonsense knowledge. The reasons are as follows: First, the generalization of industrial data is insufficient, and the large model needs to be retrained every time the scene is changed, which is costly; Stuck in a stuck state.

He Xiaodong compared the training of a large industrial model to training a person, "If a person goes out to work directly after graduating from high school, it seems okay, but the professionalism will be less professional. If you can finish a four-year undergraduate degree before going out to work, you have both general knowledge Ability and sufficient professional knowledge are the capabilities that a large industrial model should possess.”

For this reason, the data of JD Yanxi's large model is composed of 70% general data and 30% raw data of supply chain scene growth.

It is worth mentioning that these industrial data come from JD.com itself. JD.com itself is a company based on the supply chain. It is rooted in a wide range of industries. It not only has practical data in retail, logistics, finance, health, industry and other industries, but also has data on cities, government affairs, finance, manufacturing, industry, aviation, transportation, etc. The desensitized data of industries such as industrial parks, industrial parks, and energy, and the high-quality data generated each year reach 10 billion pieces.

In addition to the continuous supply of high-quality industrial data, the industrial large model also needs to understand the industry Know-How, that is, to have unique knowledge about the industry and have higher requirements for comprehension. For example, the retail industry pays more attention to the effect of marketing and recommendation, and the financial industry pays more attention to the effect of risk control, reliability and safety.

For this demand, JD.com's long-term digital intelligence supply chain has played a key role, and it has become the focus of JD.com's efforts in the application of large models. The large model can also be based on the digital intelligence of the supply chain and go deep into the physical industry.

It is reported that JD.com's digital intelligence supply chain has covered more than 10 million JD.com's self-operated product SKUs, serving more than 8 million active corporate customers, including more than 90% of the world's top 500 companies in China, and nearly 70% of the country's specialized Special new SMEs. At the same time, JD.com's digital intelligence supply chain is still in the country, and has in-depth cooperation with more than 2,000 industrial belts.

This kind of digital intelligence supply chain with longer links, more complex scenarios, and richer data is an excellent "training ground" for large models. In JD Cloud's view, the value of the large model can only be realized by thoroughly understanding the supply chain and allowing the large model to "run" on the supply chain.

In addition to the accumulation on the industrial side, JD.com's strength in basic algorithms and computing power should not be underestimated.

In 2021, JD Discovery Research Institute launched the country's first ultra-large-scale computing cluster based on the DGX SuperPOD architecture in Chongqing-Tianqin α, which increased the speed of reasoning by 6.2 times and reduced the cost of reasoning by 90%. This provides JD.com with the most basic guarantee for large-scale model training.

In the same year, JD.com launched the billion-level model K-PLUG. The product copy generated by K-PLUG has covered more than 3,000 categories in JD.com, generating a total of 3 billion words, and the manual review pass rate exceeds 95%. By 2022, JD.com's large model will be upgraded to a tens of billions model Vega, which can be widely used in various downstream natural language processing tasks such as sentiment analysis, semantic matching, grammatical error correction, intelligent question answering, and common sense reasoning.

Thanks to previous accumulation, JD.com made another technological breakthrough this year, and launched a new generation of JD.com’s large-scale model with hundreds of billions of parameters, focusing on several major tasks such as content generation, man-machine dialogue, user intent understanding, information extraction, and emotion classification. , realized the fine-tuning of base model + vertical domain model, and applied in-depth vertical scenarios such as retail, logistics, finance, health, and government affairs.

At present, the large industrial model represented by JD.com is training the large model through its accumulated industrialized and scenario-based data and knowledge, and correcting the large model based on its accumulated industry Know-How , to improve the performance of large models in specific industries and application scenarios, and to improve controllability. This is equivalent to completing a "general education" for AI.

**Jingdong’s big model is gradually going deep into various industries to improve the intelligence level of the supply chain. Conversely, the digital intelligence upgrade of the supply chain is also promoting industrial transformation, which in turn provides richer data soil for large models, forming a positive cycle. **

"Cutting into the large model from the industrial side is like climbing the technical Mount Everest from the north slope: although the road is more difficult, there are more magnificent scenery. Jingdong insists on doing 'difficult but correct things', insisting on doing practical, valuable and It is a long-term matter. In the technical field and in the large model, this is our constant commitment." Xu Ran said.

** "Difficult and correct things" need long-term accumulation. However, in terms of the construction of the industrial model, JD.com has clearly laid the foundation. **

Jingdong, born in the industry, creates industrial value with large models

"In the age of the big model, anything is worth doing over again with the big model."

Under the wave of large models, the industry quickly reached the above consensus. However, while other major manufacturers are looking for business models through various strategies, the direction of Jingdong's landing industry has never changed.

"Industrial attributes are the distinctive features of JD.com's technology. Every technology developed by JD.com stems from industrial needs, experiences in industrial scenarios, and creates industrial value." Xu Ran said.

At present, JD Yanxi’s large-scale model is moving forward in accordance with the “three-step” strategy: at present, JD Cloud has built a general-purpose large model based on internal practices; by the end of this year, JD. Solid industry services; it is expected that in early 2024, the large-scale model capabilities will be opened to external serious business scenarios.

From a practical point of view, JD.com's large model has reached the second step. JD.com is applying the capabilities of the large model to the most familiar scenarios such as retail, finance, logistics, and health, and has penetrated into various links.

For example, Jingdong is improving the ability of intelligent customer service through large models. The field of customer service is different from daily chats and conversations. It is a serious task-based dialogue scene that needs to solve various complex problems between buyers and sellers in the real world.

"When a user talks to ChatGPT, it doesn't matter even if the answer is wrong, and it will not affect any decision-making. But if in a serious business scenario, the intelligent customer service answers wrong, the consequences will be unimaginable. Therefore, the accuracy of the answer is very important." Jingdong Group Technical Committee Cao Peng, Chairman and President of JD Cloud Division, said.

For intelligent customer service scenarios, JD.com not only built a large model with basic semantic understanding and question-and-answer logic, but also polished a small model for specific scenarios. If the customer's problem involves common returns and exchanges, etc., the intelligent customer service will call a more general large model. And once the question involves the warranty policy and price protection rules of specific products, the intelligent customer service will call a more targeted small model to give the answer. Different models can take on different responsibilities.

Now, the smart customer service is working inside JD.com, helping more than 20,000 self-owned customer service employees reduce costs and increase efficiency, and continuously optimize customer service experience. JD.com has also opened up its intelligent customer service capabilities to the outside world to help more government agencies and enterprises carry out digital and intelligent transformation and upgrading.

In terms of external services for large models, JD.com still maintains its own "slow" pace, and is not in a hurry to "sell" large models to enterprises. The reason is that artificial intelligence is a very serious technological change: if used well, it can transform the industry, but if used incorrectly, it can also have serious consequences. Under such circumstances, Jingdong adheres to the long-term mentality and is an excellent choice.

"JD.com will not serve dishes that have not achieved 'full color, flavor and taste'. After the large-scale model has completed experience and practice in key internal scenes, it will be opened to partners to help the entire industry reduce costs and increase efficiency." Xu Ran express.

In JD.com's plan, Yanxi's large model will become the lowest-level technical support. Based on its capabilities, the field will produce a series of products, and the products in the same field will be aggregated into a platform, and finally output value to the industry.

For example, in the field of content generation, JD.com has built the JD Cloud AIGC content marketing platform. Based on the rich product data accumulation of JD.com’s entire category, the large model can better understand product characteristics, help merchants automatically generate product pictures, selling points and other marketing materials, and improve merchants’ operational efficiency and marketing content quality.

In other words, merchants only need to upload a product picture, and can quickly obtain multiple types of pictures such as main product pictures, marketing poster pictures, and business detail pictures needed for e-commerce operations, meeting the needs of quick store opening, product listing and marketing. These capabilities can save merchants 90% of the cost of drawing, and shorten the production cycle from 7 days to half a day.

Using more large-scale model capabilities, merchants don't even need to sell their own goods, but only need to use JD Cloud's multi-modal digital human to deliver goods 24 hours a day at low cost.

In addition to these applications, JD.com also demonstrated an AI marketing operation platform in the financial field. Through simple dialogue, users can generate marketing activities in one stop, covering operational strategy formulation, marketing task scheduling, building activity pages, batch generation of marketing copywriting and materials, digital delivery, etc. According to relevant data from JD.com, in the past this set of processes required five types of functional personnel: product, R&D, algorithm, design, and analyst, but now it has been reduced to one person; in the past, the process required 2,000 human-computer interactions, but it has also been reduced to less than 50. With the support of the large model capability, the production efficiency of the marketing plan has been significantly improved.

** It can be said that JD.com's large industrial model is becoming the base for industrial applications in various industries. At present, it has penetrated into retail, logistics, health, industry, manufacturing, finance, marketing and other industries, and has gone to industry and practical application. **

Half a year after the big model ran wild, manufacturers have realized that "large scale" or "high parameters" cannot solve practical problems. When the bubble period is over, large models must return to the industry, return to real scenarios, and solve practical problems after all. This is the ultimate destination of technology and the beginning of the benign development of large models. Those companies with solid industrial data and scenario practice began to come to the stage.

Obviously, the Jingdong Yanxi large model has set sail to the other end of the industrial value.