Large-scale models take root in the industry: from "emergence of capabilities" to "emergence of value"

Original source: Yibang Power

Image source: Generated by Unbounded AI‌

The "grand movement" of artificial intelligence AI that sounded from the beginning of 2023 is no longer so "sweet" recently, but mixed with noise and doubts.

**For example, traffic disputes. **

According to data from the overseas research organization Similarweb, after OpenAI has made great strides all the way, since May, the traffic growth has flattened. In June, the number of visits to ChatGPT even showed a month-on-month decline for the first time, with a rate of 9.7%. According to this data, some people say that AI is going to be cool again.

However, there are also public opinions that the survey traffic is only C-end data, while AI is currently working on the B-end. The traffic on the B-side even exceeds half of the current total traffic, and it is rising rapidly. It's just that the research institutions did not get the complete data.

Is the flow a short-term fluctuation? Or long-term decline? Or is it really a different emphasis?

**Another example is the false enthusiasm. **

Zhang Ying, a partner of Matrix Partners, shared two data that are quite interesting in comparison. One is that from March to May this year, among the companies in the S&P 500 Index, executives of 110 companies mentioned AI at performance exchanges, which is three times the same type of data in the past ten years.

But another set of different data is that the international investment bank Morgan Stanley recently conducted a survey of more than 2,000 people, and it turned out that 80% of them have never used ChatGPT or Google’s Bard1.

Compared with this data, it feels like "Ye Gong loves dragons". Is the enthusiasm of these corporate executives, technology giants, and analysts fake?

** Or, the confusion of users. **

C-end users are very enthusiastic and sincere, but after using it, they have doubts: We want robots to help humans sweep the floor and wash dishes because humans want to write poetry and paint. As a result, AI is now writing poems and paintings, while we humans are still sweeping the floor and washing dishes.

Could it be that the "emergence" of artificial intelligence cannot be reflected in the real world?

Controversy over traffic, false enthusiasm, and confusion among users, these controversies also directly hit the core elements of the development of large models: How can ** really become a productive force? **

Different ways of answering will shape different AI development systems, and will also become a watershed for the future development of enterprises.

01 From algorithm to product

From the day of its birth, artificial intelligence has been strengthening its "two legs" to walk: one is technology and the other is application.

Behind the improvement of AI technology is the common support of the three major elements of computing power, data, and algorithms. For example, in terms of algorithms, artificial intelligence has successively gone through stages such as rules, statistical machine learning, deep learning, and pre-training, thereby greatly expanding the amount of data; and the "pioneering" algorithm Transformer, through the attention mechanism, allows AI to "do questions" Rapid training in a fast way, thus showing significant mutation and stronger self-learning ability.

Every leap in technology will bring surprises, but if you look back at the history of artificial intelligence development, you will find that after several surprises, there is loneliness.

For example, in the heat wave that emerged in 1956, artificial intelligence could play chess and catch building blocks, but in 1973, a report in the academic community concluded: So far, any discovery in this field has not produced the original promise. Significant impact2.

In 1976, the AI-based expert system began to participate in medical diagnosis and consultation. With renewed heat waves, governments around the world are stepping up investment. However, another ten years later, it was discovered that the machine experts did not show much talent. Doctors still have to go there in person, which is not enough.

Since 2016, Google AlphaGO has challenged many human chess kings in the Go world, winning 60 consecutive victories in 5 days. Even Li Shishi and Ke Jie can only surrender. People lament the power of artificial intelligence, but in the next five years, AI has not done anything amazing.

The turning point between the ups and downs is precisely the "product": whether there is a good product, let the technology step down from the altar and enter the society, and truly become the leader of productivity and creativity.

There are many cases where technology and products promote each other. For example, the failure case is Motorola's Iridium project, which provides global satellite communication services. Its technology is leading, but because its products are not grounded, it declared bankruptcy after four years of official operation. Successful cases include electric vehicles. Although batteries and electric drives are existing technologies, the market will gradually open up only after products with a strong sense of technology come out.

Going back to the field of large models, there is an interesting point in this round of AI tide: OpenAI released the world-famous ChatGPT, but it uses Google's Transformer algorithm for continuous optimization. This shows that in this round of large-scale model competitions and the wave of artificial intelligence, algorithms alone are not enough. Only the algorithm is weak; the single-point competition of the algorithm will eventually give way to the "product competition".

And OpenAI is not a "fascinated technology". Behind it is also the support of Microsoft's powerful "product system": search Bing, office family bucket, personal assistant, and advertising marketing and other enterprise-oriented cloud services.

This is like a "golden rule" in the venture capital circle: if the founder is a technology expert or a geek, then at the same time as giving money, he must also give it to a partner who understands the market. In this way, the technology is applauded, but the income is not popular.

Therefore, while attaching importance to technology, it also pays more attention to application guidance and product drive. Especially for large-scale enterprises, instead of telling partners, I have a lot of powerful AI capabilities, and you can use them however you want; perhaps, providing some product modules is closer to reality. So, a good product, how to do it?

02 From general to industry

Even if Zhuge Liang knew astronomy at the top and geography at the bottom, even if Da Vinci could draw and dissect and build airplanes, they could only be limited to the knowledge of that era. Artificial intelligence, on the other hand, can rely on a large number of inputs to greatly expand the boundaries of knowledge.

However, the wisdom of artificial intelligence is not perfect and universal. Judging from the experience of the past few months, AI will "seriously talk nonsense" from time to time. Maybe AI is not lying on purpose, but it certainly shows that the general model is still imperfect.

Especially when it comes to some specific fields, such as finance, education, etc., the limitations of general large models will be obvious. After all, there are always many areas where ginger is still hot, and know-how is key.

However, if the large model cannot enter the industry, the value will be greatly reduced. Especially for our country, which has a huge and rich industrial chain foundation, all vertical industries should be combined with new technologies to reduce costs, improve efficiency, and generate new value.

So, in the vertical field, only need to make a small model? the answer is negative. Small industry models can solve problems in specific fields, or they can do a good job, but there are two problems.

One is the lack of generalization. Once the scene is changed, it may be necessary to do it again, which will lead to a substantial increase in costs. One dish for each person, and the dishes are not repeated. If you open a restaurant like this, it will definitely go bankrupt. So limited intelligence is not intelligence.

On the other hand, during the application process, if the user suddenly asks some cross-domain questions, the small model will also be confused. Obviously, the trend of industry crossover is becoming more and more obvious, just like electric vehicles, which are both vehicles, batteries and semiconductors. Once you think from the user's perspective, even if it is a completely irrelevant field, you still hope to get a one-stop service.

Therefore, large models need to enter vertical industries, and vertical industries also need large models. How to do? An observation sample is JD.com.

In 2021, JD.com will take the lead in injecting domain knowledge into large models, which can increase the accuracy of the model from 83% to 96%. Just yesterday, JD.com launched a 100 billion-level Yanxi scale model for the industry. According to the introduction, 70% of its training data is general big data, and the other 30% is industry know-how data accumulated in the operation process of JD’s various sectors, including retail, logistics, health, finance and other industries .

Sure enough, adults don't make choices, but want both.

In fact, this is the right thing to do. This round of generative AI is very attractive, but also because the algorithm is strong, the data is rich, and the computing power is strong enough. And the largest model is not static, it is continuous learning. Therefore, data and algorithms form a "flywheel effect". As more and more good data are available, the algorithms will become more and more advanced; the more effective the algorithm, the more users there will be, and the more data feedback will be **.

Therefore, forming a "data-algorithm" closed loop as soon as possible is not only the path to product success, but also the key to enterprise competition.

In addition, high-quality data is also scarce. In the titled "Will we run out of data?" "The report shows that good-quality natural language data may be exhausted by large language models as soon as 2026. Whoever has good data will have better "ammunition". And good data, especially in the industrial field, must come from real industrial scenarios.

Therefore, the closed loop of "data-algorithm" is interpreted as a competition of "scene-product". And only by entering the scene can the large model move from "emergence of capability" to "emergence of value".

03 From native to empowerment

One way to realize the emergence of industrial value is to cooperate with the industry, technology companies provide technology, and the industry provides know-how. And the other way is also the best way, that is from the industry.

If you own your own industrial business, you will have real and valuable "high-quality data": you have suffered losses, stepped on thunder, fought battles, won battles, and know how to fight. These data, like a catalyst, can efficiently drive the development of large models, which are closer to the business and solve problems better.

A past case is the development of cloud services in China. Regardless of domestic or overseas, the initial stage of the cloud starts from the needs of the enterprise itself, and then it is market-oriented. At the beginning of cloud services, every "product" seems to be the same, I have what you have. However, with the combination of technology and business, each company has its own characteristics.

Take Jingdong as an example. JD.com started from “marketing, trading, warehousing, distribution, after-sales” and other businesses, but along with the step-by-step improvement of the physical supply chain network, the digitalization of the internal supply chain, and JD.com’s own retail, finance, logistics, health, With the deep-rooted development of industries and other fields, JD.com has gradually completed the expansion from "the last five segments of ** sugarcane" to "the first five segments of ": ** has platforms, scenes, AI, has experience.

Subsequently, JD.com refined its experience in the supply chain into "digital intelligence supply chain" products and services based on the technology of JD Cloud, and exported them to the society. As a result, the capabilities of digital infrastructure efficiency improvement, industrial synergy efficiency improvement, and urban intelligent management have been formed.

Today, there are more than 10 million self-operated product SKUs in the supply chain of Jingdong Shuzhi, serving more than 8 million active corporate customers, of which more than 90% are the world's top 500 companies in China and nearly 70% of the country's specialized small and medium-sized enterprises , and reached in-depth cooperation with more than 2,000 industrial belts across the country.

This kind of JD.com scenario with long links, complex collaboration, and more dynamic data backflow is the best training ground for large models, and it is also the best embodiment of industrial advantages.

The experience from internal cloud to external cloud is also being applied to the development of large models. Jingdong also proposed a "three-step approach" for large models:

Image source: JD Cloud

First of all, in July this year, the Yanxi large-scale model was launched, which has a four-layer system of base layer, model layer, MaaS, and SaaS. Secondly, "sharpen" in various internal business areas for half a year, and moderately conduct benchmarking cooperation with external partners, and go through multiple cycles of "mistakes, improvements, and conclusions" to achieve product integration. Finally, in the first half of 2024, for industrial output, we will use a better attitude and a more open ecology to serve the industry and improve the efficiency of the industry.

Internal applications have also been quite effective. For example, in the field of financial marketing, this is also the "old base" of JD.com. JD Finance has accumulated a wealth of knowledge through decades of business development, and combined with AI, it can efficiently optimize key tasks, dynamic adaptability, and user experience.

For example, reduce the learning cost and operating cost of operating personnel, and increase the production efficiency of the solution by hundreds of times; reduce the process that can only be completed by more than 5 types of functions such as product/R&D/algorithm/design/analyst to one person; at the same time, A new interactive mode of an entrance reduces the number of human-machine interactions from 2,000 to less than 50, and improves operational efficiency by more than 40 times.

The significant increase in the number also shows that although from the rhythm point of view, this three-step walk seems to be a bit slow. However, considering the input cost of the large model and the significant impact on the industry, only by adopting a step-by-step approach can it be transformed into a "step-by-step profit" to allow technology to generate benefits.

In other words, it is actually not slow, because it is not easy to truly achieve an industrial breakthrough. But just like the confidence of Xu Ran, CEO of Jingdong Group, cutting into the large-scale 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 and great exploration value. Only by thoroughly understanding the physical and digital supply chain can the big model empower the industry.

As the experience summarized by the Gartner curve, the development of things will go through stages such as "technical germination-expectation expansion-breaking valley-climbing recovery-production maturity". And to sum it up in another sentence: Don't treat the rhythm problem as a structural problem.

The development of technology is an inevitable trend. Driven by the three elements of "data, algorithm, and computing power", artificial intelligence will inevitably continue to develop; however, there will inevitably be some twists and turns during this period. What is needed is the scientific rhythm of the enterprise in technology research and development and application, as well as the long-termism that sees the trend and is willing to stick to it.

JD.com's persistence and breakthrough in the supply chain is a microcosm of the victory of long-termism. Now, in the big model competition, in the wave of artificial intelligence, the same is needed.

It can be firmly believed that although the technology is implemented at a pace, as long as it takes root in the industry, huge value will inevitably be born. As Xu Ran, CEO of Jingdong Group, said, when the industrial efficiency and the boundaries of the industry are expanded and qualitatively improved, the large model will have more important practical value and significance, which will be no less than another industrial revolution.

The formula of artificial intelligence is also deduced into "scenario, product, computing power group, and industrial thickness", which is the key to promoting the big model from "emergence of capability" to "emergence of value".

[1] Jingwei Zhang Ying: The Far and Near of AI, Chaos Academy, 2023;

[2] Lighthill Report, Science Research Council of Great Britain, 1973