Shulou(Shulou.com)11/24 Report--

Since OpenAI launched ChatGPT at the end of last year, large models have stepped onto the stage of a new round of productivity revolution as a new production tool. In fact, although the AI technology, which has experienced many ups and downs for decades, has been given high expectations, it always looks unsatisfactory when it hits the industrial end. Can the outbreak of the big model bring real change to the industrial side? What great changes will take place?

On the other hand, the digital transformation of enterprises has been an important strategy for the development of the whole industry in recent years, and the key link in the digital transformation is to achieve intelligence to the greatest extent. How to empower enterprises to transform digitally into digital intelligence through AI technology represented by large models?

With the above questions, the "TVP into Mengniu" closed-door exchange meeting jointly sponsored by Tencent Cloud TVP, Tencent Wisdom Retail and Mengniu Group was successfully held on September 22. More than 50 experts from Tencent, Mengniu, Zhongshun Jierou and other enterprises shared their views and practices on the transformation of digital intelligence in enterprises. In addition, participants from different industries also expressed their views on some hot topics in the final group discussion session.

The meeting was presided over by Mr. Tang Xinlong, founder of the "Digital China" project of Wudaokou School of Finance of Tsinghua University and Tencent Cloud TVP industry ambassador: "Today, the technological development of the big model is becoming more and more hot. In addition to the endless emergence of C-end applications, the combination of B-end big models and various industries is also getting closer and closer. In the following sharing and discussion, I hope to hear voices from various industries to share and discuss freely and deeply from the perspective of industry and technology. "

The universal AI driver will be an effective grip.

In the opening session, Cheng Wei, general manager of Tencent Smart Retail Technology Architecture and delivery, said that the reason why Tencent Cloud TVP entered Mengniu is mainly because Mengniu is in the forefront of the entire AI digital industry, including the company's desire to upgrade from a dairy manufacturer to everyone's health consultant, a transformation that needs to be driven by AI.

In fact, the retail industry is facing three major integration trends: online and offline integration, brand and channel integration, and product power integration. Combined with the more obvious social trend in the future, the traffic and activity of Wechat, Mini Program and video accounts continue to grow. In such a diversified scenario, what new challenges will retail enterprises face in their digital transformation?

Cheng Wei believes that AI drivers and the digitization of the industry will be a very good starting point: "in the development of the retail industry in the next five years, there will be a deterministic key point, that is, the content of marketing and how to reach it is very important. These access methods no longer rely on personalized technology and product developers, general-purpose AI drivers will be very effective."

100 people, 1 million tons of production capacity, 10 billion output value

The second guest is Zhang Jie, Assistant Vice President of Mengniu Group and Tencent Cloud TVP industry ambassador. According to her, Mengniu currently has 68 factories around the world, eight of which have completed digital construction. It is worth mentioning that the Ningxia Lingwu factory, which was just completed and put into operation in May this year, has achieved the pure digital factory goal of "three 100s": 1 million tons of production capacity and 10 billion output value.

In Mengniu's "FIRST Development Strategy", "T" is Technology. From the beginning of its establishment, the company has attached great importance to information construction, from OA to SAP, and then to LIMS, have carried out a systematic layout.

Specifically, 2016 is the first year for Mengniu to start digital construction. The company first took three years to carry out information upgrading, including financial sharing and intelligent manufacturing (digital intelligence chemical plant), a period known as the 1.0 stage of digital intelligent governance of "one drop of milk".

Starting from 2020, it has entered the construction stage of 2.0 and 3.0. Among them, the evaluation criteria of 2.0 include the construction of business and data stations, as well as management digitization, capacity internalization and so on. Era uses AI to drive twin flywheels (supply-side flywheel + consumer-side flywheel). On August 4 this year, Ms. Li Mianjie, CDO of Mengniu Group, launched the world's first nutrition and health model MENGNIU.GPT on behalf of Mengniu, allowing data assets to continue to play an efficient insight and decision-making role.

"by working with several big models, including Tencent Mianyuan, we have completed some model-based product construction internally. For example, in the data layer, we have formed Mengniu's nutrition and health knowledge graph, which can already have a dialogue on Wow Mini Program. At the same time, it also opens the construction of domain model, including management, marketing, supply chain and other sub-areas. In the improvement of employee productivity, Mengniu has also opened its own Copilot. " Zhang Jie said.

Data + AI + Big Model: building Integrated Intelligent Marketing

Then, Yu Shuai, deputy general manager of Tencent Enterprise Point, brought the sharing of the theme "Analysis of Smart applications in the Retail Industry" to the guests at the scene.

Since the birth of the retail industry, its development has experienced three stages: physical stores, platform e-commerce, and then the rise of brand social. Nowadays, more and more online commerce platforms have made good marketing performance by using the promotion of social software such as Wechat. It has to be said that the historical evolution of the retail industry has witnessed how data move towards digital change step by step in the process of precipitation and application.

"from data precipitation to digitization is a gratifying natural trend, but we are still faced with a big question, that is, what value can data play in the real construction of digitization? what effect can it bring in the operation of social e-commerce or social private domain? there is no standard answer to this question." Yu Shuai said.

In his view, on the one hand, data as a means of production, a large number of precipitation is very important; on the other hand, the validity of data needs to be mined through AI. In other words, data is the "growth engine" and AI is the "innovation engine". So, when it comes to specific marketing scenarios, how can AI improve efficiency and what forms of business innovation can it bring?

It is important to solve the problem of data and access being separated from each other. In this regard, Yu Shuai first used four "appropriate" to explain what is meant by "accurate recommendation": "it is to push the right content to the right person at the right time and through the right channel." However, the reality is not so ideal, facing multiple fragmentation: the separation of data from different products and modules, and the separation of identities in different channels of users.

In order to solve this problem, Tencent provides a "two-oriented" solution: data integration and data scene. First of all, in the construction of data integration, integrate "two platforms, 5A, 5m". The two platforms refer to the bottom CDP customer data platform and the top AB test platform. 5A and 5m are analysis application and marketing application respectively; secondly, the data scene is to integrate the multi-contact data of the brand with the help of big data and AI technology, so as to accurately identify the user identity and journey stage, and provide differentiated marketing content and strategies in different scenarios.

In addition, through the application of the big model, Tencent is also further improving the customer experience and improving the marketing transformation effect. Yu Shuai believes that "our business analysis is very dependent on data processors. In the case of uneven human abilities, we are likely to draw conclusions or deviations in efficiency and direction." it is hoped that through our large model, when sufficient corpus is given, we can provide users with relatively qualified answers. " At the same time, Yu Shuai also pointed out that just the answer is still not enough. How to realize the interaction with several Homo sapiens, and even generate stylized UGC creative content through large models is also the focus of Tencent's future exploration.

To grasp the opportunity of new technology, the future is the competition between industrial chain and ecological circle.

In the interpretation of the theme of "Mengniu's AI Innovation Exploration", Mengniu Open Innovation Director Gao Jiulin first introduced the core architecture of Mengniu's digital intelligence strategy 3.0--AI driving twin flywheels. Based on AI, a series of applications such as Niu Huiwen, Niu Huixiu and Niu Devil King have been created in Mengniu. Employees who have been trained and assessed by prompt engineers have built more than 400 scene cards in the AI scene factory.

Externally, Mengniu released a nutritional health model MENGNIU.GPT and AI dietitian Meng Meng this year. And combined with Tencent Digital Homo sapiens technology to generate a vivid 3D cartoon image to provide nutrition and health consulting services to customers.

By reviewing the process of digital intelligence innovation and exploration, Gao Yulin tries to answer such a question: how can we seize the opportunity when new technologies emerge?

First of all, we need industry insight: to face up to the sharp increase in consumer demand for nutrition and health and the lack of nutrition and health knowledge supply. Sustained economic growth has promoted people's demand for nutrition and health, while the supply of relevant knowledge is single and insufficient. For example, the supply gap of Chinese dieticians is about 4 million. The emergence of AIGC technology provides a new possibility to break this imbalance.

Second, grasp the technological trend: once again, we stand at the turning point in history and will witness and experience a new round of "all industries are worth doing again". The large language model will become the underlying new infrastructure (MaaS), but it needs to be trained in professional areas to be of greater value.

"on the one hand, the big voice model will become a new infrastructure in the future; on the other hand, in order to really land in the professional field, it is not enough to rely on the original ability of the big language model, and more professional training is needed in the industry knowledge. only in this way can we give play to greater value."

On the basis of this judgment, Mengniu firmly chose to embrace AI in an all-round way, and joined hands with domestic and foreign technology enterprises such as Microsoft, AI, Tencent, Ali and other domestic and foreign technology enterprises to train the tuning model and integrate many algorithms in the market. And joined a group of experts and scholars in the field of nutrition and health, together with a number of nutrition and health authorities, to inject high-quality content as training materials. In addition, Mengniu also opens up the MENGNIU.GPT capabilities of nutrition and health models to explore and create scenes freely with more innovative ecological partners.

Finally, Gao Guilin introduces the path choices of different innovation from the two dimensions of closed innovation and open innovation. In his view, closed innovation and open innovation are two different paradigms: the model of closed innovation, such as Bell Lab, can produce great creations that affect human civilization, while open innovation is more active and ecological. it is related to the competition of industrial chain and ecological circle, and it is the innovation paradigm of many large international enterprises.

"the industrial chain and the supply chain ultimately determine the extent to which the product can be achieved, while the ecosystem is a higher-dimensional way of competition. I hope to create such an ecology for Mengniu through open innovation, link more possibilities on the basis of MENGNIU.GPT, and jointly create a new experience of mathematical intelligence, nutrition and health."

From the cost center to the profit center, the future is the investment center.

Yang Lin, CTO of Zhongshun Jierou, shared some different views on "how to layout enterprises in the AI era", including the application of algorithms and large models in the retail industry: "first of all, I think that in terms of the current environment, it is not the AI that changes this era, but the people who control the AI. From a technical point of view, we in China actually do not have our own technology platform, nor do we have our own technology language and operating system. That is to say, we have always been the users or assemblers of technology, and this is where we really are. So, overly complex technologies, like AI, might not be on the enterprise's agenda if it wasn't for the ChatGPT fire. "

When it comes to the current situation of enterprise digital transformation, Yang Lin's view is that the industry as a whole is not doing well. He showed participants a McCarthy report: the core data is that the success rate of corporate digital transformation is only 20 per cent. "in fact, 20 per cent is overrated, and we still need to clearly position ourselves as to what we want to do and what we can do by using AI."

So, what are the key factors in the digital transformation of enterprises? It mainly lies in three points: system, process and people.

System: no system can be used optimally in the first edition. Continuous optimization and iteration is the key to building all kinds of systems.

Process: no matter how good the system is, it can only manage 50% of the process, 50% of the process, 50% of the process, and still 20% of the process, communication, and decision-making are made outside the system. Effective and efficient docking of offline and online processes is one of the core challenges of the project.

People: even with good systems and procedures, there is a lack of people who can be used efficiently, just like the J-20 aircraft is difficult to take off if ordinary people are allowed to fly it. The ideological transformation of the users of systems and processes, and how to get out of the inherent thinking and embrace new methods is the most key factor for the success of enterprise digital transformation.

As for how to enable digitization through AI, his view is that it is better to harness AI than to embrace AI. According to its observation, the current value output of AI mainly lies in the following points:

Cost reduction: including optimization algorithms, distributed training acceleration and model compression

Improve ease of use: reduce the threshold for users by completing intuitive and easy-to-use user interface design and building simple and easy-to-use development tools and platforms

Security explainable: can improve data quality, enhance robustness, and achieve continuous monitoring and upgrade

Data security: data encryption, access control and identity authentication, as well as security audit and monitoring.

Based on the key role of "people" in the digital transformation of AI, Yang Lin believes that as the leaders of enterprise digitalization, CTO, CDO and CIO should become "hexagonal warriors" and need to understand strategy, business, technology, data, process and innovation.

"if we don't understand the underlying business, we will find that a lot of plans, whether they are models, algorithms, or systems, will float in the sky and cannot fall to the ground. In many cases, our imaginary business scenarios are completely different from the real ones. During the digital transformation, we should always focus on the situation of the enterprise itself, and complete the transformation of the digital department from a cost center to a profit center. It is possible to further become an investment center in the future. "

The driving force of technological architecture evolution: production tools, productivity and production efficiency

The last guest to share is Lu Xiao, director of data products at Tencent Cloud. The theme of his speech is "the evolution and exploration of Tencent's new generation big data technology architecture."

At the beginning of the sharing, he first introduced the driving force of the evolution of the technology architecture, which mainly lies in three points: the safety, stability, flexibility and convenience of production tools, the advanced nature of technology and low-cost use in productivity, and the need to ensure high efficiency and ease of use in production efficiency.

"the development of Tencent big data technology has gone through several stages. in the first stage, Tencent solved the demands of large-scale parallel computing and expansion of many online businesses with large-scale distributed technology in the past few years. After moving from PC to mobile, we launched user profiling, commodity recommendation and other businesses through AI capabilities, forming a variety of models, this is the second stage."

Later, when the team found that the traditional AI technology could not meet the business requirements in terms of computational efficiency and accuracy, they began to introduce machine learning to quickly build the framework and system of Tencent Cloud through deep learning and large models, which could support millions of model iterations per day, thus completing user portraits more efficiently and accurately.

"at present, we have entered the fourth stage, and we are more likely to apply cloud, big data, data lake warehouse, and even some cross-source, collaborative computing, multilateral computing, secure computing, privacy computing and other technologies to a variety of enterprise scenarios, constantly carrying out technological innovation."

Since 2009, Tencent self-developed big data platform TBDS began to evolve iteratively, from version 3.0 for commercial use, to version 4.0 for large-scale breakthrough, to version 5.0 to complete the integration of streaming and approval, integration of lakes and warehouses and real-time data lakes. Today, the platform iterates to the latest version of TBDS 5.3, which enables the technical architecture upgrade of cloud native and storage separation.

Specifically, TBDS 5.3 has been upgraded in a number of ways compared to previous versions:

Flexible improvement of architecture-separation of deposit and calculation: compared with the traditional integration of deposit and calculation, the separation of deposit and calculation can split the data, realize stateless computing, and achieve second-level flexibility and flexible scalability.

Improve production efficiency-- Yunyuan biochemistry: achieve flexible scaling of resources, improve resource utilization, carry out effective resource isolation, achieve agile and efficient deployment and management, and unify production, development, testing and other application environments.

Productivity improvement-lake-warehouse integration: open agility, flexible expansion, mixed load, cost saving

Upgrade of production tools-data development and governance platform Wedata: achieve full link coverage, efficiency improvement and multi-team collaboration.

"(Wedata) from data collection to data development, the data standards of enterprises landing in the process of data modeling and processing, at the same time, we will inspect the quality of all the data to form uninterrupted data quality feedback and improvement before, during and after the event. Ultimately, it will help us quickly reshape our business for different goals, different organizations, and data assets, which is the complete set of data lifecycle capabilities we currently provide. "

Under the current wave of localization, Tencent has completed considerable precipitation in terms of independent control and technological leadership, including applying for patents in the field of big data, and big data's soft works have also reached 70 +. In terms of industry impact and ecological construction, it fully adapts 20 kinds of domestic chips, OS and servers, and participates in the formulation of 30 national / industry / group standards. In the construction of Xinchuang, according to Lu Xiao: "both government affairs and enterprises and institutions have very high requirements for Xinchuang, and we have landed on a large scale in very large units such as the State Administration of Taxation and the Bank of China."

Group discussion

In order to enable the participants to participate in the discussion on the topic of "digital transformation and AI empowerment", after the host guest speech, host Tang Xinlong also organized a group discussion session for the on-site guests. The guests were divided into four groups, discussed in depth on different issues, and combined with their own business practice, summed up different views.

How will the big model develop in the future? What are the expectations?

Guests at the scene believed that the future large model will be divided into three layers: the top layer is a technology company like Tencent, which provides the most universal basic general large model capability; the middle layer is an industry leading company like Mengniu, which provides an industry-specific model based on industry accumulation; finally, some key enterprises in the industry will combine their own data and needs to form the final landing model. It is the basic model, the industry model and the enterprise's own model. It is also hoped that more industries can standardize the model in the future.

A hundred flowers blossom in the big model, will this technology have a far-reaching impact on the whole industry in the future?

Guests at the scene pointed out that after ChatGPT became popular at the beginning of this year, many enterprises began to try AIGC, and found that in fact, people's expectations of generative AI had undergone a curved change. At first, they thought that it could do anything and could replace almost all positions, but in practice, this idea quickly encountered a bottleneck, although it did promote productivity. But if you can say which position can be replaced is still a long way off, so it will slowly calm down. It is necessary for more enterprises to explore how to do better in the private large model in the future.

As a service enterprise in the field of large health, what new requirements have been put forward for the acquisition and application of data in the new era?

Guests at the scene believed that in the era of large models, the key to the digitization of large health service enterprises lies in data, algorithms and computing power. At present, domestic enterprises are not poor in algorithms, and the gap between domestic and foreign enterprises is mainly reflected in data and computing power. Tencent Cloud can provide us with computing support, but we still need to solve it on our own in the data layer.

So, what kind of data do we need in the pre-training process? At present, most of the data are result-oriented, and the intermediate process data is still missing, but in fact, the intermediate inference data is very important to our training, and it is also the direction that enterprises should focus on in the future.

What are the difficulties in landing the AI / big model?

In response to this problem, guests at the scene pointed out that we have a lot of fantasies about AI, but we still face two major difficulties in order to really land.

The first difficulty is the uncertainty of AI's accuracy. We have been talking about how powerful ChatGPT is, but it mainly lies in the ability that it does not involve accuracy. For example, a composition will not say whether it is correct, only good or bad, it is receptive. We don't care whether AI is accurate or not, but if we are not sure whether it is correct or not, it is a big pain point.

We have talked a lot about the second difficulty just now, that is, there is still a lack of industry models. In order to make an industry model, we still need precipitation. Only when precipitation has accumulated enough technology and data, can we rise abruptly based on accumulated strength and usher in a new round of higher growth.

Conclusion

The new round of development in the era of artificial intelligence has brought another growth impetus to the digital transformation of traditional enterprises. How to correctly treat the application of large model technology, which will bring infinite possibilities for enterprises in the future, and how to make better use of large model technology to keep up with technological development and promote the efficient growth of enterprises? this will be the key for every enterprise technology manager to make in-depth research and analysis in the future.

Since its inception, Tencent Cloud TVP has been carrying out the original intention and original heart of Tech for Good with the bright vision of "influencing the world with science and technology", hoping to embody the practice and thinking of more experts and promote the digital construction of various industries to a new height.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.