Dialogue with Fante Technology: Large model services in the future will be subscription-based payment models | Annual AI Dialogue

Annual AI Dialogue

文|Yang Jun

Editor |Shi Yaqiong

Mastering artificial intelligence means mastering the productivity code of the new era.

As ChatGPT explodes in popularity, large models are popular all over the world, and Chinese industries have also stimulated new enthusiasm for artificial intelligence applications.

Therefore, at this point in time, the 36 Krypton Digital Time Krypton team officially launched the "Annual AI Dialogue" column, hoping to discuss hot topics with domestic professionals who have in-depth research in the field of large models, and restore AI through a series of reports New technical capabilities and application potential.

In this issue, we invited Wu Shengyang, CEO of Fante Technology.

What impact does the emergence of large models have on China’s digitalization process? What impact will it have on corporate digital budgets? In addition, will large models change the business models of Chinese companies? These are all issues worthy of attention.

Fant Technology is an enterprise digital transformation and metaverse infrastructure provider, focusing on the research of enterprise digital transformation and artificial intelligence. Wu Shengyang, CEO of Fant Technology, believes that Digital China is actually closely related to the Metaverse. In the context of the new era where models are productivity, every city will have a large underlying city-level model; in the future, companies will pay like SaaS now. The model is the same, with large model capabilities customized on demand from large model service providers; after the emergence of large models, companies that are deeply integrated with new technologies will replace traditional integrators and redefine and divide business scenarios.

The following content is the original text of the interview, edited and compiled by Digital Times:

Big Model and Enterprise Digital Transformation

Large-scale models will be applied in actual production and operations of enterprises, especially at the digital level, which can accelerate the implementation of digital solutions. Fante Technology believes that on both the B and G sides, enterprises and government departments can use the end-to-end capabilities of large models to customize large models in their own subdivisions, allowing various solutions to be quickly implemented.

Digital Times: What impact will the emergence of big models have on China’s digital transformation process?

Fante Technology: Digital transformation first solves the problem of infrastructure construction, and the second is large-scale capacity building. In the past, capacity building required a lot of manpower, material resources and time costs, because the development and application of algorithms The development cycle is relatively long. Forming a team, training a group of people, and accomplishing one thing requires a lot of investment, a long cycle, and slow results. Big models promise to change that. The emergence of large models has subverted the traditional algorithm research and development and application development models from the bottom up. A large number of manually designed algorithms have been directly replaced by large models, and application development has been greatly simplified and even automated. Enterprises with large models are equivalent to hiring an additional algorithm team and application development team, which not only reduces costs but also greatly improves efficiency. Some people would say that the threshold for large models is very high and ordinary companies may not be able to master it. This is actually not a problem. Companies like Fant, which have been deeply involved in vertical fields for many years, can help customers create and implement large models. Fant's new generation artificial intelligence MaaS platform integrates large model-as-a-service capabilities, algorithm self-training capabilities, application customization capabilities, and deployment automation capabilities to solve the implementation problems of large models end-to-end and facilitate digital transformation.

We once made a comparison chart. Digital China is actually closely related to the Metaverse. In cities like Zhejiang and Chongqing, we are already witnessing the booming development of Digital China. The Digital China architecture in these two places has a large underlying platform at the government level, which can be simply understood as a general capability. Based on this general capability, each business unit is empowered to develop small capabilities. This is actually a natural combination with large models. In the context of the new era where models are productivity, I believe that in the near future, every city will have a large underlying city-level model. On this basis, everyone can customize models for various vertical fields or industry applications. . On the C side, residents’ daily necessities, food, housing, and transportation will become more convenient based on deep digitalization. On the B-side and G-side, enterprises and government departments can use the end-to-end capabilities of large models to customize large models in their own subdivisions, allowing various solutions to be implemented quickly. These will contribute to the early realization of Digital China’s grand blueprint.

Digital Time Krypton: After the emergence of big models, what changes have occurred in the digital budget of enterprises?

Fante Technology:In the short term, it is foreseeable that the price and hardware investment of large models, especially privatized large models, will be higher. Moreover, the application of large models still requires opening up internal processes, standards, etc., and integrating with the original business, which is a considerable expenditure. Therefore, in the short term, corporate digitalization budgets may increase.

In the long term, after the company has fully digested the fixed costs of the large model and the docking costs of its own business integration, the overall cost of corporate operations and innovation may be reduced due to staff streamlining and efficiency improvement, but it has tasted the benefits of the large model. Sweet companies will definitely invest more energy in large models and innovations based on large models to maintain their own advantages and competitiveness, so we believe that the MaaS model will rise.

In the future, enterprises will customize large model capabilities on demand from large model service providers like Fante, just like the current SaaS payment model, and pay corresponding service fees on demand. Business budgets will be significantly relieved, and the ratio of expenses to income will be more accurately matched and controlled.

Digital Times: What are the difficulties in implementing large models in digitalization?

Fante Technology: The first direct difficulty is the computing power and prompt word engineering issues. Computing power is essentially an infrastructure issue. Not only for large models, but for any digital project, infrastructure is a big problem. Large models have relatively large requirements for the memory and quantity of graphics cards. Currently, this requires a certain amount of invested.

Prompt word engineering is a problem unique to large models. After the large model is trained, how to use this large model is very particular. Bad input will get bad results. Therefore, it is particularly important to do a good job in prompt word engineering in landing applications. At present, there is a shortage of engineers in prompt word engineering, and both academia and industry are constantly exploring. In other words, large models have come out, but the understanding of large models is still being strengthened step by step.

I talked about the direct difficulties earlier. What needs to be added is that digitization is a systematic project, and large models are the underlying infrastructure. Large models solve the problem of algorithm innovation, but there is still a lot to do between algorithms and applications. For example, although large models are effective, they are difficult to run directly on edge devices. In many cases, large models need to be used to supervise and train smaller models in vertical fields that are lighter and more efficient and run on edge computing devices. On the other hand, the algorithm is only the core module of the application. In addition to the algorithm itself, the application also has a large amount of business logic that needs to be implemented.

Finally, with algorithms and applications, they must be easily deployed on a large number of devices, preferably with automated deployment and operation and maintenance. In order to truly solve the problem of large model implementation end-to-end, a new generation artificial intelligence service platform is needed that integrates large model-as-a-service capabilities, algorithm self-training capabilities, application customization capabilities, and deployment automation capabilities, which is the so-called MaaS platform. Fant has created such a MaaS platform. The bottom layer is the base of the large model, the middle layer is the customization and automated deployment capabilities of AI, AR, algorithms and applications, and the upper layer is specific applications for vertical industries. We believe that only in this way can we truly solve the problem of digital implementation of large models.

Large models bring more opportunities for Chinese companies to “break the circle”

The emergence of large models has brought greater opportunities to Chinese industries and Chinese enterprises. Fant Technology said that in the ToB and ToG fields, the original business territory was basically divided among major integrators, and most of the technology manufacturers or algorithm manufacturers were integrated. With the emergence of large-scale models, companies that can deeply integrate new technologies into their products will replace the old-school system integrators and redefine and occupy the market.

Digital Times: What impact will the emergence of large models have on Chinese industry?

Fante Technology: The training of large models requires data, algorithms and computing power. The current situation is that there is a lack of high-quality training data in China, and a large amount of raw data has not been sorted out and is of low quality; at the algorithm level It’s not too backward, but the graphics card resources used to train the algorithm are very limited. At present, the leading AI companies have basically entered the large model track, and many startups are also involved. Overall, they are currently in the catching-up stage. I believe that as time goes by, we will be able to solve the data and computing power problems and truly have originality. The large model that understands Chinese is comparable to GPT4.

The emergence of large models will accelerate the upgrading of China's industry. On the one hand, the current situation has made us realize that we still have many weak links in the field of technological innovation, and it is very easy to get stuck. Then the industry will definitely invest more in data and AI chips. Upgrading large-scale models is a productivity improvement, and the improvement of bottom-level productivity will inevitably bring about the progress of the superstructure. The emergence of large models will provide more opportunities and motivation for China's scientific and technological innovation.

Through the capabilities of large models, Chinese companies can conduct more in-depth research and innovation in various fields, promote the development of science and technology, and achieve breakthroughs in fields such as artificial intelligence and machine learning. China's manufacturing industry can use large models to predict and optimize production plans, improve supply chain management, and achieve higher levels of automated and intelligent production to improve competitiveness. China's hotel, tourism, finance, medical and other service industries can use large models to improve customer service and experience, provide more intelligent and personalized services, and enhance competitiveness. As the pace of innovation continues to accelerate, a large number of simple repetitive tasks are replaced by AI, and the emergence of new engineers with AI skills will further promote innovation and technological development. In one sentence, large models will bring about huge upgrades in productivity and greatly promote the digitalization process of various industries. In this wave of technological revolution, if our country can seize the opportunity, it will be possible to further narrow the gap with technological powers. , comprehensively enhance international competitiveness.

Digital Times: What impact will the emergence of large models have on the business models of Chinese companies?

Fante Technology: In the fields of ToB and ToG, the original commercial territory is basically divided up by major integrators, and most of the technology manufacturers or algorithm manufacturers are integrated. After the emergence of big models, companies that are deeply integrated with new technologies will replace traditional integrators and redefine and carve up scenarios. Under this trend, all companies have the opportunity to break out of the circle. Companies that cannot keep up with the times will be eliminated. Internet companies bloomed more than ten years ago, and the history of subverting traditional industries will be repeated again. What is competing now is the understanding of new technologies and new things, as well as the ability to quickly adapt and iterate in the new era.

In the field of ToC, large models may also bring about changes in the level of traffic entry. Giants like BATJ that have almost achieved traffic monopoly in the past, and even world-class hegemons like Google, may be replaced in the new era. Subversion, no company can say that it now has a moat. For entrepreneurs who dare to innovate, this may be a once-in-a-decade or even a century opportunity. The Eight Immortals have shown their magical powers across the sea, and MidJourney is by no means the only company with eleven employees to achieve profits of over 100 million years.

Digital Times: What impact will the emergence of large models have on specific industries?

Fante Technology: The big model is a change in the underlying productivity, and its impact is on the entire industry. Take the financial industry, which we are deeply involved in, as an example. On the day ChatGPT came out, we received inquiries from bank customers, which shows that the industry is particularly keen on the innovation brought by large models. Large models are in strong demand in specific businesses such as bank customer service, financial advice, robo-advisory, fraud detection, and digital intelligence. However, due to policy reasons, ChatGPT and Internet-based language large models cannot be used directly, involving users. Privacy and other issues. At present, domestic leading companies in the field of basic large models have successively launched their own large language models, but there is still a big gap compared with ChatGPT, and it is difficult to achieve large-scale product launch in the short term.

The R&D investment in basic large models is huge and not friendly enough for startups; of course, with the development of technology, various open source large models have flourished, and new large models appear almost every day. Many models have The effect has also reached the level of more than 90% of GPT-4. The emergence of these models has made it possible for startups to have their own large models, especially large models for vertical fields. We are currently following up in this direction, thinking that Customers mainly provide customized optimization services. We focus more on large models in vertical areas of the industry. Because the Fante team has been deeply involved in vertical fields such as finance, governance, emergency, and cultural tourism for many years, and has a group of tireless and continuous innovators, we not only know what kind of large models are needed in vertical fields, but also know how to implement large models. In a word, we will use large model capabilities, or MaaS capabilities, to help our industry customers have the ability to build vertical AI models that are most suitable for their fields.

Vertical large models are opportunities for artificial intelligence startups

As a startup company in the field of AI, Fante Technology adjusted its role to a MaaS service provider after the emergence of large models. It focuses more on the technical level rather than focusing on everything from business to bottom layer. Fant Technology also said that in China, due to cost and talent reasons, there are actually no more than three companies that can make basic large-scale models. With Fant's current size, it will not enter the competition in this field. The vertical large model is a real opportunity for artificial intelligence startups like Fant to stand out in the new era.

Digital Time Krypton: How do you view the opportunities for startups to develop large models in vertical fields?

Fante Technology:As Lu Qi said, only China and the United States can do basic large-scale models in the world. In China, due to cost and talent reasons, no one can do basic large-scale models. In fact, there will not be more than three companies. With Fant's current size, it will not enter the competition in this field. The vertical large model is a real opportunity for artificial intelligence startups like Fant to stand out in the new era. Large models in vertical fields are usually at the tens of billions level (basic large models are generally at the hundreds of billions level), and the goal is to solve problems in vertical fields. The vertical domain large model is based on a mature and stable basic large model base, and the capabilities are combined and optimized for specific problems in the vertical domain, and can solve practical problems quickly and efficiently at low cost.

What kind of company can build a large vertical model? In our opinion, to have both the ability to accumulate technology and understand the industry, it is still the saying "the one who captures the scene is king". This sentence is also applicable in the field of vertical large models. To make large models in vertical fields, startups must first have sufficient knowledge of the industry. They must not only fully master the capabilities of large models but also know where the pain points of the industry are. Therefore, large models must be built in vertical fields that they are familiar with. In order to realize the implementation of large-scale models, startups need to have strong engineering implementation capabilities, which requires the blacksmiths themselves to have sufficient hard power. The ability to implement this project is reflected in multiple levels such as algorithm, application, deployment, operation and maintenance. It is a systematic project and does not happen overnight. Fant has been solving the above problems since the first day it was founded. We started from the smallest inference unit and continued iteration, gradually forming our own development and operation system. Only in this way can we implement high-efficiency and high-quality operations. Large model applications are delivered to customers.

Digital Times: What impact will this wave of large models have on Fante Technology?

Fante Technology:Large models, for all companies, including AI manufacturers like Fante, first bring tension. The terrifying capabilities of large models are based on our previous cognition and AI capabilities. With the impact of dimensionality reduction, every company needs to reposition and develop itself. But soon our more emotions turned to surprise. At present, almost all AI startups are put on the same starting line, and this time we may outperform everyone.

We have been paying close attention to the development of large models internally since GPT3 in 2020. The explosion of ChatGPT in February this year allowed us to see the opportunities of the large model track. We have also organized internal meetings for AIGC and large models. At the discussion, everyone was unanimously optimistic about the development of large models. They also agreed that using the vertical field as a breakthrough to enter the large model track is the best choice at present, and they quickly started research and development work. We have adjusted the R&D organizational structure, established a large model research group headed by the chief scientist, adjusted the R&D direction, and focused on strengthening the R&D capabilities of large models in vertical fields and the integration of large model capabilities in each product line. It can be said that large models have changed the way AI companies operate themselves, changed the production and design ideas of algorithms and applications, and broadened the capabilities of AI implementation.

In the process of serving our customers, our positioning has also changed a lot. Before the advent of large models, we need to collect data on-site according to user needs, train algorithms, implement engineering implementation and connect user systems. We need to provide a complete set of solutions. From a manufacturer's perspective, if the solution does not If it can be replicated on a large scale, the input-output ratio is not high. With the large model, our role becomes a MaaS service provider. After selecting the most appropriate large model for the underlying algorithm, we provide users with the ability to customize their own algorithms, so that Fant can focus more on the technical level. Not everyone from the business to the bottom level needs to be focused. Customers can use large models to continuously innovate business models and take advantage of their familiarity with the business. To give a simple example, if there is another battle over shared bicycles, we don’t care whether Ofo or Mobike will win. We are only responsible for making good bicycles and selling them.

Digital Times: What are Fante Technology’s current plans? What are you doing?

Fante Technology: Lu Qi mentioned in his speech some time ago that the era that mankind is currently in is the beginning of the second generation system, that is, it has just begun to complete the stage of transformation from information to model. Fante The MaaS platform is committed to efficient empowerment at this stage. In the future third-generation system, human society will not only be information-model, but will be transformed into information-model-action. Spatial computing, combination of virtual and real, and Web3 are all inevitable trends in the future, and this part is exactly in line with Fant. The future layout is a perfect match, that is, using the capabilities of AR on the basis of the MaaS platform to realize the connection between model and action capabilities.

From the perspective of short-term goals, Fant is more focused on two things. One is the construction of MaaS capabilities, and the other is using the AR capabilities we already have to complete the last mile of the scene. In terms of MaaS capabilities, Fant has completed technical pre-research on large language models, large visual models, and large multi-modal models, and has mastered the full-process customization capabilities of training, tuning, and inference. And it has applied large model technology in the process of product development, completed the integration and upgrade of large model capabilities for products such as AI rapid development platforms, and will launch related products in the near future. During the process of technical pre-research, we also discovered the shortcomings and flaws of general large models in vertical subdivision fields. Next, we will optimize and iterate large models to address these issues to make large models more versatile and easy to use in vertical fields. There will be more applications based on large models in finance, governance, emergency and cultural tourism.

At the same time, we also have a prediction. As mentioned above, after the arrival of the big model, all scenes will be reshaped under the new productivity revolution. That is, all the rules in the past will be broken. So how to use technology to quickly occupy The commanding heights of the scene will become the key to commercialization. AR is a direction we are optimistic about and insist on. First of all, the evolution of various products from 2D to 3D is in line with the needs of human nature. When informatization, digitization, RPA and other technological means liberate people's hands, large models will gradually liberate people's brains. Then humans will most likely seek to enter the metaverse to experience more parallel worlds and enjoy different things. life.

Scan the code to get more content

This article comes from the WeChat public account "Digital 36kr" (ID: digital36kr), author: Yang Jun, 36kr is published with authorization.

The above is the detailed content of Dialogue with Fante Technology: Large model services in the future will be subscription-based payment models | Annual AI Dialogue. For more information, please follow other related articles on the PHP Chinese website!