Home About us

Clear the fog and find the fulcrum for the application of large models

Insight into the New Research Institute 2024/07/09 17:26

<em> WeChat </em> picture _20240709163249.pngAuthor | Chenwen

Source | Insight into the New Research Institute

The temperature in Shanghai is very high, close to 40 degrees, and what is even hotter than Shanghai is the heat of AI.

On July 4, the 2024 World Artificial Intelligence Conference and High-level Conference on Global Governance of Artificial Intelligence (WAIC 2024) opened at the Shanghai World Expo Center, with more than 500 exhibitors and more than 1,500 exhibits.

At last year's WAIC, the focus of the industry was "whether it can be done", and this year, the wind direction has changed to "whether it can be used".

Among them, leading manufacturers such as iFLYTEK Xinghuo, Baidu Wenxin Yiyan, Ali Tongyi Qianwen, Tencent Yuanbao, Huawei Pangu and other leading manufacturers competed on the same stage, and startups such as Baichuan Intelligence, Minimax, Zhipu AI, and Step Xingchen also brought large-scale model-related products and the latest applications.

The scene is very lively, and the application of large models is also showing a situation of blooming everywhere, but it can be found that most of the landing of each company is still in the early trial and pilot stage, not only the homogenization phenomenon is serious, but also the lack of a relatively clear business path.

From "usable" to really easy to use and practical, until advanced to love to use, there is still a gap that needs to be crossed, and this gap is also the fulcrum of the application of large models that the industry has been exploring.

01

How to break the game and apply the large model

Liang Zheng, vice president of the Institute of Artificial Intelligence International Governance at Tsinghua University, said when talking about the development trend of generative artificial intelligence, "AI large models are becoming a new type of industrial infrastructure, and provide intelligent services for thousands of industries, and promote AI R&D and production from traditional small-scale workshops to large-scale industrial mass production with the paradigm of "basic large model + industry data fine-tuning". ”

In line with the industry trend, Liu Qingfeng, chairman of iFLYTEK, is full of confidence in the implementation of China's large model application, "We must not only rationally see the gap between China and the United States in science and technology, but also have the confidence to do the best in the application, today we have been able to fully benchmark GPT-4, in actual commercial use, from 100 billion to 10 billion to billion parameters and other models of various sizes, in the special field we can do the best in the world." ”

The question that emerges from the WAIC field is, how can large models be applied on a large scale?

There are two ways to solve this problem in the industry.

The first is to move from a general model to a vertical model, and use more professional models and applications to solve the problems of subdivided tracks and specific scenarios. In the fields of healthcare, finance, education, automobiles, and industry, Huawei, iFLYTEK, Ant, NetEase, and Ctrip have all launched industry models.

The second is to combine large models with agents to provide users with a series of intelligent services in the form of personal AI assistants or agents, and agents are particularly active in software development.

For example, ByteDance has launched a one-stop AI development platform "Coze Button", which can support 30-second no-code generation of AI bots, integrating more than 60 different plug-ins, covering news reading, travel planning, productivity tools and other fields. Baidu has upgraded its industrial-level AI native application development platform "Qianfan AppBuilder", which greatly lowers the threshold for AI native application development and improves development efficiency. iFLYTEK launched the Spark Enterprise Intelligent Twins Platform, which provides core capabilities to enable enterprises to quickly build landable intelligent twins.

In addition, humanoid robots, which are highly related to artificial intelligence and intelligent manufacturing, also broke out in this conference and became one of the highlights.

The robots shown are very cool, but whether these robots can do their work and whether they can really replace humans are still subject to question marks, which also reflects the fact that natural intelligence is the same as large models.

Overall, this year's WAIC scene is still hot, and the large model is welcoming the upcoming era of its own with a new attitude.

However, behind the prosperity, the product differentiation and innovation points are relatively similar, mainly in the form of software, and there are many non-rigid applications such as general assistants, chatting assistants, Wensheng diagrams, and AI searches.

In the context that application landing has become an industry consensus, how to land, whether it is an industry model or a personal agent, needs to move from virtual to real, and use the personalization of application experience to fight the homogenization of industry competition, so as to reflect the value of the application and form a value cycle in the industry ecology.

02

Take a peek at the landing sample of the iFLYTEK Xinghuo

We noticed that among the many exhibitors at WAIC, iFLYTEK showed a different way of thinking from other manufacturers in terms of large-scale model application.

The eye-catching "AI assistant that understands you" at its booth is the core of iFLYTEK's Xinghuo display. Not long ago (June 27th), the iFLYTEK Xinghuo large model has just completed the upgrade of the V4.0 version, in addition to the overall technical progress over GPT-4 Turbo, the biggest attraction is to make the large model more user-friendly and practical through personalization.< em> WeChat </em> picture _20240709163305.jpg

At the scene, iFLYTEK demonstrated how to create an "AI assistant that understands you" in life assistants, work assistants, teaching assistants, enterprise assistants and C-end products.

For example, in the field of life, iFLYTEK has launched the AI health assistant iFLYTEK Xiaoyi APP, which has no commercial advertisements and will not confuse information, as users upload personal information such as electronic medical records, examination reports, and physical examination reports, it can build a personal digital health space, which can further analyze the cause of the disease before seeing a doctor, give a personalized judgment of drug contraindications when taking medication, give data changes in joint comparison after examination, and understand the health status of other family members through role switching.

Among the teaching assistants, the iFLYTEK Xinghuo smart blackboard exhibited this time won the highest award of the 2024 WAIC - Outstanding Artificial Intelligence Leadership Award. It is reported that this is the only award-winning project in the field of education of this year's Sail Award. It is equipped with the iFLYTEK Xinghuo model, which can deeply empower classroom teaching to improve quality and efficiency, and make the blackboard change from a board tool to an AI assistant for teachers.

For enterprise scenarios, iFLYTEK focused on the Xinghuo enterprise intelligent twins platform to solve the "last mile" problem of enterprise large model applications. The platform focuses on the three key capabilities of building agents, covering 400+ AI atomic capabilities, integrating 90+ external sources, and opening up 100+ internal IT systems, so that enterprises can quickly build intelligent applications that can be implemented with less computing power and higher efficiency in combination with business scenarios.

For example, the Spark Business Opportunity Assistant can realize the knowledge of business opportunity clues, improve the quality and efficiency of customer visits, and intelligent research and judgment of sales management, helping to improve the efficiency of front-line sales and business opportunity management. Through functions such as pre-bid sourcing, intelligent bid evaluation, and bid calibration review, the Spark bid evaluation assistant greatly improves the efficiency of enterprise bid evaluation and reduces procurement costs.

In addition, iFLYTEK also displayed C-end hardware products including iFLYTEK AI learning machine, iFLYTEK intelligent voice recorder, iFLYTEK translator, iFLYTEK smart office notebook, etc., through the application of large models, these hardware products have been greatly improved in function and experience compared with traditional products.

It is not difficult to see that the large-scale model application of iFLYTEK, starting from the rigid demand, grasping both the B-end and the C-end, and truly realizing the value landing through personalization, is currently the most thorough application of the rigid demand application in the domestic large-scale model application.

Among them, the landing of iFLYTEK Xinghuo shows the following very distinctive characteristics.

First, it has completed a comprehensive evolution from a general-purpose AI assistant to an "AI assistant that understands you", making the large model easier to use and more practical.

Second, iFLYTEK uses the "cloud-edge-end" overall solution to turn the large model into a "wishful golden hoop stick", providing full-size large models to meet the needs of large models in various complex and multi-faceted scenarios.

Third, the integration of "cloud-edge-end" also brings the combination of software and hardware integration, which is not only reflected in the dazzling array of C-end hardware products, but also in smart home appliances, smart cars, operators, robots and other industry scenarios.

For example, the Epoch ET smart cockpit jointly built by iFLYTEK and Chery uses the Xinghuo model as a smart middle platform, opening up internal and external devices such as heart rate sensors, which can monitor and remind drivers of heart rate to ensure the safety of long-distance driving. There is also the industry's first Xinghuo intelligent marking machine, which looks like a printer, but has the "magic" of automatically correcting test papers, allowing teachers to reduce their burden and increase efficiency.

03

Take the lead in creating a personalized development path for large models

Judging from the trend shown by WAIC, it is not difficult to implement large model applications, but it is difficult to achieve valuable implementation in multiple scenarios.

iFLYTEK Xinghuo can take different paths among a group of players, first of all, because of the strong support of the two bases of technology and computing power, that is, the performance of the application is based on the ability of the basic model.

As a national team of artificial intelligence, iFLYTEK insisted on benchmarking against the international advanced level when it released the iFLYTEK Xinghuo large model in May last year, and when iFLYTEK Xinghuo upgraded to V3.0 version, it has achieved the overall surpassing of ChatGPT, and the medical field has surpassed GPT-4.

On the eve of WAIC, the upgraded iFLYTEK Xinghuo V4.0 won the first place in 8 international authoritative test sets, surpassing GPT-4 Turbo as a whole. During this year's college entrance examination, a number of media and professionals used Chinese and English composition questions and math questions in the college entrance examination to evaluate the ability of large models.

What we see on the front-end is the ability presentation of large models, and in order to output results, large models need to undergo complex training and calculations on the back-end, in which computing power plays a very important role.

In the context of the Sino-US science and technology game, the independent and controllable supply of computing power is particularly critical, which largely determines how far the domestic large model can go.

iFLYTEK Xinghuo is precisely China's first large-scale model that is open to the whole people based on domestic computing power, so it can bring higher voice and security to the industry under the base of domestic computing power.

Returning to the essence of the landing of large model applications, the second difference path of iFLYTEK Xinghuo is that it is not for the sake of landing, but before each application lands, it is thinking and answering - what problems can I solve, and whether I can provide different services and experiences.

Translated into this, the personalized landing of large model applications in rigid demand scenarios should be both practical and easy to use.

In public, Liu Qingfeng, chairman of iFLYTEK, has more than once expounded on the "important value of application scenarios", and in the field of large models, "whoever can land in the application scenarios that have the rigid needs of society will be the first to form a virtuous circle of self-hematopoiesis".

Let's take a look at the C-side first, iFLYTEK believes that in order to make the large model easier to use, the most important thing is that the AI assistant should be able to carry out personalized expression based on user portraits, which is also the first time that iFLYTEK has systematically proposed the development of personalized large models in the industry.<em> WeChat </em> picture _20240709163327.jpg

Therefore, when the iFLYTEK Xinghuo V4.0 was upgraded, the "personal space" function was released in the iFLYTEK Xinghuo APP and desktop version, and users uploaded their work, study, life, health and other information to form a personal exclusive knowledge base, and the large model can generate personalized content belonging to individuals according to the persona.

In order to break through the barrier of software and hardware ecology, iFLYTEK Xinghuo has launched the "Xinghuo Family Bucket", which can synchronize hardware data such as iFLYTEK smart office and intelligent voice recorder to Xinghuo's personal space with one click, and through data exchange and operation linkage, you can use Xinghuo's capabilities for efficient office.

Reflected in the data, since the full opening in September last year, the Xunfei Xinghuo APP has been downloaded 131 million times in the Android open market, ranking first in the domestic tool category general large model App.

In the B-end scenario, the ToB business in China in the past was usually "people adapt to machines", but with the arrival of large models, iFLYTEK thinks more about "making models better applicable to people and industries", and makes large models truly practical and easy to use through the overall solution of the cloud edge end and the enterprise intelligent twins platform.

Among the central state-owned enterprises with rich application scenarios of large models, iFLYTEK Xinghuo has become the choice of leading enterprises in many fields such as National Energy Group, PetroChina, China Mobile, Chinese People's Insurance Company, Pacific Insurance, and Bank of Communications by virtue of its basic capabilities and domestic computing power base.

In addition, after the comparison and selection of customers, iFLYTEK Xinghuo has become the preferred brand for large-scale model applications in the fields of automobiles, home appliances and robots.

In the automotive field, domestic brands such as Chery, NIO, GAC, and FAW are cooperating with iFLYTEK, and the international auto brand Volkswagen has also quickly applied it after evaluation, and released the first model "ID.UNYX" equipped with the Xinghuo model on March 13.

In the field of home appliances, Haier's home appliance model chose iFLYTEK after bidding. The company is also carrying out all-round cooperation with Midea Group, Hisense Group, TCL Group and other home appliance companies.

iFLYTEK has cooperated with UBTECH, Fourier Intelligence, Unitree Technology, Pujin Intelligence and other brands to empower nearly 400 robot companies.

These landings are also reflected in the bidding of large model projects, and some media have counted the winning projects of the large model in June, showing that among the 75 large model winning projects, iFLYTEK won the bid for 8 projects, and the number of winning bids and the amount of winning bids are the first.

Just like Liu Qingfeng said: "These are the results of the careful choice of customers and the PK of each company, which represents the real hard power of iFLYTEK Xinghuo." ”

04

epilogue

From the technology, to the product.

This is the concept that iFLYTEK has adhered to since its establishment, and in this WAIC, iFLYTEK has demonstrated to the industry its understanding and differentiated path of large-scale model application, and through the landing ability of "independent and controllable large-scale model base + personalization + rigid demand scenario", it will take the lead in forming a differentiated path and competitive advantage in the inflection point of change in the large-scale model era, and the potential of commercial landing may be accelerated.

* The picture comes from the Internet, please contact to delete the infringement

This article is from Xinzhi self-media and does not represent the views and positions of Business Xinzhi.If there is any suspicion of infringement, please contact the administrator of the Business News Platform.Contact: system@shangyexinzhi.com