Text: White Pigeon
Edited by Wang Yisu
The Beijing Auto Show, which has returned after four years, has encountered the boom of smart cars in China.
The continuous baptism of the price war at the beginning of the year not only made a number of Chinese car companies panic, but also made the global giant Tesla face a sharp drop in sales in the first quarter.
At the same time, smart cars are constantly rolling up new technologies.
First of all, each company has played "high-level intelligent driving", and competed under the route of heavy perception and light map, "intelligent driving can be driven all over the country". In addition, under the wave of large models, the "onboarding" of large models has also had a new substantive landing this year.
As early as 2023, the popularity of ChatGPT has already driven large models to the car.
Autonomous driving solution providers such as Momo Zhixing and car companies such as Xpeng Motors have followed Tesla closely and applied the Transformer architecture to intelligent driving systems, Zhiji's model LS6 has launched the GPT large model, and GAC has released an AI large model platform, which will be installed in the Haobo GT.
In general, there are two types of large models "on the car", one is in the field of intelligent driving, which uses an end-to-end way to improve the perception ability of the intelligent driving system, and the other is in the field of intelligent cockpit, which completes the natural language interaction between people and vehicles.
However, an important problem facing large models is that the application scenarios that have been implemented are still relatively narrow, and the main mature application scenarios are in the fields of intelligent customer service and cockpit interaction.
However, as the automotive industry enters the "new intelligence" stage, the breakthrough of AI large models is far more than these.
"The application value of large models in the automotive travel industry is far more than the obvious 'on-car' experience for users, and there is also a lot of practice and application space in more 'off-car' scenarios such as R&D, production, marketing, and customer service. Zhong Xuedan, vice president of Tencent Smart Mobility and head of Tencent Smart Mobility, said.
It is foreseeable that with the continuous implementation of these scenario applications, the automotive industry will truly enter the "high-level intelligence" of the whole industry chain.
From getting in the car to getting off, car companies rushed to make large models
Competition in the automotive industry is becoming increasingly fierce, and the price war is intensifying.
When homogeneous suppliers meet homogeneous design, car companies need to continuously reduce costs and increase efficiency, and at the same time, they urgently need to build new competitiveness in the intelligent era.
The AI model has become the key for car companies to build differentiated capabilities, and it is also the basic foundation capability for car companies to build global intelligence.
Last year, Geely Automobile Research Institute set up an AI intelligence research department, Geely Group CEO Gan Jiayue said that Geely already has a full-stack self-developed large model technology, Li Xiang, founder of Li Auto, said, "The development and training of large models is a necessary capability for smart electric vehicle companies." ”
It can be seen that car companies are actively building large model capabilities, and there are two main ways to build them:
The first is independent research and development of large models, led by new car-making forces such as Wei Xiaoli, such as Li Auto's self-developed Mind GPT Chinese large model; Wei's self-developed AI large model NOMI GPT; Xpeng Motors' self-developed XGPT Lingxi large model.
For example, Changan, Chery and other companies are working with Tencent to explore the development and application of cockpit vertical models, and Great Wall Motors and iFLYTEK have built a knowledge model "Great Wall Motors Knowledge Brain".
At present, the focus of the application of car companies in the field of AI large models is more on "getting on the car", that is, continuously improving the level of vehicle intelligence through the capabilities of large models.
For example, in the field of intelligent cockpits, the powerful computing and processing capabilities, production and multi-modal interaction capabilities behind large models can bring a higher level of intelligence and richer interaction capabilities to the intelligent cockpit.
One of the most obvious manifestations is the on-board voice interaction system, when the on-board voice interaction system is connected to the AI model, it can realize the intent understanding of multiple rounds of dialogue, and can chat with the user, rather than just a machine that executes voice commands.
In the field of autonomous driving, Tesla was the first to introduce the Transformer large model in the training of autonomous driving systems.
It is understood that in terms of image recognition capabilities, Transformer has a higher upper limit, with the growth of the amount of training data, the traditional CNN model recognition ability is saturated, while Transformer has better performance in the case of larger data volume.
Autonomous driving is a scenario with massive data, and it is also necessary to continuously train higher-quality data from massive data to promote the upgrade and iteration of the autonomous driving system.
In addition, with the intensification of competition in the automotive industry, car companies not only need to "get on the car", but also "get off" in the application of large models. The AI large-scale model capability will be integrated into all business scenarios of car companies and all links of the industrial chain to build global intelligence and continuously improve R&D production efficiency.
After all, the efficient operation of a company from the inside out is also the key to competing in the industry.
At present, major automobile companies have incorporated large models into key investment plans, and superimposed AI productivity in various links such as R&D, production, sales, service, and collaborative management of automobiles to improve quality, reduce costs, and increase efficiency.
It can be seen that car companies attach great importance to the implementation of AI, and they are also making positive changes to this end. Chen Rui, deputy general manager of Guangzhou Automobile Group Daisheng Technology Co., Ltd., said that AI will be quickly applied to various fields such as production and life. But at the same time, the full application of AI also requires comprehensive transformation of enterprises, organizations, and people.
Yao Zhen believes that the new changes brought by AI models to the automotive industry are mainly reflected in two aspects: cost reduction and efficiency increase.
At the cost reduction level, it is for the internal R&D, generation, management and other links of car companies, such as code assistants, conference document AI assistants, etc., to reduce labor costs.
"Because of the introduction of Code Assistant, some of our customers have seen a 7% increase in labor efficiency. Yao Zhen said to the light cone intelligence.
Efficiency is more reflected in sales and service. For example, in the process of consumer car purchase, the ability to assist sales consultants and auxiliary knowledge bases based on large models will allow sales and service consumers to have a clear data strategy for reference and improve sales conversion rate.
Since the large model is so important, how to take the lead in landing the ability of the large model has also become the direction of thinking for car companies.
How do car companies use large models to practice "internal skills"?
So, how can car companies quickly implement large-scale model capabilities?
With the help of strength, compared with the investment of complete self-research and slow results, it is the key to choose the right large-scale model service provider in the automotive industry.
"We found that the application of the general model in the automotive industry is not so good. Yao Zhen said.
The strength of the general model focuses on the word "wide", which is very wide for people and scenarios. However, for specific scenarios, enterprises do not need the "all-round" capabilities of general large models, but more need the accuracy and quality of the models.
Therefore, based on Tencent's self-developed hybrid model, Tencent Smart Mobility has launched a large model for the automotive industry, landing in five scenarios: automobile R&D, production, marketing, service, and enterprise collaboration.
Yao Zhen said: "Based on this large model of the automotive industry, the application effect is very good. Because of the addition of massive professional data in the automotive industry, pre-training, fine-tuning and reinforcement learning of vertical tasks in the automotive field are carried out, especially in Chinese reading comprehension, end-to-end Q&A, and automotive industry-related tasks. ”
For example, in R&D, Tencent's large model can provide AI code assistant capabilities, which can assist software engineers in writing code, supplementing code, diagnosing code, and test cases; in the production process, Tencent's AI quality inspection function can screen for defective auto parts, alarm for illegal operations, and visualize the operation and production progress of production lines and production equipment in real time.
At the marketing level, based on Tencent's large-scale model capabilities in the automotive industry, it can intelligently generate promotional copy, posters, and videos through AIGC, and can produce exclusive digital intelligent avatars within 1 hour, and can sell cars live online, not only familiar with the selling points of the model, but also accurately interact with the audience, effectively promoting the conversion of clues from online to offline test drives. For B-end salespeople, the sales AI assistant can help salespeople quickly obtain accurate business analysis of automobile sales.
At the service level, based on the intelligent cockpit model created by Tencent Mix, by adding professional car knowledge and fine-tuning the model, it can achieve high-level, professional and free Q&A, for example, in the intelligent customer service scenario, it can provide more accurate and detailed answers to car use and maintenance questions. At the same time, it can also assist human customer service, automatically generate a summary of the previous session, assist in retrieving difficult problems, and automatically summarize and fill in the form after the session.
In the face of such a huge systematic project, how will car companies prioritize to land?
Yao Zhen said: "The fastest landing speed is in the fields of marketing and sales, because it can help car companies sell cars well and serve customers well, so whether it is budget or priority, car companies must pay the highest attention." ”
Secondly, it is the intelligent cockpit, which mainly involves the improvement of product power, and even whether users will buy this car, so car companies attach great importance to this area. "However, at present, there is no company in the entire industry that can make the application of large models in the field of intelligent cockpit unique, and the capabilities of each company are relatively scattered. Therefore, it is more important to build an ecosystem that can serve the intelligent cockpit based on the open capabilities of large models. ”
Finally, it is to help car companies build a large model base horizontally, such as in collaboration and assistant code. "Generally speaking, car companies will allocate part of their budget to build the entire AI large model base, and build up the large model base with the characteristics of the car company and the upper-level scene capacity to carry out a complete systematic connection, and the overall project will be very deep. ”
"Although the current priority of timeliness is not high, the landing can be more in-depth to help car companies do the construction of internal strength. Yao Zhen said to the light cone intelligence, "From the actual effect, it is not necessarily possible to quickly see the effect in a certain field, but more to build a basic capability." ”
Behind the implementation of these five application scenarios, Tencent's complete capabilities from models, computing power, AI engineering platforms (tool chains) to AI applications are inseparable.
Taking the intermediate tool chain layer as an example, Tencent can effectively reduce the difficulty of large model training and improve R&D efficiency by providing platform-based capabilities.
Tencent Cloud's TI platform's industry large model fine-tuning solution presets data from 140 different types of large-scale model fine-tuning task scenarios, covering data processing links such as data cleaning, prompt optimization, data filtering, and data augmentation, which can quickly prepare high-quality fine-tuning data and support the use of matching in the fine-tuning process to achieve better model fine-tuning results.
In addition to the AI tool chain, Tencent Cloud also has a ready-to-use large-model knowledge engine, which can provide large-scale model capabilities for many scenarios such as intelligent customer service, sales AI assistants, and enterprise training, so that consumers and automotive professionals can gain in-depth knowledge about vehicle performance, maintenance, and market trends through simple interactions.
In general, in order to support car companies to make large models on the car, cloud vendors have done a full set of support: improve the utilization rate of computing power (cost reduction), and do efficient development and application on the platform (efficiency increase).
However, Yao Zhen also said: "The technology upgrade and iteration of the large model is very fast, and the ability of the hybrid large model in terms of openness and industry adaptability has been continuously improved. Therefore, starting from the first half of this year, we found that it is more cost-effective to directly call hybrid capabilities in the automotive industry, so we will also directly implement some hybrid cloud privatization capabilities this year. ”
In fact, car companies will have different needs for large models.
For example, in response to the needs from business departments, including technology centers, sales companies, brand departments, IT businesses, etc., it will focus on subdividing needs in the fields of intelligent cockpit, intelligent customer service, and marketing.
In response to the needs of segmentation, the solution provided by Tencent is to quickly help car companies build large models from the application point, "with a very light closed loop, in a specific scenario, to provide large model services for car companies, avoiding their own infrastructure large models, and it is also more cost-effective for car companies." ”
In addition, some car companies hope to build their own large-scale model base through heavy investment, and for this part of the demand, Tencent can provide full-stack technology closed-loop capabilities from the underlying infrastructure, to the tool chain, to the upper-layer application.
Tencent's full-stack closed-loop underlying technology service capabilities enable car companies to better "get off" in large-scale model applications. In addition to the large model "getting off", Tencent can also make the application on the car more abundant based on the combination of the large model and the ecosystem.
The king of socializing, the app "Get in the Car"
In the process of competing for car companies, cloud vendors have homogeneous competition, and they also need to think about their own differentiation.
For Tencent, the C-end characteristics of entertainment and social networking are an important starting point for car applications.
The characteristics of the C-side here mainly include two concepts:
One is Tencent's own huge content ecological resources, such as music, WeChat, mini programs, etc., which Tencent can directly move into the intelligent cockpit system;
One is Tencent's technical capabilities that have been precipitated in the process of doing To C business, not only large models, but also Tencent Security, Tencent Cloud, etc., and through these technologies, it provides underlying technical capability support for the automotive industry;
For example, Tencent will move WeChat to the car, using WeChat as a super ID, as a bridge between the mobile phone and the car machine, users can scan the code, the information on the mobile phone WeChat can be sent to the car machine with one click, including music on the mobile phone, audio books, official account content, etc.
In addition to the super ID of social networking, Tencent has also implanted its best game apps into the car. Mercedes-Benz said it will join hands with Tencent at the Beijing Auto Show to integrate a world-class racing game into Mercedes-Benz's in-vehicle system.
The idea of using C-end applications to open up the differentiation of Tencent's solutions in the automotive industry comes from Tencent's thinking about the relevance of technology and scenarios.
"There is a very simple principle, we want to improve efficiency and solve practical problems through digital intelligence. Yao Zhen said.
For example, in the field of autonomous driving, Tencent's accumulated technical experience is mainly focused on the collection, storage, labeling, and simulation training of autonomous driving compliance data. ”
However, the integration of C-end applications and large models is a capability that Tencent is good at.
Yao Zhen believes that if the intelligent cockpit wants to create a better user experience, it cannot only rely on the capabilities of large models, but also the large model, the underlying computing power and ecological resources.
Tencent's advantage lies in its ability to connect a wide range of ecological Internet resources with the cockpit car machine system, and integrate C-end product capabilities including maps, music, WeChat, games, and mini programs into the car machine system, and realize the seamless connection between the mobile phone and the car machine experience, providing users with a richer experience.
Before AIGC was widely used, Tencent already had a complete and mature AI system in the field of intelligent cockpits, and at this stage, Tencent focuses on combining this system with large models to build a set of intelligent recommendation systems with Internet ecology in the intelligent cockpit.
Sun Jue, general manager of Tencent's smart travel intelligent cockpit, said that based on Tencent's APP Agent (intelligent agent) capability, under the mature process, it only takes one week to learn and proficiently use hundreds of applications or applets, and does not require API to achieve in-depth voice interaction with in-vehicle applets and apps, which can fully understand user intentions, replace users to intelligently operate various complex applications, and improve service efficiency.
According to Yao Zhen, Tencent's intelligent cockpit solution will reach 15 million units by the end of this year.
It can be seen that based on its own ecological characteristics on the C-end and the ability of AI large models, Tencent's intelligent cockpit has found its own position in the industry.
Tang Daosheng, Senior Executive Vice President of Tencent Group and CEO of Cloud and Smart Industry Business Group, concluded that Tencent will continue to consolidate its core technical capabilities such as cloud and graph to build a solid foundation for the development of intelligent vehicles. On the other hand, we will work with upstream and downstream partners in the industrial chain to actively explore the innovative application of cutting-edge technologies such as AI large models in various scenarios of the automotive industry, and drive the "new intelligence" development of the automotive industry with AI.
#2024北京车展# #腾讯云#