After ChatGPT became popular, the popularity of large models continued unabated, and a situation of "thousand models war" gradually formed. In the first two months, Baidu, Ali, and Tencent joined; in May, Netease Youdao first released a spoiler video of an AI oral teacher developed based on the "Ziyue" model, introducing its ChatGPT-like product based on educational scenarios; Later, there was HKUST Xunfei, which released a large-scale cognitive model "Xunfei Xinghuo"... In short, various large and small press conferences are proving to the outside world how much room for imagination they have.
With the development of large-scale models in full swing, its industrialization has also been mentioned more and more, and the previously hyped "AI+" has once again stood on the cusp. In this context, the application of AI large models in the automotive field has also begun to soar.
The "ChatGPT moment" of smart cars is here
At present, although it has become an industry consensus to reconstruct all walks of life with large models, ChatGPT "boarding" is much faster than expected. As the "third space" besides the home and office, the car is becoming a new type of smart terminal. After the arrival of ChatGPT, the car-machine relationship has also been more affected.
One is the impact of ChatGPT on the automatic driving of cars. According to industry analysts, the large model has the ability to process massive data and multi-dimensional analysis capabilities, which can provide more accurate and comprehensive data analysis and prediction capabilities, continuously optimize the model, and improve the accuracy and reliability of intelligent driving. Especially with the emergence of ChatGPT, people found that when the parameter amount of the model reaches a certain level, the effect presented is not "better performance", but "surprisingly good".
Specific to the application level, the impact of large models on autonomous driving is: in the cloud, car companies can take advantage of the large capacity of model parameters, complete most of the data labeling and data mining work through large models, save data labeling costs, and Able to build empowerment with the help of simulation scenarios. On the vehicle side, it can combine small models in charge of different sub-tasks into a large model, saving inference calculation time on the vehicle side and increasing vehicle safety. Most importantly, the bottleneck of the end-to-end perception and decision-making integration algorithm, which is considered to be the end of the autonomous driving algorithm, may be effectively solved after the car is connected to the large model, and the upgrade of the autonomous driving algorithm is just around the corner.
The second is the impact on the smart cockpit of the car. The on-board ChatGPT voice assistant can handle complete conversations, such as asking questions, and can maintain an understanding of the context, forming a relatively good voice interaction experience. For example, Microsoft and Mercedes-Benz are exploring ChatGPT's plug-in ecosystem to open up possibilities for third-party service integration. In the future, drivers are expected to complete tasks such as booking restaurants and movie tickets through the on-board system, further improving convenience and productivity, and greatly enriching the interactive experience between smart cars and people.
In addition, in the case of changing the interactive capabilities of intelligent driving and intelligent cockpit, it will also have a new impact on the research and development methods and business models of automobiles. In terms of research and development methods, due to the high-efficiency labeling capabilities of the machine, the data labeling task that takes a year only takes a few hours, and the research and development cycle is greatly shortened. Moreover, the rich data of multi-modality (vision, voice, gesture, etc.) It can further improve the overall R&D efficiency and reduce R&D costs. As far as the business model is concerned, after the in-vehicle AI voice interaction has emotional intelligence, it will evolve from an "employment relationship" to a "accompanying relationship". It will have a better understanding of people's preferences and habits, which will lead to a new business value.
Various genres of industry exploration
Perhaps because of this awareness, more and more car companies are now choosing to access AI large models. In addition to the announcement of ChatGPT by Mercedes-Benz abroad, the domestic ideal car also released MindGPT, a large self-developed model, and Baidu’s Wenxin Yiyan has also been connected to many car companies such as Changan, Geely, Lantu, Hongqi, and Leap. Even Ferrari, which "suffocates dreams" in the future, has incorporated large-scale models in new cars. It is not difficult to foresee that large-scale models will be popularized in smart cars in the future, which will be a high probability event. From the point of view of the entire participating car companies, the direction and focus of their development of large models are not the same.
From a functional point of view, it can be mainly divided into the following two categories: one is used in the field of artificial intelligence communication and dialogue, and most of them are used in smart cockpits. For example, Baidu’s Wenxin Yiyan, nearly ten car companies such as Dongfeng Nissan, Hongqi, and Great Wall, have announced their access; during the Shanghai Auto Show, SenseTime unveiled the Ririxin Sensenova large model, showing its Chinese language model "Discussion SenseChat" and "Ruying SenseAvatar" are combined with the cockpit; Alibaba also announced that the AliOS smart car operating system has been connected to Tongyi Qianwen large model for testing.
The other category is large-scale model applications that focus on intelligent driving. For example, Momo Zhixing released DriveGPT, a large generative model for autonomous driving, to help solve cognitive decision-making problems and ultimately realize end-to-end autonomous driving. Ideal Auto's self-developed large-scale model MindGPT gets rid of the dependence on high-definition maps, allowing the car to achieve driving performance closer to that of a human driver. Weilai, Xiaopeng, Great Wall, and Chery have also registered and applied for multiple trademarks related to GPT.
From the point of view of the participants, it can be divided into two categories: one is that the car companies do it themselves, such as Ideal Automobile and Baidu, which apply their own large models to their own car products; the other is that external manufacturers provide large models to the car. Enterprises, such as Huawei Pangu large model, Baidu Wenxin Yiyan, etc., are used by other car companies.
Compared with general-purpose large-scale models, large-scale model training and use costs in vertical fields are lower, and may become an area that is easier to achieve commercialization. According to industry analysts, automobiles have clear interaction requirements, and compared with general-purpose large models, the application scenarios in the vertical field are relatively small, and the magnitude requirements for parameters are not as large as general-purpose AI. Therefore, whether it is a traditional car factory or a new force, or a large-scale technology manufacturer, it is generally believed that smart cars are most likely to become the first B-end scene to realize the landing of large models.
It is not yet ripe for large-scale models to be put on the car
At present, although ChatGPT's first "get on the car" has opened the prelude to the competition of large models in the field of smart cars. But at this stage, there is still a long way to go before the real scale of the large model.
First of all, the collection, processing, and training of multi-modal data is itself a difficult problem for car companies to build large-scale models. The sensor data required for autonomous driving includes lidar, millimeter-wave radar, ultrasonic radar, as well as high-definition cameras, GPS, etc. These data come from different coordinate systems, with different trigger timestamps, and hardware damage and other issues must be considered; at the same time, a large amount of scene data is required, such as traffic sign lines, traffic flow, behavior models, etc. This makes the development and training threshold of large car models very high.
Industry insiders believe that model data mobilization management needs to use many platforms such as intelligent networked vehicles, computing technology platforms, cloud control technology platforms, etc. Only when massive data aggregation can be done, especially in the vertical field, this is different from other systems in the car. It is not the same. If the basic platform capabilities cannot be opened up, it will be difficult to develop to a deeper level. In addition, although generative AI has achieved a breakthrough in information acquisition, it is still far from actually landing on the car in terms of decision-making and execution control.
In addition, the end-to-end AI large model training needs to build a new algorithm based on the AI large model in the smart car, which also requires a process. The industry believes that it will take at least 3-5 years or even longer to complete this. a process.
Secondly, limited by the hardware conditions of the on-board equipment, the hardware configuration required by the large model in the car may be limited, making it difficult to effectively play a good role. Specifically, large models require high-standard hardware configurations, including high-performance computing capabilities, large-capacity memory, and low latency. However, the hardware conditions of on-board equipment are relatively limited and cannot provide sufficient computing resources to support large models. For example, the GPT-3 model in the field of natural language processing requires trillions of TOPS of computing power. This requires that the computing power of the chip must be at least above 10,000 TOPS to be able to handle the computing tasks of large models. However, in the vehicle deployment scenario, the computing power of the chip is often only a few hundred TOPS, which is far from meeting the requirements of large-scale models.
In this context, upgrading the computing power infrastructure in cars has become an inevitable trend. At present, the intelligent computing center may become the "standard configuration" of smart cars in the future. For example, Tesla released Dojo, an independent cloud computing center, which used a total of 14,000 Nvidia GPUs to train AI models. In China, Xiaopeng Motors and Alibaba Cloud jointly built the intelligent computing center "Fuyao", which is specially used for autonomous driving model training, and the computing power can reach 600PFLOPS; Momo Zhixing and Volcano Engine jointly built the largest intelligent computing center in the domestic autonomous driving industry "Snow Lake. Oasis", the floating-point operations per second can reach 6.7 trillion times. However, these constructions are still in the stage of exploration and application, and the timing of large-scale application of large-scale models has not yet fully matured.
The underlying technology is the focus of future competition
At present, the most essential change of the AI model to the car may be that it will further drive the car, from the attribute of manufacturing to the attribute of technology + consumer electronics. In this context, the underlying technological capabilities of the car factory itself will become the key to success in the future.
On the one hand, the AI large model will redefine "human-computer interaction" and "service ecology", and accelerate the electronic consumption of automotive application service ecology, which will change the underlying product definition. The bottom layer of the current in-vehicle voice system is an immature task-based dialogue system, which cannot truly realize personalized, emotional, and free interaction capabilities. However, large models can use deep learning + voice generation to usher in open scenarios + natural interaction. A new human-computer interaction experience. At the same time, in terms of R&D and design, with the OSization of large models, the necessity and importance of traditional APP on-boarding, touch functions, and HMI delivery is worth considering, which means the redefinition of automotive "products".
As Su Qing, the former director of Huawei's autonomous driving product department, said: "Traditional car manufacturers think that my base is the car. Now there are some single points of computers. Then I use the car as a base and try to embed the computer. This is the traditional The view of the car factory. Our views are different. The foundation is the computer, and the car is a peripheral device controlled by the computer. A large computer does the work, and the car is attached to it. , the car's computing center, intelligence, and consumer electronics attributes will become more obvious, and its product definition will follow the iterative evolution of the underlying logic of consumer electronics (such as APP, ecological services, etc.).
On the other hand, the AI large model will change the existing algorithm of the car. Considering the hardware cost required for the application of the large model and the external environment, the technical strength of the large model of domestic enterprises in the future will depend more on the AI operating system, etc. underlying technology. Specifically, the original automatic driving algorithm of the car is more dependent on manual work, but after accessing the large model, it needs to rely on the large model to drive the formation of a new intelligent algorithm. To build such an algorithm, in addition to making a good product, it is also necessary to platform.
For example, computing platforms around chips, cloud, etc. need to be low-cost and standard configuration; only with good products can there be enough sales, and only with enough sales can we increase the "end-to-end" Closed-loop data (from the cloud to the car end); as the basic support for AI chips with large computing power, it also needs to reconsider the situation of the car companies themselves. Who will run to the front.
Take Nvidia's training chip as an example. A chip was fired for 100,000 yuan. The key is that this high-cost chip training solution is not the best solution for car companies. Moreover, affected by external sanctions, this kind of "external supply dependence" may be pinched off at any time, which is very vulnerable. Therefore, in the long run, if car companies want to go to the forefront of the industry in this field, they must increase their efforts in self-development of underlying chips to reduce procurement costs, such as Baidu's self-developed AI chips, or improve AI algorithms and operating systems. Seek to break through to find the best application path.