Source: "Chebai Think Tank" (ID: EV100_Plus), author: Chen Chongshan

According to Tianyancha, in the first half of 2023, there will be more than 20 financing events directly related to large-scale models, and the number of various large-scale models that have been released in China exceeds 100, showing the trend of a "hundred-model war".

Source of large-scale model financing events in the first half of this year: Sohu Technology

The mobile phone industry has already begun to apply large models. The support of large models on the device side and the cloud side is promoting the innovation of mobile phone acceleration experience. Some media even said that the deep integration of large models and mobile phone systems will make mobile phones a new species. "

Compared with the mobile phone industry, in the new energy vehicle industry, it has also become a trend to "get on the car" with large models, and it is also very popular.

This is because, as a relatively vertical field, large models are more widely used in new energy vehicles. Smart driving, smart cockpit, and even the resulting changes in business models, as well as car research and development, all have large-scale application scenarios, which will help cars quickly upgrade from transportation attributes to smart terminal attributes.

Recently, GAC Group recently announced the official launch of the latest research and development results of AI large-scale model technology - GAC AI large-scale model platform. Geely Automobile also stated that the first full-stack self-developed full-scene AI large model will be carried on the Geely Galaxy L6. Moreover, automotive supply chain companies have also set foot in large-scale models, such as the Ririxin large-scale model released by SenseTime; Momo Zhixing released the autonomous driving generative large-scale model DriveGPT-Xuehu Hairuo, and so on.

"The one thing that currently restricts the development of the domestic large-scale model industry is 'too hot'", this is the recent observation of Huang Tiejun, dean of Beijing Zhiyuan Artificial Intelligence Research Institute and professor of the School of Computer Science, Peking University. In fact, the current large model "on the car" is also overheated.

Some car companies not only apply large models in technology, but also have registered "GPT" related logos on their trademarks. For example, Chery New Energy Automobile has registered the "ICAR GPT" trademark, Great Wall Motor has registered trademarks such as "Great Wall Motor SPACEGPT", and NIO and Xiaopeng Motors have also applied for the registration of GPT-related trademarks.

As for the role of large models in the automotive industry, industry insiders position them more as "infrastructure".

Recently, Zhang Yongwei, vice chairman and secretary-general of the China Electric Vehicle 100, pointed out in the article "Several Strategic Issues in the Development of Smart Vehicles in my country" that the intelligent stage will require more new infrastructure, including large models. He pointed out that it is necessary to rely on the intelligent computing center to build a large model of the automobile industry.

1. Large model "boarding"

The large model currently referred to is the abbreviation of "large-scale deep learning model", which refers to a machine learning model with a large number of parameters and complex structure, which can handle large-scale data and complex problems, and is mostly used in natural language processing and computer vision. , speech recognition and other fields.

Previous traditional machine learning models were small in scale and could only handle small amounts of data, for example, only one data modality, such as text, images, or audio. Deep learning models can contain millions of parameters and process massive amounts of data. However, large models can reach tens of billions of parameters and require supercomputers for training.

And because the large model can more accurately express the data distribution and learn more complex features, it can improve the accuracy of decision-making.

The Transformer architecture is the basis of the mainstream algorithm architecture in the current large model field, thus forming two main technical routes of GPT and BERT. After the release of GPT3.0, GPT has gradually become the mainstream route for large models. At present, almost all large models with a parameter scale of more than 100 billion adopt the GPT mode.

Screenshots of the interface that has been iterated to GPT4 are on the official website of "openai"

Large models will play a more important role in cross-modal applications, especially in large-scale scenarios. "Cars" that are undergoing intelligence and manufacturing innovations have many large-scale model application scenarios that can expand unlimited imagination.

Specifically, in the field of new energy vehicles, the smart cockpit is the most convenient application scenario for large models. The application of large models in the field of in-vehicle AI communication and dialogue can change the situation in the past that the intelligence of in-vehicle AI voice is generally weak and the experience is not good.

If in the past there was a "command relationship" between the car owner and the AI assistant, where the car owner assigns tasks and the AI assistant passively executes them; Car owners communicate and output content to people. For example, today's in-vehicle ChatGPT voice assistant can already handle complete conversations, such as asking questions, and can maintain an understanding of the context and form a relatively good voice interaction experience.

GAC has announced that its AI large-scale model platform can better understand user intentions, realize deeper personalization and "human-like" interaction, and it is a "mobile encyclopedia", knowing astronomy from top to bottom and geography from bottom. Not only has "super intelligence", but also has "high emotional intelligence", which can change the mechanical voice interaction and upgrade the form of "one question and one answer" to an intuitive and natural dialogue without threshold.

GAC AI Large Model Platform Interactive Interface Image Source: Bitauto

Xunfei also announced that the combination of its "Spark Cognition" large model and the smart cockpit can realize the free communication between people and vehicles in the car across business and scenes.

Second, or change the auto industry

If extended to the entire industry level, it can be predicted that the large model may change the entire new energy vehicle industry.

The large model can not only change the smart cockpit and improve its sense of experience, but also promote the smart driving that is currently in full swing. In this regard, Li Keqiang, academician of the Chinese Academy of Engineering, professor of Tsinghua University, and chief scientist of the National Intelligent Connected Vehicle Innovation Center, said that the advantages of large models in processing text, acquiring and processing data, and establishing scenarios for training and iteration will be beneficial to human-machine development. Interactive intelligence and intelligent driving play an accelerating role.

Large models can process massive amounts of data, and can improve the accuracy and performance of the model, which has a great effect on intelligent driving. Smart cars are inherently a natural data producer. The road network, road conditions, environment, and driving behavior are all massive data. Processing these data and multi-dimensional analysis capabilities through large models can not only continuously optimize the model, but also Improve the accuracy and reliability of intelligent driving.

In the past, intelligent driving perception used the "stacking" method of various small models. The principle of recognition is to look at it first, and then compare it in the knowledge base. If you have not learned it before, you may not be able to recognize it accurately. The key role of the self-learning algorithm ability of the large model in the automatic driving process of the car is mainly reflected in the perception and decision-making levels. It can reason when dealing with complex road conditions and the ability to think like a human.

Therefore, industry insiders believe that 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 large model will also have a new impact on the research and development methods and business models of new energy vehicles.

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 development cycle is greatly shortened, and the multi-modal (visual, voice, gesture, etc.) rich data , can further improve the overall R&D efficiency and reduce R&D costs.

Li Xiang of Ideal Automobile once said: "In the past, we had to manually calibrate about 10 million frames of self-driving images a year, so we hired many outsourcing companies to calibrate, about 6 yuan to 8 yuan per piece, and the cost per year Nearly 100 million, this is only for image calibration of automatic driving. But when we use the large model of software 2.0 to perform automatic calibration through training, the results and effects will be terrible. What used to take a year to do, Basically, it can be completed in 3 hours, and the efficiency is 1000 times that of human beings. The work in this field is completely different.”

For automobile developers, how to cooperate with the learning ability and generalization ability of the large model to form a more efficient research and development method is one of the current core issues. In the future, large models have the potential to revolutionize design, engineering, and manufacturing in the car-making process.

As far as the business model is concerned, after the large model is "boarded" through the smart cockpit, the vehicle-mounted large model will have a better understanding of the preferences and habits of "people" through the interaction between people and vehicles, which will inevitably generate new commercial value.

Third, start a hundred "model" war

The application of large models in the field of smart cars has become the focus of collective attention of the entire pan-auto industry. Many car companies have begun to "get on the car" with large models. Some media even said that the new competition for large models to "get on the car" Shots fired.

Recently, GAC Group announced the official launch of the latest research and development results of AI large-scale model technology - GAC AI large-scale model platform. Geely Automobile also stated that the first full-stack self-developed full-scene AI large model will be carried on the Geely Galaxy L6. Prior to this, Li Auto also released the large model algorithm MindGPT. At Huawei's nova11 series and full-scenario new product launch conference, Yu Chengdong announced that the AITO M9 will be equipped with AI large models, and Xiaoyi's smart assistant will be able to bring users the industry's strongest in-vehicle AI experience.

In addition to car companies, some automotive supply chain companies also develop large-scale models, such as the Ririxin large-scale model released by SenseTime; Momo Zhixing released a large-scale model DriveGPT-Xuehu Hairuo.

Car companies not only independently develop large models, but also adopt a cooperative approach to apply large models. For example, Baidu Apollo announced that Great Wall Motors and Yikatong Technology (in no particular order) have become the first batch of Wenxin large-scale smart cabin application exploration partners to explore smart cockpits and autonomous driving. Geely Automobile, Zhiji Automobile, Chery New Energy Automobile and many other companies also said that they will cooperate with Alibaba Cloud in large-scale model-related scenarios.

The application of large models by car companies is not only applied to smart cockpits, but also has begun to explore large models in the field of intelligent driving. A typical example is Xpeng Motors and Ali’s large models to create an autonomous driving AI intelligent computing center "Fuyao", which is used for autonomous driving models train. He Xiaopeng, CEO of Xiaopeng Motors, said that "Fuyao" is currently the largest intelligent computing center for autonomous driving in my country's auto industry, which has laid a computing power foundation for Xiaopeng's training in the all-scenario intelligent assisted driving system. The "Snow Lake Hairuo" released by Momo Zhixing is the industry's first self-driving generative large-scale model.

Car companies' enthusiasm for large models is even reflected in trademark registration. For example, Chery New Energy Automobile has registered the "ICAR GPT" trademark, Great Wall Motor has registered "Great Wall Motor SPACEGPT" and other trademarks, and Weilai Automobile and Xiaopeng Automobile have also registered trademarks such as "ICAR GPT". Apply for the registration of GPT related trademarks.

In the second half of intelligentization, some people regard a smart car as a "large mobile phone", so the actions of car companies in the field of large models can be seen as a competition for the "entry" of a new generation of smart terminals. However, some people have said that such a rush of car companies to lay out large models is too hot, and there is suspicion of rubbing the heat.

At present, the popularity of large models has decreased month-on-month. According to the latest data released by the network analysis company Similarweb, in June ChatGPT’s global visits saw its first negative month-on-month growth since its launch, with a drop of 9.7%, and the month-on-month growth rate also gradually declined from January to May.

Fourth, positioning "infrastructure"

The overheating of large models is not limited to the automotive industry. Recently, Huang Tiejun, the dean of Beijing Zhiyuan Artificial Intelligence Research Institute and a professor at the School of Computer Science of Peking University, said: "The one thing that currently restricts the development of the domestic large model industry is 'too hot'."

The "China Artificial Intelligence Large Model Map Research Report" shows that as of May 28 this year, 79 large models with a parameter scale of more than 1 billion in China have been released. This is because the phenomenon of low-level duplication and fragmentation in the industry is relatively serious. This phenomenon will cause the dispersion of resources, and it is difficult to form a systemic breakthrough with a major impact. Therefore, Huang Tiejun believes that the industry should avoid repeated efforts and concentrate on major tasks.

Robin Li, the founder of Baidu, has long publicly stated: "It doesn't make much sense for a start-up company to recreate ChatGPT. I think there is a great opportunity to develop applications based on this large language model. There is no need to reinvent the wheel. After having the wheel, it is possible to make a car." , Aircraft, the value may be much greater than the wheel."

Industry insiders said that after the survival of the fittest, there may only be a very small number of two or three large-scale model ecology in the world in the future.

Therefore, large models should be viewed from an infrastructure perspective. Zhu Xiaohu, managing director of GSR Venture Capital, wrote in Moments: "Don't be superstitious about the general model, because next year GPT-3.5 will become commodity (general infrastructure), and three years later, GPT-4 will also be."

For the automotive industry, large models will only become industry infrastructure. Recently, Zhang Yongwei, vice chairman and secretary-general of the China Electric Vehicle 100, pointed out in the article "Several Strategic Issues in the Development of Smart Vehicles in my country" that the intelligent stage will require more new infrastructure, including large models. He pointed out that it is necessary to rely on the intelligent computing center to build a large model of the automobile industry.

In short, the large model has achieved good commercial implementation in the smart cockpit, making the smart cockpit an important selling point of new energy vehicles. In the future, at the level of intelligent driving, the empowerment of the large model will promote its rapid development; it is predictable that the large model will promote automobile research and development, and even lead to a new business model. However, all of this may be based on the positioning of the large-scale model "infrastructure". Flocking to make large models, or even gossip, may lead to low-level duplication and waste of resources.

Full text reference

[1] "AI large models are "on the car", the prospect remains to be seen", China Business Daily

[2] "Smart cars usher in the "ChatGPT moment"? ", China Electronics News

[3] "Large models are going to the automotive industry, and they are the first to land in the smart cockpit? ", Intelligent Relativity

[4] "From a small task-oriented assistant to a hexagonal co-driver, the new competition of AI large-scale models "on the car" has been fired", Auto Market Ruijian