Rebuilding a car with AI, you can’t just rub the heat

"Access to DeepSeek" was the hottest word in the car circle in the past few months, but it was the hottest word in the car circle. Now that the noise has subsided and returned to reason, it is time to seriously discuss how AI will reshape the car.

Let’s put a conclusion here first:AI may be the key point of qualitative change of automobile intelligence.

The reason lies in the evolution direction of several core plates of automobile intelligence, all pointing to the ability demand of AI big model: intelligent cockpit should be able to provide more personalized customization or active service; The "end-to-end" required by intelligent driving requires faster and more flexible understanding and decision-making ability; As the bottom support, automotive electronic and electrical architecture is turning to domain centralization and central computing in an all-round way, which requires centralized and flexible scheduling of computing power.

The characteristics of AI big model are also the key solution: the universal base provides a computing base combining end and cloud, and the big model is good at coordinating multi-party information and making quick decisions, especially in the face of common travel scenarios, which can form solutions and give direction suggestions and references in the face of special situations such as corner case.

This has made clear the key to the next generation competition of automobile intelligence, and it may even be the biggest change that automobile, as a tool for human travel, faces in a century of development:AI will become the infrastructure of the next generation of smart cars.

However, it will still take time for AI to reshape the car: the construction of infrastructure is by no means a day’s work. If there is no fundamental integration and transformation of the underlying technology, just like this wave of car companies focusing on marketing, it is neither decent nor honest to make a simple dialogue interface between people and cars, or even to fool directly with a web browser.

Going around, the serious evolution from the beginning of infrastructure depends on the determination and thinking of the giants.

On March 3rd, Geely held a conference on AI intelligent technology, which revitalized the seemingly irrelevant "idle chess" in various fields, demonstrating the first in-depth thinking result of AI as a smart car infrastructure.

Dimension reduction from the beginning of infrastructure

Geely may be the first car company in the automobile circle to seriously discuss the construction of AI infrastructure.

Compared with the "grand strategy" that friends like to announce in the AI field, Geely spent nearly one-third of the conference space, and invited Yin Qi, an AI industry expert and chairman of Qianli Technology, to share the basic judgment on the development of AI big model from the perspective of automobile industry as a member of Geely Technology Ecology, based on objective technical facts.

There are a lot of dry goods. Let’s first look at the final answer of the evolution of AI+ cars given by Geely.

Geely called it "the birth of supernatural human-computer interaction". The logic behind it is that, unlike PC as the core carrier of digitalization and mobile phone as the ultimate entrance to the Internet, AI needs more energy consumption and more access to real information, so cars naturally have the advantage of becoming a "mobile terminal" in the AI ? ? era. The most direct change brought about by AI’s natural "multimodal" interaction and reasoning ability is to connect various services in series across user travel scenarios.

Imagine that in a go on road trip, vehicles can take over the whole driving control according to the accumulation of your usual daily habits, and intelligently help you book the destination food in advance, intelligently recommend different services to couples, family, friends and elders during the trip, and even link with your mobile phone after the trip to intelligently generate travel notes, which can be shared with one click … ..

The so-called "supernatural human-computer interaction" is not a subversion of starting a new stove, but a one-way instruction of traditional people to equipment. It is also a two-way suggestion and feedback between people and equipment, and finally people can communicate with each other and even understand your feelings better.

With a clear goal, starting with the end, we can find the greatest common denominator of infrastructure.Geely’s infrastructure advantages in AI big model can be understood from two key words: computing power+general model.

Here, a few figures can be used to visually show Geely’s achievements: Geely’s "Smart Car Computing Alliance" is based on Star Smart Computing Center 2.0, and its comprehensive computing power has increased to 23.5 EFLOPS. According to "China Computing Network Market Operation Status Survey and Investment Strategy Forecast Report (2023-2030)",By the end of 2023, the scale of AI in China will reach 200 EFLOPS, and Geely’s AI computing power layout is enough to rival nearly 1/10 of the country.

Sufficient computing power corresponds to Geely’s higher goal of building a universal base model.

What is common in the industry is the vertical intelligent driving model, which is short, flat and quick, and easy to produce results. You can understand the difference between the two models like this:Vertical intelligent driving model is a kind of intelligent driving technical scheme that integrates all kinds of road driving scenes, and emphasizes the familiar ones such as laser radar and visual perception, and its relative sources are also limited.

But it is different from the information input and feedback of people’s real driving. People are natural "multi-modal input". You will form a trend judgment according to the change of traffic flow ahead, the command of pedestrians or traffic police, and even the seemingly unrelated scene changes on the road ahead, and make a subconscious reaction and judge the next road conditions and operations.

The universal base model is born for this multi-modal nature, but the exponential explosion of its data volume means that it needs stronger computational support.

An intuitive data comparison is that the specifications of the intelligent driving model widely used in the automobile industry are about 10 million video clips, and a large number of repeated scenes will be mixed in the middle, while the data volume of the universal base model can easily reach 1 billion, even for similar scenes, it will identify the internal trend changes more carefully. Only in the recognition degree of the general obstacle, compared with the industry model, the general model is the gap between "100,000 categories" and "50 categories", and even can identify new obstacles through reasoning, forming the ability of independent learning and decision-making.

To put it bluntly, the universal base model is essentially observed in a mode closer to the real human thinking, and the result is closer to the real human road judgment and driving mode, and truly realizes the anthropomorphic "fully automatic driving" above L3.

On this basis, in order to better understand the real world, Geely and Eco-Partners have created a model matrix that integrates language model, multimodal model and pan-world model. Among them, the language model represented by DeepSeek is aimed at the smooth interaction between human and computer, and the rapid understanding of the world abstract information into words by intelligent driving.

The multi-modal model is represented by Geely’s AI ecological partner Step Star, which can understand and generate images, music and videos. In the end, the foundation of the future industry includes VLA and a world model that can understand the objective reality. Geely has simultaneously laid out the AI Drive model at this level, laying a good foundation for intelligent driving above L3.

According to the computing power demand of the common large model in the industry, the universal base model needs at least tens of thousands of cards for a single training, so this is a "dimension reduction blow" that a "head player" like Geely can achieve. Looking back at Geely’s "Smart Geely 2025" plan put forward at the Longwan conference in 2021, it was now curious and questioned by many people at that time. Geely’s heavy investment and extensive layout in the digital field now have a new answer.

The competition for "AI+ cars" has just begun.

Why is it called a thousand miles?

In the matter of AI reshaping the car, Geely played its own stubbornness as building a car:Starting from the bottom structure, we should pursue the accumulation of wealth and thin hair in the final landing. Only in this way can we meet the goal of "high value" and achieve the blow of dimension reduction.

At the press conference, as the first landing of the AI strategy, Geely released a self-developed "thousands of miles" intelligent driving system, and took the lead in carrying it on two cars: the 2025 modified version of Geely Yinhe E8 and Geely Yinhe Xingyao 8EM.

As a unified intelligent travel solution formed by Geely through AI technologies such as end-to-end, VLM and world model, Qianli Vast is divided into five different levels, namely H1, H3, H5, H7 and H9, with a minimum standard computing power of over 100TOPS.Highest H9 is the industry’s first L3 intelligent driving solution with mass production landing capability, which will be carried on all Geely models.

The most intuitive benefit brought by the vastness of thousands of miles is that users no longer have to choose between the inconsistent experience of intelligent driving and the complicated packaging words of different manufacturers. At the turning point when industrial technology enters the popularization period from the rising period, it is obviously much more important to let more people use it with low threshold than the dazzling skills and packaging of high-end technology.

Geely named this intelligent driving system "Qianli", and the strength behind it is the industry-leading general model and computing power. This is one aspect of technology, and the vastness of the other side emphasizes that this intelligent driving system is not only rich enough to support intelligent driving in various scenarios, but also stable and reliable. As the "safest intelligent driving system", Qianli Vast can be integrated with Geely’s technical standards in the field of car building.

At the press conference, Geely showed a typical scene: the AEB of 120km/h stopped automatically. As one of the most risky dangerous driving scenarios that most users may encounter, compared with many brands who still use this ability as a gimmick for high-level intelligent driving, Geely can directly declare the vast expanse of thousands of miles as a standard, and even avoid the risks in the same lane, and can also avoid the risk scenarios that appear continuously in two lanes.

Qianli Vast integrates the self-research power of Geely’s ecology, which is essentially based on the brand-new anchoring of the industry layout for many years, the VLM model of vision and integration of language, and the universal model base based on the understanding of the real world, and on this basis, it promotes the first mass production of L3.It indicates that China enterprises, represented by Geely, are setting up a real moat in the field of intelligence, especially in terms of technology and experience standards integrated with the whole vehicle, which means that they are ready to promote the maturity of laws and regulations and promote them globally.

This is the competition for upgrading the dimension after laying the foundation first. Compared with most enterprises in the industry, which are still using smart driving to create profit space and avoid responsibility, Geely has made the vast expanse of thousands of miles into the best and most time-tested value in the century-old automobile industry: standard.

AI defines the first player of a car.

Geely’s years of development actually follows similar thinking:Under the grand strategy, the grey line of grass snake will be continuously laid out, and then it will be applied first when the industry changes its qualitative point, so as to realize the great leap and the establishment of moat.

From the acquisition of Volvo, the layout of smart Geely 2025, the layout of vast architecture, and the release of Ningbo Declaration and Taizhou Declaration, Geely AI technology conference marks the first time that Geely has completed the first transformation to "AI defines cars".

The industrial leap of any technology must be the first move of the explorer and the follow-up with the industrial giants, so as to jointly promote the application for thousands of households. The entry of an industrial head enterprise like Geely marks the seeds of innovative technology brought by DeepSeek or other AI ecological partners, with the soil to land.

Similarly, they all need to go through the process of long-term exploration in core technologies and bear huge sunk costs and risks; What a giant like Geely has to do is to discard the false from the true as soon as possible and complete the popularization and application of engineering. In this process, advanced technology has to find true value for users and popularize more scenarios, which is bound to be a long-term process.

But this may also be the "high-value" experience that users can reach at their fingertips as soon as possible: once AI builds a car, it will not be far from the mature products such as "It just works" once it enters the unknown technical engineering.