Click to follow
Text｜Yao Yue, editor｜Wang Yisu
"Although we have been building cars for more than 30 years, we are now facing a new problem and challenge." Yang Jifeng, head of Great Wall Motors AI Lab, said, "In the AI era, every problem is an AI problem."
The AI Lab that Yang Jifeng is responsible for is a new department recently established by Great Wall Motors. It is centrally responsible for "delivering" products and technologies based on AI large models to the entire vehicle field and R&D field of Great Wall Motors.
Great Wall Motors has been laying out intelligence for many years, including the basic base of algorithms, computing power, data, and technology brand Coffee Intelligence. However, this time the AI Lab is an organizational system change that "coordinates" past AI achievements and is Great Wall Motors' first choice. A major strategic upgrade in automobile intelligence.
Coincidentally, just last month, new power Xpeng Motors also announced that it plans to establish a company-level intelligent R&D, planning and operations team in the next step. And according to the latest news from 36Kr, Ji Yu, former vice president of the Internet Center of Xpeng Motors, has rejoined the company in a low-key manner, and Xpeng Motors’ intelligent department is already planning organizational adjustments.
Recently, the intelligentization of automobile products has become increasingly intensive. AITO Wenjie has launched the most important OTA system upgrade since its launch; Weilai released the intelligent system Banyan 2.0.0 CN; BYD’s latest Seal DM-i has also been significantly upgraded in terms of intelligence.
A traditional car company that has been manufacturing cars for more than 30 years and a new force that has been established for 9 years have made major organizational adjustments towards intelligence. The weight of intelligence in automobile products is increasing, and the intelligence of the automobile industry has ushered in a new stage.
01 In the era of big models, development model upgrades force organizational changes
"When we deliver the car to the user today, it is not an end, but a beginning." Yang Jifeng said that in the AI era, the car development model will be completely changed.
In the past, car development was 95% completed in the SOP (product preparation for mass production) stage, and the rest was fine-tuning. Therefore, the development model is aimed at vehicle model delivery, with one project team responsible for one vehicle model.
However, as users have strict requirements for the OTA (over-the-air update download) capabilities of cars, the one-car-one-project development model cannot meet users' needs for rapid changes.
In the era of large models, this development model is destined to end.
Not only because terminal applications change faster, but also because the most prominent ability of large models is generalization capabilities, which requires a high degree of collaborative operations among various teams. Moreover, in the era of large models, vehicle data is needed to simultaneously support multiple business lines such as autonomous driving AI development, cockpit space AI development, and even intelligent power and energy consumption management in parallel.
Therefore, the platform-based development model emerged as the times require.
Great Wall Motor's AI Lab integrates Great Wall Motor's original voice, recommendation algorithm, autonomous driving and other teams to collaboratively develop a technology base based on large AI models, providing a large-model universal solution for Great Wall Motor's vehicle end and R&D end. solution.
The AI Lab is more like a "pump" that delivers the "blood of large models" to Great Wall Motors' entire body.
"There will be an integration trend in vehicle development." Yang Jifeng said, "The next organizational changes in the industry will basically discuss computing architecture, data systems, infrastructure, etc. as separate topics. We are also doing this today. .”
However, Great Wall Motor’s AI Lab is not completely defined as a middle office.
The middle office provides data capabilities, computing power, and system capabilities; the front office is responsible for the application of algorithm capabilities in intelligent spaces, including voice experience, visual experience, multi-modal experience, etc., as well as the application of intelligent transformation in enterprises.
"AI Lab does not tend to only be a middle platform. The premise of a middle platform is to have several front desks to provide peer-to-peer AI development capabilities." Yang Jifeng said, "We hope to integrate the entire AI data capabilities, computing capabilities, algorithm capabilities and application scenarios in Solved end-to-end within an organization.”
With the establishment of the AI Lab, Great Wall Motors' intelligent system is more completely presented.
The AI large-model technology platform responsible for AI Lab is the base of the entire intelligent system. On top of the base are the two main components of Coffee Smart Cockpit and Coffee Smart Driving built around the technology brand Coffee Smart, as well as Coffee Smart Cloud Platform and Coffee Electronics. Architect two auxiliary platforms.
Among them, the Coffee Smart Cockpit operating system is about to be upgraded to the third generation, and the Coffee Smart Driving software system includes high-speed NOH, urban NOH, automatic parking, memory parking, etc. On the vehicle side, AI Lab's large model technology will be widely used in cockpit function development, intelligent driving upgrades, etc. in the future.
02 How does AI redefine cars?
"The large model makes us feel that we have really entered a new AI era." Yang Jifeng said.
In the past 4-5 years, Great Wall Motors has successively established algorithms, data, and computing capabilities, but it has always focused on AI development and task-solving capabilities in a series of specific scenarios. The generalization ability of large models has allowed Great Wall Motors to see new AI technologies. opportunity.
How does AI define cars? Reflected in smart cockpit and smart driving.
The smart cockpit is where consumers are currently most aware, and it is also where car companies are looking for differentiation.
Yang Jifeng believes that "the concept of smart cockpit was first proposed, and it was more like a functional machine similar to a mobile phone. This does not involve AI issues, but when it is regarded as an intelligent space, AI issues are involved." This AI issue is , can machines form a driving space like humans?
One of the most important issues that smart cockpits solve is driving safety.
Relevant statistics show that there are more than 100,000 road traffic accidents directly caused by fatigue driving in my country every year, causing more than 90,000 people to be seriously injured or killed.
For fatigue driving, the smart cockpit is equipped with a DMS (driver monitoring system) that can assess the driver's fatigue level. However, the current evaluation logic is to judge the degree of fatigue based on the degree of eyelid opening and closing. That is, the smaller the eyelid opening and closing, the higher the degree of fatigue. But there are many problems. Is my eyes tired? Is squinting under bright light considered fatigue?
When a large model is used to assess fatigue driving, multiple information such as the driver's lane changes, facial features, and body movements can be sensed, and then the historical data can be compared to determine whether the driver is fatigued. The result will be more accurate. Then, the large model will also recommend a wake-up mode that is more suitable for the driver.
In addition to driving safety, another aspect that smart cockpits strive for is "personality and thoughtfulness." By applying AIGC to the smart cockpit, the driver can not only have Vincent's AI assistant, but also Vincent's drawing assistant, as well as large model + knowledge base consulting experts.
"The large model of Wenshengtu is a user experience in the short term, and an evolution of interactive strategies in the long term." Yang Jifeng believes that AIGC can create a more personalized smart cockpit.
The smart cockpit in the era of large models is capable of more humane human-computer interaction.
In terms of intelligent driving, Great Wall Motors internally incubated an autonomous driving company, Haomo Zhixing, in 2019 ("After completing large models and using large computing power, autonomous driving will usher in the battle for cities in 2023"), Great Wall Motors is also conducting a full-stack intelligent driving Self-developed, it is one of the first domestic car companies to introduce Transformer and propose an emphasis on perception rather than maps.
Great Wall Motors has developed large-scale models such as perception and cognition, focusing on urban NOH (urban pilot assisted driving), and plans to expand to 100 cities by 2024.
Large models can already better solve the problem of output of perception tasks and the interpretability problem of path planning. However, Yang Jifeng believes that the revolution in intelligent driving brought about by large models has just begun. “Whether the cognitive emergence of large models can effectively solve the generalization problem of complex and diverse scenarios is the next important intelligence in the paradigm capabilities of large models. Driving issues.”
03 AI subverts “traditional car manufacturing”
In addition to car intelligence, car companies are also using AI to reshape the research and development process, hoping to transform productivity from the bottom up.
From the outside in, what can large models do inside car companies?
For traditional car companies, automotive industry design is one of the core competitiveness. AIGC is the first to change the design industry, and automotive design is no exception.
Design creativity is a core business secret of car companies. If outsourced, there will be a risk of leakage, so you must build your own large car design model.
Building a large car design model by yourself and inputting the design drawings accumulated over the years for training can not only improve efficiency, but also provide inspiration for design.
In addition to car appearance design, with the advent of the trend of "software-defined cars", the importance of car software is increasing day by day. Regardless of the functional level or the architectural level, the complexity of automotive software is increasing, but the efficiency of development work has not kept up at the same rate.
The report shows that software complexity has increased four times in the past ten years, while software development efficiency has only increased by 1 to 1.5 times. This problem is most acute in large modules that are becoming increasingly complex, such as infotainment systems and advanced driver assistance systems (ADAS). Compared with traditional deeply embedded software, the efficiency of developing these modules is about 25% to 35% lower.
Car companies need to invest more resources in developing software and maintaining it during the software life cycle. Using large models to generate code not only provides automatic code supplementation, but also provides intelligent error correction reminders and suggested code solutions.
In addition, the data accumulated by car companies over the years can also be activated by large models to the greatest extent. Using large models and knowledge base structures, a series of applications such as standard queries, intelligent customer service, user manuals, intelligent diagnosis, etc. can be developed.
However, it is not easy for car companies to complete the AI transformation of productivity.
Large models need to "eat" a lot of data. Although traditional car companies have accumulated a large amount of text, pictures and other data, these data cannot be used directly. This is because these large data models cannot understand it. Therefore, it is necessary to convert these texts, pictures and other data into data that can be understood by large models, that is, vectorized data to form a knowledge base structure.
However, if large models want to become production assistants for car companies, they still need a lot of "human feedback" for training. Therefore, Great Wall Motors asked designers, programmers, etc. to "train" themselves to train large models that meet the production requirements of car companies. .
"Use a sufficiently robust and generalized algorithm base to form them into everyone's assistant." Yang Jifeng said that this is what Great Wall will do next.
Once the underlying capabilities are established, the next step is to incorporate them into SOPs in the R&D process to form standard operating procedures.
For example, for designing a creative large model, first teach the large model to draw a sketch. At this time, you must pay attention to the spatial structure; then turn it into a floor plan, and at this time, you must pay attention to the previous design features and color applications; when rendering, you must consider the frame, style, etc. wait.
Currently, in the R&D process of Great Wall Motors, among 60% of active designers, 60%-70% of design sketches are generated using large models; almost all internal code development areas include tools for automatic code generation.
Yang Jifeng, head of Great Wall Motors AI Lab
"We have been making cars for more than 30 years, but now we are facing a new problem and a new challenge. In the AI era, what is a car? What is a car company?" Yang Jifeng said, "We found that in the AI era, everyone The problems are all AI problems. ”
We see that car giants represented by Great Wall Motors are undergoing a self-revolution. From smart cockpits to smart driving on the product side, to creative design and code generation on the R&D side, car companies are using AI to reconstruct themselves.
In the era of smart cars, the automobile industry must be undergoing a reshuffle of “if you don’t advance, you will retreat.”
Welcome to follow Light Cone Intelligence and get more cutting-edge knowledge of science and technology!