Written by / Tu Yanping
Editor / Zhang Nan
Design / Teacher
On the day OpenAI's chatbot ChatGPT came out, the gears of AI destiny began to spin faster.
In less than a year, domestic AI models have sprung up and are difficult to count.
The Interim Measures for the Management of Generative AI Services came into force on August 15, 2023, clarifying for the first time that generative AI services are included in the management. Up to now, more than ten large models such as Baidu Wenxin Yiyan, Douyin Skylark, iFLYTEK Xinghuo, and Tongyi Qianwen have passed the first batch of filings.
By filing and providing services to the public, this means that the large model is about to enter the stage of large-scale promotion.
For example, 360 revealed in its 2023 half-year report that 360's cognitive general AI large model "360 Wisdom Brain" has generated nearly 20 million yuan for the company.
The future development direction of AI big models must be vertical, specialized and industrialized, empowering thousands of industries. In the automotive field, the application of large models is also accelerating.
Large models get on board
In June, Mercedes-Benz announced that it would integrate ChatGPT into its MBUX voice assistant through Microsoft's Azure OpenAI service. More than 900,000 U.S. car owners can participate in the MBUX intelligent human-computer interaction system test program through the Mercedes me mobile app or voice "Hey, Mercedes".
GM is also working with Microsoft to explore the use of ChatGPT in vehicles, trying to develop automotive application models on the model architecture of ChatGPT, and will integrate functions such as autonomous driving, vehicle management and human-computer interaction in the future.
"ChatGPT will permeate everything. This shift isn't just a single feature like the evolution of voice commands, but means customers can expect their future vehicles to be more robust in terms of emerging technologies. GM Vice President Scott Miller has said.
In China, many large models have begun to get on the car or announced the car plan.
In November 2022, Baidu and Geely released the industry's first knowledge-enhancing model for the automotive industry, Geely-Baidu Wenxin, with application scenarios including intelligent customer service systems, in-vehicle voice systems, and domain knowledge base construction. According to Baidu's data, the manual evaluation availability of the intelligent customer service knowledge base expansion task of the model jointly built by the two parties increased by 24.37% compared with the baseline. The effect on generation tasks such as short answer generation of in-vehicle voice system and knowledge base construction in the automotive field has been improved by 20%-35%.
In addition to Geely, car companies such as Changan, Great Wall, Hongqi, Lantu, Haima, and Dongfeng Nissan also announced the connection to Wenxin.
In September this year, HUAWEI CLOUD released a large model of Pangu Automobile, covering business scenarios such as vehicle design, production, marketing, and R&D. Pangu can shorten the cycle time for autonomous driving to learn new complex scenarios from more than two weeks to two days. In the commercial car scenario, it can simulate scenes such as dusty mining environment, long slopes up and down, and large curvature turns and automatically annotate, and adapt to new heavy-duty truck models in 4 months.
More professional vertical large models in the automotive field are also emerging.
For example, in April this year, DriveGPT Snow Lake Hairuo was released as a generative model for autonomous driving. Gu Weihao, CEO of Milli Mobility, believes that the training and deployment of autonomous driving perception and cognitive algorithms based on large models will make the realization of end-to-end autonomous driving possible.
DriveGPT input fusion perception scene sequence, output generated prediction scene sequence, related technical achievements will gradually land on the new Mocha DHT-PHEV equipped with Milli HPilot 3.0, Weipai Blue Mountain DHT-PHEV and terminal logistics automatic distribution vehicle Xiaomo Camel 3.0.
For example, in August this year, Ganymede released a large model of car safety pendant "Butterfly", positioned as a co-driver of an automotive safety engineer. Based on the standard Transformer structure, Butterfly trained a 13 billion parameter large model on the training data of dozens of high-quality tokens related to automotive exception handling.
Baidu, Huawei, Tencent, Milli Zhixing, Qingzhou Intelligent Navigation, Spichi, SenseTime, Yunzhisheng, iFLYTEK, Megvii Technology, China Science and Technology Chuangda and many other large model R&D enterprises are cooperating with car companies to promote large models on the car to provide more intelligent and efficient automotive products and industry solutions.
Three paths
To promote large models on the car, some car companies choose to cooperate with suppliers, and some car companies choose to self-research.
The Ideal Car's spatial algorithm team had already started training the entire large-language model before the release of ChatGPT. In June this year, Li Auto released its self-developed cognitive large model Mind GPT at the Family Technology Day held at the Changzhou base.
Later, Ma Donghui, president and chief engineer of Ideal Automobile, revealed that in terms of research and development, they already have the full-stack self-research capabilities of Mind GPT. For pedestal model training, they used 1.3 trillion tokens, and the first version used 16 billion parameters.
In July this year, Geely Automobile revealed that it will launch self-developed AI models, including painting models, music models, language models and autonomous driving models. Geely Galaxy L6 is already equipped with AI intelligent interaction technologies such as external voice interaction, AI music MV, AI children's picture book and WOW wallpaper.
Chery Automobile's choice represents the third path.
On August 17, Chery and iFLYTEK jointly held a press conference, announcing that the Spark model will be the first to ride the Star Age ES model, and also released the LION AI large model platform. Star Era ES will be launched in the fourth quarter, when it will be enabled by the Spark cognitive big model and the LION AI big model platform.
In the future, more car companies may follow the same path as Chery, not only having their own large models, but also not excluding the large models of suppliers.
Great Wall Motor is the first batch of intelligent cockpit application exploration partners of Baidu Wenxin Big Model, and works with Baidu Apollo to create large-model artificial intelligence interaction based on intelligent car scenarios, and at the same time, it is also accelerating the creation of its own large-model service capabilities.
In September, Great Wall Motor announced the establishment of AI Lab, an intelligent frontier organization under the technical center, based on computing power, algorithm and large model capacity building, to pull through the power, chassis, styling, upper body and other organizations to build Great Wall Motor's large model service platform. Great Wall Motor's AI model will be installed on its models.
According to the planning of the Great Wall, the future large model technology will be widely used in cockpit function development, intelligent driving, and even the use of AIGC to do model design. The head of the organization department of AI Lab is Yang Jifeng, the former head of salon brand intelligence, and reports directly to Wei Jianjun, chairman of Great Wall Motor, and has fixed weekly meetings every week.
The state of chaos is to be solved
The application of large models in automobiles is still further expanding, and the future may span all links such as automobile research, production, supply and marketing.
The large model of the automotive industry is also developing in a more subdivided direction, not only the intelligent cockpit model, the intelligent driving training model, but also the intelligent customer service model, the safety model and other subdivided models.
However, a groundbreaking technology can also be a double-edged sword. While large models improve driving safety, there is also the possibility of increasing the risk of privacy leakage.
At present, the large model on the car is still in a chaotic state, and it is urgent to have unified standards and theoretical foundations to guide the development of the industry.
In May this year, the state officially launched the standardization of large models, and set up a large model task force in the overall group of artificial intelligence standardization. The large model task force will undertake the standardization of large models, with the aim of promoting the practical integration of large model technology and standardization.
In July, Geely Automobile, together with the China Academy of Information and Communications Technology, took the lead in formulating the standard for the "Evaluation Method for Large Models in the Automotive Industry".
A consensus reached by all parties is that the future automotive industry big model standard will set standard indicators for multiple application scenarios such as intelligent cockpit, vehicle-road coordination, assisted driving, and automotive knowledge base from the future automotive industry to comprehensively evaluate the support and application efficiency of the large model for automobile R&D, production, marketing, use, after-sales and other links.
Recall that in 2021, when the metaverse became popular, car companies have registered metaverse trademarks. Nowadays, large models are on the car, and car companies have begun to register large model trademarks.
Popular things are always alarming. Whether the large model is a short-term grasp for product marketing or a weapon to completely revolutionize the travel experience remains to be seen. (The title image is automatically generated by Baidu Wenxin)