On December 27, Li Auto introduced the '2024 Ideal AI Talk' for three consecutive days, sharing their latest insights on artificial intelligence, including advancements in smart driving and the Ideal Classmate AI technology. Li Xiang announced that the Ideal Classmate, based on the self-developed Mind GPT large model, has moved from the in-car system to smartphones, with the app fully launched on December 27. Additionally, Li Auto announced that the OTA 7.0 version of its in-car system will be fully pushed to AD Max users by the end of December, featuring AI inference visualization, high-speed end-to-end capabilities, and upgrades to Mind GPT-3o and Mind Diffusion V2.0.
During the three-day livestream, Li Xiang announced that Li Auto will transform into an AI company. Specific details can be found through the provided links. On the second day of the livestream, Li Xiang and the autonomous driving leader, Lang Xianpeng, discussed the development trends of Li Auto's autonomous driving. Following the current end-to-end + VLM system iteration, it is expected to achieve L3 level autonomous driving by 2025. Further details can be found through the link.
● Ideal Classmate and Autonomous Driving Are Li Auto's Two Core AI Products
Li Xiang, Chairman and CEO of Li Auto, stated: 'Ideal Classmate and autonomous driving are often viewed as independent fields. Our large language model, Mind GPT, is cognitive intelligence connecting the digital world, while autonomous driving is spatial intelligence related to the physical world. We are exploring both fields simultaneously and firmly believe that the combination of cognitive and spatial intelligence—what we call VLA (Vision Language Action Model)—is a more promising and achievable opportunity.'
● L3 Supervised Autonomous Driving: Not an Extension of L2 Assisted Driving, but a Precursor to L4 Autonomous Driving
In different stages of autonomous driving, L3 refers to supervised autonomous driving, which is not an extension of L2 assisted driving, but a precursor to L4 high-level autonomous driving. Assisted driving only implements specific functions, while autonomous driving involves overall capability. Traditional L2 assisted driving relies on previous-generation autonomous driving solutions, executing smart driving functions based on preset conditions in different scenarios, but unable to handle all corner cases. Li Auto has developed an end-to-end + VLM dual-system solution, using artificial intelligence to improve autonomous driving capabilities, continuously iterating and enhancing with the Scaling Law to adapt to all driving environments.
With the continuous iteration of the end-to-end + VLM dual-system, Li Auto aims to achieve L3 supervised autonomous driving by 2025 and provide users with an integrated, end-to-end product. As of December 25, Li Auto's smart driving total mileage reached 2.9 billion kilometers, and training computing power increased to 8.1 EFLOPS.
● Electric Vehicles Are Not the End of Li Xiang's Entrepreneurial Journey
Li Xiang believes that after many years of development, the competition with traditional car manufacturers may have ended, and many new entrants have emerged. Initially, the competition was between new forces and traditional automakers, but now companies like Huawei and Xiaomi have entered the game, changing the competitive landscape. This is what makes the world so interesting and rich.
● What About Xiaomi's Electric Car? Did You Give Lei Jun Any Advice?
Li Xiang: I told him 'You must go all in,' and if Xiaomi follows this path, the electric car will succeed. Lei Jun is very skilled at hardware, which is indisputable. He doesn't just make good cars; his TVs and air conditioners are also excellent, which is his inherent advantage, and he approaches these projects with a passionate mindset. We have a good relationship with Xiaomi, and Lei Jun helped us a lot by supporting our Ideal MEGA and L6 models. We are very grateful for his help.
Full Q&A Transcript:
01. The 'iPhone 4 Moment' Comes in the Agent Stage
Zhang Xiaojun: When did you first use ChatGPT, and how did it feel?
Li Xiang: I used it when it was released. My biggest impression was that it looked like what AI should be.
Zhang Xiaojun: If you were the CEO of OpenAI, would you do a better job than Sam?
Li Xiang: No, I think Sam Altman and the team have done an excellent job.
Zhang Xiaojun: If you were the CEO of OpenAI now, what would you do?
Li Xiang: Today, OpenAI is defining the first stage of AGI (Artificial General Intelligence): the chatbot. I believe OpenAI has done the best in providing this product. The second stage is the reasoner, and in the third stage, the Agent (AI agent) will mark the 'iPhone 4 moment,' where ordinary people can use it, completing tasks independently and continuously without relying on intensive prompts. The interaction method at that time should be considered by all leading companies.
02. Making AI Interaction as Natural as Human Communication
Zhang Xiaojun: Why did an automaker decide to develop its own large model? How was this decision made?
Chen Wei: It was a gradual consensus. By the end of 2022, we had already transitioned to a pre-training model for natural language processing tasks. This allowed us to quickly and efficiently cover tasks such as car control, media, and navigation. After seeing the rapid advancements of large models, we were deeply inspired. Initially, we didn't consider creating such a large model, but later Li Xiang suggested we should focus on enhancing the cognitive intelligence of Ideal Classmate, raising the ceiling. This guided our future work on foundational models.
Zhang Xiaojun: As a latecomer, how do you plan to catch up with ChatGPT?
Chen Wei: OpenAI is the industry's benchmark, and most teams are still in the L1 stage (chatbot). However, we are playing an infinite game and will focus on the first principles Scaling Law to ensure rapid iteration. Our Mind GPT model has undergone more than 30 iterations since the release of OTA 5.0 in December 2022.
Zhang Xiaojun: How has Mind GPT evolved?
Chen Wei: Mind GPT has gone through three generations. Version 1.0 was released in April 2023, and by the end of 2023, the OTA 5.0 update pushed this large model to in-car systems. In mid-2023, we launched version 2.0, optimizing both model performance and inference efficiency. Mind GPT's architecture will continue to evolve with a mixture of experts (MoE) and Transformer structure.
Comments0