Image source: Generated by Unbounded AI

Li Xiang, chairman and CEO of Li Auto, said in a recent live broadcast after disappearing for nearly nine months.

As early as January 2023, in an all-staff letter, Li Xiang mentioned that Li Auto’s vision is to become the world’s leading artificial intelligence company by 2030, when ChatGPT was born just two months ago .

But Ideal has already regarded artificial intelligence as a truly important direction as early as its first strategy meeting in September 2022. Li Xiang believes that artificial intelligence should be changed from a hidden strategy to an open conspiracy strategy. Only in this way can more talents be attracted to join Ideal.

In 2022, when the new power car companies gradually recovered from the life-and-death situation and began to hit the monthly sales target of 10,000, the ideal goal became "an artificial intelligence company"-neither Becoming an artificial intelligence-driven electric vehicle company is not about becoming an autonomous driving company with artificial intelligence technology.

Different directions lead to different endings. Cars are carriers for different goals, but the role artificial intelligence plays in the entire process is different.

To become an autonomous driving company with artificial intelligence technology, the ultimate goal is to do autonomous driving. Li Xiang firmly believes that Ideal will not do Robotaxi.

He made an analogy. Nowadays, renting a house is much cheaper than buying a house, but some people will still buy a house. The core reason is that “we need high-quality companionship and create a stable, safe, and comfortable environment for our families.” Environment. "The logic of the car is the same. When a space becomes better, more efficient, better in experience, and more utilized, users want to own the space, not just rent it.

To become an artificial intelligence-driven electric vehicle company, the ultimate goal is to make electric vehicles, and return to the logic of manufacturing. Li Xiang believes that the ideal is not a car manufacturing company in the traditional sense.

Why must it be an artificial intelligence company?

Li Xiang believes that cars are the largest application of artificial intelligence in the physical world, but they cannot just stop at building cars. Hardware is the starting point, but it cannot be the end. “We cannot be a pure hardware company, otherwise we will miss many opportunities.”

The experience of starting a business many times in a row told him, “I want to choose a forest and do The biggest tree in the forest, not just a tree. "Until now, Li Auto has never added "car" after its logo.

The ideal of becoming an artificial intelligence company is not just talk.

Currently, of Ideal’s annual R&D investment of 10 billion, nearly half of its “bullets” are devoted to artificial intelligence. In addition to funding, Li Xiang, as a non-technical entrepreneur, is also making up for his perception of AI. "I attend four to five AI meetings a week - one is a meeting every other day, and one is a weekly AI meeting."

On the road to artificial intelligence, Li Xiang took an atypical path.

"The two most important subsets"

In Li Xiang's division, ideals are both It is not a pure car-making company, nor a pure hardware company. The ideal should be an ecological company like Apple, Huawei, and Xiaomi.

“Huawei is not just a carrier company. Everything behind it started when it became a relatively stable carrier. Xiaomi is not just a mobile phone company. It has IoT and its own ecosystem. There are even cars. ”

Building cars and making their own hardware, Li said, is to make the entire hardware system and performance “a little bit higher” and the experience better. But these cannot widen the gap, because the hardware can be done by suppliers.

Hardware is only the starting point and pillar, but it should not be the end point. If you want to become an artificial intelligence company, Li Xiang believes that the most important thing is to build a base model, just like in the era of smartphones, self-developed operating systems, app stores and cloud services. This requires enterprises to have huge software capabilities.

In the long run, under the premise of mastering the base model, a truly large-model product "must be able to use all equipment independently and have all services. It is not software intelligence, but real artificial intelligence." Intelligence.”

Based on the base model, Li Auto currently has two To C AI products, namely Ideal Classmates and Intelligent Driving. Li said that these are the two most important aspects of Li Auto’s AI technology. subset.

According to industry standards, smart driving and smart cockpit are two products, and they are supported by two different sets of technical architectures. Ideal classmates and autonomous driving made by Ideal are also separated according to industry standards. But Li Xiang believes that this is only an early stage and not yet its final form.

The two sets of large models currently developed by Ideal represent two different sets of intelligent systems: the Mind GPT based on the large language model is called language intelligence internally, and the end-to-end large model of intelligent driving being developed is called Internally recognized as behavioral intelligence.

In Li Xiang’s opinion, these two sets of large models will definitely be connected together one day in the future, as they “reach the same destination by different routes.”

What AI needs to test is not just a single technology, but a complete combat system from technology to products, and products to business. For AI to be like humans, it must be both smart enough and capable enough.

In Li Xiang’s view, the key to the connection between the two is that the base model can evolve into VLA (Vision, Language, Action). "When spatial intelligence realizes L4 and language intelligence realizes Agent, it will become a model."

This is because the language model cannot truly restore the physical world and requires vectors and language and recognition. Understand the three-dimensional world.

But why haven’t these two models come together at this stage?

Answer given by Li XiangThe time has not come yet. "Although everyone believes in artificial intelligence, the technology is mismatched. It uses a method that is not really consistent with artificial intelligence. But this is progress. We may be full after eating the eighth bun, but every bun in front is full. It is valuable. It is precisely because of the gradual progress that today’s artificial intelligence is relatively prosperous.”

Some time ago, Ideal has launched the official APP of “Ideal Classmates” and started to move from cars to machines. Mobile assistant. Extending the intelligent experience of Li Auto from inside the car to the mobile phone; adopting a one-stage end-to-end technology architecture, Li Li Auto’s intelligent driving has steadily entered the first echelon.

The two businesses Li Xiang mentioned refer to Li Auto’s two core AI products: Li Auto’s classmates and intelligent driving. These two businesses are the core competitiveness of Li Auto. “No matter how we subtract, Neither of these two businesses will be cut.”

Get priority

The Ideal Classmate App that has been launched is based on Ideal’s self-development. An artificial intelligence application built based on the base model. "The artificial intelligence super product built by the base model will be the entrance to a new generation - it will be on all devices and on all services." Li Xiang explained.

Why do you let your ideal classmates enter the mobile phone?

The ideal entry point is still the family, and the core is still to provide a seamless user experience.

Unlike traditional voice assistants, Ideal Classmates not only answer questions, but can also see the world. It can identify objects through visual perception, provide homework help to children, and even create emotional connections as part of the family. Li wanted to say that for many children of ideal users, the first artificial intelligence product they come into contact with is their ideal classmates. “Why don’t I also put this thing on the devices they use every day? For them, artificial intelligence is no longer a A cold tool, but an integral part of family life.”

Li Xiang believes that artificial intelligence has its own evolutionary path, and AGI (artificial general intelligence) will achieve three stages as its capabilities evolve. iterate.

The first stage is called "enhancing my abilities". It means it is an assistant, but the final decision-making power rests with the user. For example, L3 intelligent driving is called "supervised intelligent driving", which also requires users to supervise in the car for final verification.

The second stage is called "Become my assistant". Users only need to send it tasks, or even continuous tasks, and it can complete them independently and take responsibility for the results. For example, the user can ask an L4 car to help pick up the child. The user does not need to sit in the car. It can go to school to pick up the child for me, perform facial recognition, open the door and let the child get on the car, and then complete the next step. A task is to send children to learn swimming and Lego. "The better situation at this stage is that it will become a large-scale application and is the real iPhone 4 stage."

The third stage is the ultimate stage - "It is my silicon-based family." AI can continue and use human memory to make decisions autonomouslyMake decisions and take the initiative to perform multiple tasks to become an important member of the family.

Currently, many companies are competing in the smart language assistant track. The startup Dark Side of the Moon launched Kimi, ByteDance launched Doubao, Alibaba Cloud launched Tongyi Qianwen, and Baidu Wenxinyiyan was launched, Tencent launched Tencent Yuanbao, etc. There are many participants in this track, and the competition is as fierce as the automobile industry. But in Li Xiang’s view, it is still at a very early stage, because there has not yet been a “to C intelligent agent available to everyone”. Today, everyone is trying to get “AGI’s L3 and autonomous driving L4 Tickets."

At present, these smart software are highly homogeneous. What are the advantages of launching the "Ideal Classmates APP" at this stage?

Li Xiang gave three advantages: First, it is ideal to have a user base of at least 1 million and a basic user activation volume. Secondly, in terms of spatial intelligence, Ideal has accumulated a lot of papers and practices, and currently ranks first in the autonomous driving business. In terms of language models, at the beginning, the ideal positioning is to be a follower, but later on, we will want to move from the tenth place to the fifth place, and then from the fifth place to the third place.

In 2023, Li Xiang mentioned the artificial intelligence company for the first time in an internal letter. Many front-line employees did not understand this vision, and the outside world did not believe that Ideal was serious about doing AI.

“When technology changes, it is normal not to believe it.” Li Xiang said. He took ideal range extension as an example. In 2018, range extension was considered by the industry to be a backward technology, and there were very few car companies using range extension power. "When we first started to extend the range, you asked our employees, what is the success rate of this car? How many sales were there? Most employees thought that we could only sell a few hundred cars a month. This is because everyone has not experienced the new technology. I didn’t see the value that this technology could bring.”

But Ideal became a “guinea pig” and successfully put the range extender on the car, and successfully made Ideal ONE and Ideal L9 popular products. car model.

“It’s important to be willing to be a guinea pig. One of its benefits is that you get first choice. If I experience the Volt before others, I will understand the extended range better than those who have not experienced the Volt. What it is like. We can also better understand why those unsuccessful efforts are not successful. Only by experiencing and trying can we gain such a knowledge." Li Xiang said.

The secret to winning this tough battle: keep investing and go all out

Intelligent driving, as another important subset of ideal AI technology , currently in the first echelon.

On the first generation of Ideal ONE in 2019, Ideal built a data closed-loop system - Poseidon, a set of tool chains for collecting, mining, labeling, and training data. As of December 25, the total mileage of autonomous driving by ideal users has reached 2.9 billion kilometers.

In the second half of this year, Ideal will become the second company after Huawei to open a globalThe automaker with smart driving in Guocheng District overtook NIO and Xpeng, which were originally leading. It may seem like a late start, but with a clear rhythm, Ideal has delivered a report card that is not inferior.

On October 23, 2024, Li Auto announced the full launch of end-to-end + VLM. Immediately afterwards, on November 28, Li Auto’s parking space-to-parking space intelligent driving was launched in full. In addition, the full-scenario upgrade of high-speed cities with end-to-end + VLM, as well as the innovative AI inference visualization interaction will also be fully pushed to AD Max users through OTA.

But this is not enough. The ultimate goal of ideal intelligent driving is L4. "Currently end-to-end can only solve L3, but not L4." Li said.

He believes that for intelligent driving to reach the L4 level, three conditions need to be met: first, there must be enough cars on the road - more than 5 million vehicles; second, the VLA base must be mastered The ability of the model; third, you must have enough money, recruit top talents, and have enough computing power.

In December 2024, Li Auto will deliver more than 58,000 new cars in a single month, and will deliver more than 500,000 vehicles in the whole year of 2024. The cumulative delivery of new cars in history is more than 1.13 million units. According to the financial report data for the third quarter of 2024 released by Li Auto, quarterly revenue reached 42.9 billion yuan, a year-on-year increase of 23.6%, net profit was 2.8 billion yuan, a month-on-month increase of 156.2%, and cash on hand exceeded 100 billion. Only by crossing the stage from 0 to 1 and moving towards a scale of 100 billion can we support the next AI ideal.

Since September last year, Ideal has organized a weekly meeting on artificial intelligence. The weekly meeting brings together all the people related to AI in the entire company, including those related to smart spaces, infrastructure, and training platforms, to share two things: one is the latest papers, and the other is the best practices of different teams on AI. In addition, Li Xiang will also discuss AI topics with Lu Qi, Wang Xing and others, using the "external brain" to understand the core and substantive technologies of AI.

But Li Xiang also understands that AI will be a big battle. "I will go all out to ensure that we have the ability to become the world's leading artificial intelligence enterprise and do a good job in every link. But if In my lifetime, I will not be able to achieve it, or we will not be able to achieve it because we have made huge mistakes. I am also very happy to see that some of China's top companies can do very well in artificial intelligence and realize the fifth stage of artificial intelligence."