In recent years, Li Auto has provided the Chinese automotive industry with a product methodology that has been widely imitated. Now, the company is turning its gaze to the future.

On June 17th, Li Auto hosted a 'Livis Day' event focused on AI and embodied intelligence. Gone were the industry staples of refrigerators, TVs, and reclining seats. In their place were discussions of VLA, Agent, World Models, 3D ViT, and the in-house developed Mahe chips.

This content seemed more fitting for an OpenAI, Google, or Nvidia developer conference. Many attendees' initial reaction was confusion over why Li Auto was suddenly discussing such complex, unfamiliar topics.

For an automaker targeting annual sales in the millions, this focus might appear to be a distraction from its core business. However, Li Auto's vision now extends beyond just automobiles. The car is no longer the entirety of the story.

It's important to note that the domestic auto market is currently contracting, with sales for the first five months down nearly 20% year-over-year. While competitors scramble to protect sales and deliveries, Li Auto's aggressive pivot is an attempt to break free from the intense price competition and find a new path.

Looking back, over the past few years, nearly every company in China's auto industry has started to resemble Li Auto more closely.

The company pioneered an incredibly successful product formula. From extended-range technology to family SUVs, from six-seat layouts to in-car entertainment systems, more and more new vehicles began to emulate Li Auto's design. What was once considered an unconventional choice has now become mainstream.

This success is precisely why Li Auto must now seek its next breakthrough. All product innovations face the same fate: once proven successful, they are replicated, and the competitive moat begins to erode.

A decade ago, Elon Musk navigated a similar transition.

As more automakers began learning how to build electric vehicles from Tesla, Musk shifted his focus to autonomous driving. As more companies began chasing self-driving technology, Tesla started researching robots. Today, Tesla's head of Autopilot, Ashok Elluswamy, spoke at CVPR not about cars, but about a unified foundational model for robots.

Ashok stated that autonomous driving, the Optimus robot, and intelligent agents are essentially projections of the same foundational model onto different physical forms. What matters is the unified model capable of understanding the world, reasoning, and taking action.

Comparing this speech with Li Auto's Livis Day reveals a clear parallel: as the entire industry learns from Li Auto, Li Auto is now learning from Tesla.

Foundations of Intelligence

The discussion began with the "brain." During a media Q&A after the event, Zhan Kun, head of the foundational model team, was asked why Li Auto places such emphasis on language model capabilities. He revealed the company's true objective.

"We believe that as we progress towards L3 and L4 autonomy, the problems we solve will increasingly approach the 90%, 95%, 98% mark—those edge cases you've never encountered before. This requires a model capable of thinking like a human," he explained.

Over the past decade, the dominant logic in autonomous driving has been scaling through data—feeding models more data, driving more miles, and covering more corner cases. This is an empirical approach. Li Auto is now exploring a different path.

What happens if a car encounters a completely novel scenario? What if the answer isn't in the training data? To illustrate, Zhan Kun gave an example: "For instance, how should it handle encountering an ostrich versus an elephant? If it's an ostrich, a slight bump might be okay. If it's an elephant, a bump could cause a rollover."

For humans, this is common sense. For a machine, it's not. Common sense is based on understanding, not just memory. Therefore, Li Auto is shifting its focus from pure perception to reasoning.

"The biggest difference between humans and animals is that humans use language as symbols for high-level thinking. These capabilities come from language, not from vision," Zhan Kun stated.

This perspective diverges from traditional autonomous driving companies. The industry has long focused on cameras, LiDAR, and computing power. By discussing language, reasoning, and chain-of-thought, Li Auto is effectively redefining intelligence. In their view, the most critical capability for future vehicles may no longer be just "seeing."

Redefining the Agent

Li Auto's definition of an Agent points in the same direction.

Many view an Agent as simply a smarter in-car assistant. However, based on Li Auto's description, it's clearly more than a chatbot. Zhan Kun highlighted key attributes for an Agent: memory, planning, reasoning, and execution. These capabilities, while seemingly belonging to an Agent, are also fundamental to autonomous driving and, indeed, to all future robots.

Whether it's helping a user plan their daily schedule or charting a driving route, the underlying capability is the same: understanding intent, decomposing tasks, and executing them.

MindGPT, Agent, VLA, and World Models may appear as separate business lines. In reality, they converge on a single goal: a unified intelligent agent capable of understanding the world and taking action.

"We break down physical robots into three key tasks," Zhan Kun said. "First, embodied interaction. Second, mobility. Third, manipulation."

He added, "The car contains language intelligence, which has a high probability of being directly transferable to robots—interaction, thinking, long-term planning." Li Auto is now re-examining the automobile within the framework of embodied intelligence.

The industry traditionally viewed cars, robots, and Agents as distinct categories. In the converging worldviews of Li Auto and Tesla, they are gradually merging.

In this framework, cars, Agents, and robots share a common "brain," responsible for mobility, interaction, and manipulation respectively. Li Auto aims to build a unified system possessing all three capabilities.

Every technology announced at Livis Day contributes a piece to this unified intelligent agent—language intelligence, interactive intelligence, action intelligence, and an understanding of the physical world.

The Neural System

If Zhan Kun discussed the brain, then Xie Yan, Li Auto's CTO, addressed the nervous system.

Many view chip development as a cost issue. Xie Yan repeatedly emphasized a different term: full-stack capability. "When moving from L2 to L3, there are many problems that no supplier can solve today. To tackle unknown problems and achieve higher standards, leading companies will inevitably choose to integrate vertically," he said.

In Xie Yan's view, future competition is no longer about simply procuring components, but about competing on system capabilities. Models need to synergize with chips, chips with systems, and systems with vehicles, ultimately forming a complete, closed loop.

This mirrors what Tesla has been doing for years. Dojo, in-house chips, FSD, and Optimus, while seemingly disparate, follow the same logic: keeping critical capabilities in-house. When the industry ventures into uncharted territory, the supply chain can no longer provide ready-made solutions.

Xie Yan later pinpointed the fundamental reason for Li Auto's transformation: "The intense competition in the auto industry stems from homogeneity. To break free from homogeneity, you must do what others find difficult or cannot do."

In a sense, Li Auto is increasingly resembling Tesla. They are no longer competing for the next generation of cars, but for the next generation of intelligent terminals. Li Auto aims to transform from an automaker into a genuine AI company.

Building a Unified System

When asked about the differentiated moat of Li Auto's integrated approach compared to Tesla's, Xie Yan highlighted two key aspects: rapid iteration and vertical integration.

"AI is developing very quickly. If you want to build strong competitiveness around AI, you need to iterate rapidly," he stated. Integrating teams for chips, foundational models, and intelligent driving allows for tighter collaboration and faster iteration.

"Secondly, vertical integration is necessary. In this era, if you only develop your own models without your own chips, you miss opportunities where chip and model co-design could solve problems better, especially for major innovations."

He drew a parallel to Nvidia, which started with chips but now extends into packaging and server racks to solve more advanced problems. "When technology is in a phase of rapid development, ambitious companies aiming for leadership must take this path," Xie Yan concluded.

Data as a Strategic Asset

On the topic of data, Zhan Kun explained that the industry's understanding is converging. First, data volume must be sufficiently large to capture more corner cases, which requires a large fleet. Second, data quality—specifically the quality of driving behaviors—is paramount, especially as the industry moves towards end-to-end paradigms.

With a large fleet, Li Auto can filter for high-quality user behaviors and data. "As our model capabilities push towards 100, it will inevitably follow a logarithmic curve, with diminishing returns. However, our data quality can improve with fleet scale, helping to counteract this curve," he noted.

Chasing the Benchmark

Regarding catching up to Tesla's FSD V14, Zhan Kun outlined two levels. The first is foundational experience: matching FSD's sense of safety, efficiency, and comfort. The second is capability: matching unique FSD features like yielding to special vehicles or navigating extremely narrow spaces.

"To catch up on foundational experience, we need a robust evaluation system. For capabilities, there are opportunities for architectural upgrades. We are making improvements to elevate these abilities," he said.

The Roadmap for In-House Chips

Xie Yan confirmed that the long-term vision is for the in-car AI computing center to fully utilize self-developed Mahe chips. "The ultimate form in our roadmap is a single AI computing center in the vehicle, where all AI tasks are processed," he explained. This allows for high efficiency and task isolation, ensuring critical functions like autonomous driving are not interfered with.

He detailed the rationale behind adopting a dynamic dataflow architecture for the Mahe M100 chip four years ago, moving away from the traditional von Neumann architecture to pursue greater efficiency tailored for AI computation.

Defining the Future

Xie Yan emphasized that the intense competition, or "involution," in the auto industry is due to product homogeneity. "To escape homogeneity, you must do what others find difficult or cannot do. If a supplier could do it, then everyone would have it, leading back to homogeneity. That's the basic logic," he stated.

The ultimate goal, as articulated by Li Auto's leadership, is not just to build better cars, but to develop a unified intelligent system that encompasses mobility, interaction, and manipulation—a system where the car is one embodiment of a broader AI and robotics platform.