Dialogue with Pony AI CEO Peng Jun: L4 is a different species from L2, and the scaling of L4 is about to arrive

Author | Zhou Zhiyu

Editor | Zhang Xiaoling

In 2016, when Peng Jun and Lou Tiancheng founded Pony AI, the autonomous driving industry was embroiled in a battle of approaches: Tesla's incremental route (gradually upgrading from L2 to L4) versus Waymo's leapfrogging route (directly tackling L4). Peng Jun and his team chose the latter.

However, for a long time, representatives of the leapfrogging route, such as Robotaxi, faced unclear commercialization prospects, leading many L4 autonomous driving companies to pivot towards L2. Compared to the relatively long return cycle of L4, the incremental route aims to achieve commercial value at each stage, gradually iterating assisted driving to higher levels of driving.

Pony AI also experienced moments of "temptation." Around 2020, the industry faced a capital winter, and even investors believed that Pony AI would shift towards exploring L2++ assisted driving.

Pony AI co-founder and CEO Peng Jun regarded this period as one of the most challenging times since he started his entrepreneurial journey.

The current sprint for scaling up is another point that Peng Jun finds "most difficult." In terms of regulations and market acceptance, Robotaxi is still some distance away from entering the harvest period.

After gaining recognition from the capital market, Pony AI went public on NASDAQ last November, and now it is entering a critical period—having transitioned from 0 to 1, enabling autonomous vehicles to operate on the road, the next step for Pony AI is to sprint from 1 to 10. Scaling up is a threshold it must overcome.

At this Shanghai Auto Show, Pony AI showcased its seventh-generation Robotaxi, co-developed with major manufacturers such as Toyota, BAIC, and GAC, for the first time. The cost of the autonomous driving kit for this solution has decreased by 70% compared to the previous generation. This lays the groundwork for Pony AI's current battle for mass production.

On the eve of the auto show, Wall Street Journal had an exclusive conversation with Peng Jun. When discussing the current trends in technology and the market, he expressed firm confidence in the L4 route.

Peng Jun believes that today's L2 has become a red ocean, and the market will only become more competitive, while L4 is on the brink of scaling up, facing a blue ocean.

In Peng Jun's view, the ultimate goals of L2 and L4 are different, leading to different demands for technology and costs; the two are entirely different species and business models.

This difference also means that for companies following the L2 route, transitioning to L4 is almost equivalent to starting over. Pony AI, however, has firmly pursued the L4 route and has initially explored a commercialization path, allowing it to go further down this road.

On the policy front, cities like Beijing, Shenzhen, Guangzhou, and Shanghai have issued "city permits" for autonomous driving; on the capital front, L4 and Robotaxi have emerged from their lows and gained capital recognition.

After the policy breakthrough, the real test is to establish public trust. Only by maintaining an accident rate consistently lower than that of human drivers can regulators and users lower their guard Next, Peng Jun hopes that Pony AI can build a "supply-demand network" while moving towards a scaled future. When users' wait times for rides are shortened, the experience is good enough, and the costs are low enough, a commercial closed loop will be formed.

At this moment, the autonomous driving industry is standing at the "recovery phase" of the technology maturity curve: capital is returning to rationality, technology is entering the deep waters of application scenarios, and the dawn of commercialization is beginning to appear.

Peng Jun and his team are still climbing on the long slopes and thick snow of Mount Everest. Their next camp may be the breakeven point of a thousand vehicles, or the data flywheel of ten thousand vehicles, or perhaps some unnamed technological breakthrough. The only certainty is that this mountain must be climbed.

On the way to the "Everest" of fully autonomous driving, many people will still make different choices due to differences in technical routes and business models. However, climbers like Peng Jun never doubt the existence of the summit; they only care about where to step next.

The following is the full text of the dialogue between Wall Street Insight and Peng Jun (edited):

Technology is Just a Means

Wall Street Insight: The autonomous driving industry has been like a roller coaster over the past decade, experiencing ups and downs before welcoming a new round of capital frenzy. What stage do you think the current autonomous driving industry is in?

Peng Jun: The entire autonomous driving industry has entered a "great development" stage.

If we separate assisted driving and autonomous driving, assisted driving has already entered a stage of large-scale rollout and application, while autonomous driving has moved past the 0-1 stage and entered the 1-10 stage, achieving scalability.

With rapid technological accumulation and breakthroughs, the autonomous driving industry had already reached a certain scale three years ago, with over a hundred fully autonomous vehicles operating regularly. The exploratory phase of autonomous driving has passed; it has proven that autonomous driving is technically feasible, safety is guaranteed, and costs are controllable and still rapidly declining. Essentially, the autonomous driving industry is at a stage where scalability is about to arrive.

Wall Street Insight: There are now obvious differences in technical routes within the industry, with various paradigms emerging. Pony AI has also undergone a reconstruction of its technical route; what is your view on the current situation?

Peng Jun: Technology is just a "means to an end." The goal is more important than the method.

For assisted driving, its core goals are cost and universality. For autonomous driving, the biggest core goals are sufficient safety and commercialization. The current discussion of autonomous driving commercialization is about using it within a certain ODD (Operational Design Domain) and a certain operational range.

All technical choices actually arise from subtle differences in goals. While methods are important, the core is driven by the goals.

Self-learning methods like end-to-end reinforcement learning are also widely used in our system.

Technology is not the fundamental key to victory; how to use it is crucial, not the method itself.

There is a Gap Between L4 and L2

Wall Street Insights: How to balance the pressure brought by commercialization with the safety redundancy required for hardware and software?

Peng Jun: Different demands and usage scenarios lead to different costs.

In my view, assisted driving and autonomous driving are two very different products. The technology itself serves the product, which also leads to differentiated technological routes.

Currently, the problems with assisted driving mainly stem from excessive promotion, causing users to treat assisted driving as autonomous driving, ultimately leading to accidents.

The current assisted driving systems, due to scalability and cost demands, often use low-cost solutions with insufficient computing power, resulting in a series of issues. This is actually a problem caused by not using the product according to the correct definition.

For Level 4 (L4), a high level of redundancy is required for safety.

Wall Street Insights: So you believe that Level 2 (L2) and Level 4 (L4) are completely different things, with a gap between them?

Peng Jun: I think they are actually two different species, simply because their goals are different.

A simple analogy is that both long-distance running and sprinting require two legs to move, but a long-distance champion will not become a sprint champion, and vice versa. This is actually a difference caused by the differing demands of their goals.

Although both long-distance and sprinting are running, the different demands for explosiveness and endurance lead to different muscle structures required by the athletes.

Wall Street Insights: Many L2 manufacturers and OEMs are moving towards L3. What do you think of this trend?

Peng Jun: In fact, all transitions from L2 to L4 are essentially starting from scratch. Just like Tesla, which previously did L2 and is now saying it wants to do L4, this transition is not a simple extension of the original method.

When transitioning from L2 to L3, there are also many technical and optimization goals that need to be adjusted.

I’m not saying that L2 companies cannot do L4, but the barriers between the two make this transition require starting anew, and vice versa.

Wall Street Insights: Will Xiaoma consider this direction (L2, L3)?

Peng Jun: I think today’s L2 is already a red ocean. Currently, everyone talks about "equal rights in intelligent driving," which is basically a pricing issue, or that today’s L2 has become very difficult to achieve product differentiation.

Many solutions in this market can no longer provide users with perceivable differentiation, ultimately leading to intense price competition. I believe that unless there is a significant change in the business model, the competition in this market will only become more brutal, or very cutthroat.

Wall Street Insights: Once a viable commercialization scenario for L4 appears, what do you think will be the biggest difficulty or challenge for those L2 companies if they want to enter L4?

Peng Jun: Many of them have never done it before, it’s like starting over with a new company. Of course, L2 companies may progress faster than a completely new company in doing L4, but aspects such as data collection, sensor layout, and safety requirements are all different, essentially requiring a complete redo.

The real-world driving data accumulated by L2 companies is actually completely different from that of L4 companies; the two fields have different demands for data density. To make an analogy, data collection for L2 might be addition and subtraction, while transitioning to L4 requires multiplication; the data is completely different

Scaling is the Key to Overcoming Challenges

Wall Street Insights: What progress has been made in cooperation with automakers?

Peng Jun: After achieving 0-1, Pony AI has been focusing on how to quickly reduce costs and achieve mass production.

At the Shanghai Auto Show, we launched three vehicles in collaboration with automakers including GAC, BAIC, and Toyota.

I believe there are several highlights in our cooperation with automakers.

First, the rapid decrease in costs. The maturity of China's smart automotive supply chain, especially in automotive-grade sensors and computing chips, has facilitated the widespread use of advanced driver-assistance systems, leading to a very fast decline in costs.

Previously, people said lidar was very expensive, but now it is not expensive at all. Currently, the hardware cost of our seventh-generation autonomous driving system solution has decreased by 70% compared to the previous generation, which is less than one-third of the previous cost.

Second, higher durability. Thanks to the rapid development of L2, while hardware costs are decreasing, these components are all automotive-grade products, which also enhances their durability.

Third, a higher degree of industrialization. After deep cooperation with automakers, all modifications and productions are done on a single production line, leading to higher consistency and lower costs. This is somewhat similar to the revolution in production processes brought about by the mass production of the Ford Model T, which further reduced costs and facilitated the popularization of automobiles.

What we are doing with automakers now is somewhat akin to the changes brought by the Model T, producing cheaper and more scalable products for the market.

Wall Street Insights: What will the future cooperation relationship look like between technology companies like Pony AI and automakers?

Peng Jun: I believe it will be closer and mutually dependent.

I don't think automakers believe they can achieve L4; they might be able to handle L2, but they are still hesitant about L4. For automakers, the current scale of the L4 market is not large, and what they can see is only in the thousands. Compared to the millions of units produced by complete vehicle manufacturers, this number is too small.

As L4 becomes more mature and the scale of the L4 market expands, I believe many automakers will embrace L4, but this will take time.

Wall Street Insights: There are many players in the L4 market now; how do you view the current competitive landscape?

Peng Jun: The L4 market has a long way to go, but it is very challenging.

Today, I say L4 has already passed the 0-1 stage, but regarding how to move from 1 to 10, there is still much exploration to be done.

The competition in the L4 market is fierce, and the track is so long that it is actually in a blue ocean state compared to L2. Today, the competition is less about price or market share and more about who can climb the hill first.

Wall Street Insights: Is the current competition in L4 focused on technical feasibility, cost control, or trying to establish a commercial model?

Peng Jun: I think it encompasses all aspects, which makes it difficult. L4 is not a single-point optimization; it requires systematic optimization. This "wanting this and that" leads to a longer uphill process

Chinese Companies Have Advantages

Wall Street Insights: What unique advantages do Chinese autonomous driving technology companies have compared to foreign companies like Tesla and Waymo, such as in data and computing power?

Peng Jun: Chinese autonomous driving technology companies may have started a bit later, but they have actually developed very quickly in recent years.

From a technological development perspective, I believe there are no secrets in the methodologies of domestic and foreign companies; the main differences lie in the refinement of details, and in this regard, Chinese companies have many advantages.

These advantages come from training; the complex scenarios and road conditions in China lead to differences in detail refinement. It's like learning to swim; training in rough seas every day compared to training in a pool will result in a stronger ability to cope with the impact of rough seas.

The differences in data and scenarios lead to different capabilities. The top autonomous driving companies globally will also have unique capabilities due to differences in detail refinement.

We also know that China's manufacturing industry is very strong and has a strong cost-reduction capability. Therefore, Chinese companies like Pony AI have certain advantages in commercialization compared to foreign companies and will be competitive globally.

Wall Street Insights: The cost of mass-produced vehicles has decreased significantly. What is our current status in terms of operations?

Peng Jun: In terms of operations, we need to create model projects in some areas to show everyone that L4 autonomous vehicles are safe, do not obstruct traffic, are user-friendly, and can bring convenience to people's lives.

Only by creating good model projects can we promote user adoption and gradually ease regulations.

Wall Street Insights: In the short term, with the scale not being able to increase significantly, how does Pony AI operate on the user side to enhance user awareness?

Peng Jun: On the user side, I think one aspect is building trust, and the other is in our product design.

Previously, our R&D efforts may have focused mainly on how to make the car drive better. Over the past three to four years, we have placed great emphasis on the user experience in human-machine interaction, the experience of getting in and out of the car, picking up and dropping off passengers, and the entire lifecycle and product experience, which involves a lot of new developments. This also requires a process of building reputation and word-of-mouth.

Wall Street Insights: Are we currently facing more regulatory issues or issues related to social awareness?

Peng Jun: I think these two issues are interconnected and complement each other.

Wall Street Insights: Is there any data that can help us understand safety more directly?

Peng Jun: We have not experienced any serious accidents. Even when minor collisions occur, they are due to other vehicles hitting ours, and our accident rate is more than ten times lower than the average traffic accident rate for humans. The insurance premiums for our autonomous vehicles are also cheaper upon renewal, reflecting the safety of our vehicles.

Wall Street Insights: What is our target for the scale of operational vehicles in the next phase?

Peng Jun: Our next goal is to have over a thousand operational vehicles globally, which we expect to achieve by the end of this year or early next year

Challenges Beyond Technology

Wall Street Insights: At this stage, our main focus is on usability. When the L4 market enters a relatively mature and competitive state, what is the most important thing for an autonomous driving company?

Peng Jun: The most important things are twofold.

First, forming a network. This means that companies need to establish a supply-demand relationship network, and the scale must be large enough so that users have a short wait time for rides and a good experience.

Second, costs must be low enough. The costs referred to here are broad, including the overall cost of hardware and software, operational efficiency, etc., which is a comprehensive cost competition.

Wall Street Insights: Do you have an expected timeline for Pony AI to move from the 10 to 100 stage?

Peng Jun: Achieving operational breakeven or positive gross profit, in my view, is the 1-10 stage, which is our current goal and should be achievable in a relatively short time.

In terms of fleet size, to move from 10 to 100, I believe we need to reach at least tens of thousands or even over 100,000 to say that the L4 market has entered a mature stage. From this perspective, I think it will take at least another 15 years.

Wall Street Insights: Does this mean that Pony AI, which has primarily been a team of technical engineers, will now focus more on operations, or even manufacturing?

Peng Jun: Operating a fleet of over a thousand vehicles and then over ten thousand certainly presents challenges. These capabilities can be said to be lacking in others; operating an autonomous driving fleet is quite different from operating a taxi fleet.

Next, we can supplement some of the existing capabilities. For example, in production, we can draw on some of our automotive manufacturing capabilities; in operations, we can learn from the experiences of taxi and ride-hailing services or recruit some talent. However, how to do this in a new business model will definitely require a lot of exploration.

So we are actually taking two steps: on one hand, supplementing some experienced talent in the existing business model, and on the other hand, continuously accumulating our own "Know How" in the process.

Wall Street Insights: Pony AI has gained a lot of operational experience in the Chinese and American markets. Is there further room for expansion in other overseas markets?

Peng Jun: The operations in China and the United States are more about creating model projects. The next step is to hope to operate over a thousand vehicles in individual cities.

Other countries or regions also have demand for the convenience and economy brought by autonomous driving. So there are some attempts.

Pony AI has already begun to roll out autonomous driving technology and landing cooperation in regions such as South Korea, Saudi Arabia, and the UAE, and recently received the first batch of Robotaxi testing permits issued by Luxembourg. Of course, operating in these markets is different from in China; in China, we operate many of our services ourselves, while locally we need to find local partners to explore the market.

These markets will not quickly reach thousands of vehicles, but it is necessary to enter early, find partners, and work with local regulatory authorities to gradually establish the market and habits Wall Street Journal: Many automotive-related companies are now turning to the robotics field. Will Pony AI consider this?

Peng Jun: From a technical perspective, autonomous driving and robotics indeed have many similarities, or we could say that an autonomous vehicle is just a walking robot! In other words, autonomous vehicles represent a more regulated, sufficiently large, and valuable area of robotics.

I believe the larger robotics industry is still in the 0-1 stage, and there is still much to be done.

I wouldn't say that I will never venture into other fields, but currently, all my energy and resources are focused on achieving L4 at least from 1-10, or even 10-100, before I might consider other opportunities. Startups certainly have many temptations, but focus is very important.

Wall Street Journal: After Pony AI went public, how have investor expectations for the company changed? What metrics are they mainly focusing on?

Peng Jun: In the earlier stages before going public, people were more focused on technological development and breakthroughs in laws and regulations, needing to see L4 becoming increasingly feasible.

At this stage, the 0-1 phase has passed, and people are more concerned with production scale, user feedback, and optimization of usage costs, focusing more on metrics corresponding to large-scale commercialization.

Wall Street Journal: Currently, most of Pony AI's revenue comes from Robotruck. Will its proportion gradually decrease in the future?

Peng Jun: Pony AI's original intention is to change transportation. Transportation mainly involves two aspects: carrying people and carrying goods.

We started laying out the cargo aspect early on and have achieved some commercial operations in specific scenarios. The economic model of Robotruck, in terms of revenue, will be much higher than that of passenger vehicles. Compared to C-end businesses like Robotaxi, it achieves a commercial closed loop more quickly.

However, for trucks to be fully unmanned, the safety requirements will be much higher than for passenger vehicles, and the more economically viable scenarios for Robotruck are interprovincial transportation, which leads to greater regulatory complexity.

Therefore, achieving full autonomy for Robotruck will take longer than for Robotaxi. Currently, Robo Truck is still in the 0-0.8 stage and needs further improvements in operational capability, technology, and scenario optimization.

Achieving Milestones is Important

Wall Street Journal: What do you think was the most difficult moment in the past nine years?

Peng Jun: Any major change is difficult. I think one challenge was right at the beginning when we had nothing.

There were also challenges before and after the pandemic, as the industry cooled down and the economy faced uncertainties. At that time, there were temptations from the prospects of the L2 or cargo vehicle market, which brought challenges.

Now is also a significant challenge, specifically how to truly achieve scalability. I think these three are considerable difficulties.

Wall Street Journal: In 2021, with a poor capital environment and no clear direction for autonomous driving, how did you get through that phase? Peng Jun: First is the focus on goals, and secondly, there has been some slimming down, which definitely includes the pace of hiring and some small benefits, making some reductions in these areas. I think these are the main two, relying on increasing revenue and reducing expenses.

Wall Street Insights: Did you have any self-doubt during that time?

Peng Jun: It wasn't self-doubt; I still had strong confidence in the overall direction. However, I was uncertain about how long it would take. I believe spring will definitely come, but I don't know how long winter will last, so I should think about how to endure longer during this time.

So at that time, I was actually thinking not about changing direction, but about how to make the "takeoff track" longer and how to get through it.

Wall Street Insights: What do you think you did right that allowed you to persist?

Peng Jun: Communication is very important. During difficult times, I communicated more with the team and investors, and through further understanding, everyone came together for support or drew positive energy from each other.

Another thing is to exercise more; exercise itself helps relieve stress and increase dopamine.

Wall Street Insights: You have repeatedly expressed optimism about the L4 direction. Why do you firmly believe that we can see L4 scaled deployment within 10 years, rather than 50 years later?

Peng Jun: It's like climbing a ladder; it’s done step by step.

To digress a bit, many people have asked me how to mobilize the team to focus when full autonomy hasn't been achieved. Or over the years, what I think is the most correct thing I've done.

Many people say that a CEO needs to find people, find money, and find direction, and I think all of these are important. But for Pony AI, my biggest contribution is setting a phased goal for something that could take 6, 7 years, or even longer to achieve, according to the general direction, and continuously achieving these phased goals over half a year or a year.

There should be a lot of data and indicators to show our capability improvement, which in turn brings confidence step by step. Otherwise, even if I have confidence, the team won't