Is Level 3 intelligent driving a "false proposition"? The second episode of AI Relativity is here. Two industry leaders reveal on-site: There have been too many marketing terms in the past, and consumers only care about the results.

As urban NOA (Navigate on Autopilot) starts to rapidly penetrate from high - end models priced over 300,000 yuan to the mass market of 100,000 - 200,000 yuan, the intelligent driving industry in China is entering a new competitive stage: data - driven, end - to - end, large models, and vehicle - level intelligent agents are gradually becoming new keywords in the industry.

On one hand, the parallel evolution of pure vision, multi - sensor fusion, and vehicle - road - cloud collaboration continues; on the other hand, discussions around L3, L4, and Robotaxi (autonomous taxis) are moving from "concept verification" to "commercial implementation". The industry needs to address the challenges of computing power, models, and engineering, and also balance safety, cost, and scale.

Recently, the second episode of "AI Relativity" by "National Business Daily" (hereinafter referred to as NBD) had a dialogue with Lv Peng, the vice - president of Horizon and the head of the Strategy Department, Intelligent Driving Product Planning, and Marketing Department, and Yan Jing, the executive secretary of the Intelligent Driving Professional Committee of the Chinese Association for Artificial Intelligence, on the current hot issues in the industry. The former, standing at the front line of chip mass production, focuses on software - hardware collaboration and vehicle - level intelligent agents; the latter, deeply involved in industry research and standard discussions, focuses on the technology rhythm and industry regulations.

From "effects are more important than buzzwords" to "a truly implementable L3 essentially requires L4 capabilities"; from end - to - end, VLA (Vision - Language - Action), to "the vehicle will become the largest personal computing center", the two guests had an in - depth discussion around the technology path, evolution logic, and the development pattern in the next 3 - 5 years.

Guests of this episode:

Lv Peng, Vice - President of Horizon and Head of the Strategy Department, Intelligent Driving Product Planning, and Marketing Department

◼ Yan Jing, Executive Secretary of the Intelligent Driving Professional Committee of the Chinese Association for Artificial Intelligence

Effects are more important than buzzwords, and the core is data - driven

NBD: From the perspective of the technology route of urban NOA, there are currently different paths such as multi - sensor fusion, pure vision, and data - driven, with many disputes. How do you evaluate the advantages, disadvantages, and future trends of these routes?

Yan Jing: The coverage and consumer acceptance of urban NOA are both increasing. Multi - sensor fusion has been developing. In the early stage, there was a solution of lidar plus cameras. Later, different technical solutions will be launched by comprehensively considering cost, consumer acceptance, safety, and scenario adaptation. This is the comprehensive result from R & D to engineering implementation. Currently, high - end models have more multi - sensor fusion. To make the technology widely used in the mass market, a cost - acceptable solution is needed. Different sensor configurations will be matched according to different costs and scenarios, which is the current mainstream choice.

The pure vision route is used in many models. It has high requirements for algorithms and performs well in good weather and working conditions. However, in poor weather or working conditions (such as heavy rain, heavy snow, thick fog, complex road conditions, etc.), more caution is needed. Data - driven and data fusion are essential. The data includes multi - source data collected at the vehicle end, as well as data from vehicle - road collaboration and vehicle - road - cloud integration. The data from roadside infrastructure can be used as a supplement to cover visual blind spots and optimize traffic scheduling.

NBD: What considerations does Horizon have for different routes and market demands?

Lv Peng:  Horizon has always designed chips and systems in a software - hardware combination way. From a technological trend perspective, in the past, we moved from rule - based systems to hybrid systems, and now to data - driven systems. The difference lies in whether we pre - define rules to tell the vehicle how to drive or let sensor data come in and generate real - time vehicle control trajectories.

The more data - driven the system is, the larger the model, the higher the computing efficiency, and the greater the bandwidth are required. We design chips from this dimension and optimize core operators such as Transformer. The most difficult part of making computing chips is whether there is strong software know - how. It takes at least 2 - 3 years from chip definition to vehicle installation, and you must predict the technological paradigm in advance.

As for whether to use lidar in the data - driven paradigm, it is not the core consideration for us. As long as it is data - driven, using more or fewer sensors is just a choice of R & D path. We won't explicitly say that it must be pure vision or must use lidar. Our R & D focuses more on pure vision, hoping not to rely on the "crutch" too early, otherwise, it will be difficult to reach the upper limit of pure vision performance. However, we also actively embrace the supplementation of different sensors. So, we mainly focus on pure vision development while being compatible with more sensors.

More importantly, whether the computing efficiency of the chip is improved. Computing power does not represent everything, and real computing efficiency is very important. Just like buying a house, if the shared area is large, the actual usable area rate is very low. Making the computing in the data - driven paradigm efficient and highly utilized is our design concept.

NBD: Is the core bottleneck for the implementation of end - to - end technology the black - box problem, or the threshold of computing power and data?

Lv Peng: Data is not the biggest challenge. The amount of high - quality data required for end - to - end is not particularly large. The core is to build a foundation model and form human - like and anthropomorphic characteristics by learning human driving behavior. The biggest challenges are divided into two parts: software and computing platform.

In terms of software, the ability of the model architecture is the biggest challenge, which requires top - notch AI talents, training methods, high training costs, and engineering implementation capabilities. In terms of hardware, it is necessary to consider whether the computing efficiency, bandwidth support, and operator support required for end - to - end are well - thought - out at the chip design stage. Computing power itself is not a huge threshold, but the computing architecture and Pipeline are. Once the Foundation Model is built, the model can be distilled.

Therefore, first, you need to have the ability to build the model, which has a high threshold; second, if the computing efficiency on the chip is not supported, the deployment cost will be extremely high, and the development cycle will be prolonged. All these require software - hardware collaboration capabilities.

Yan Jing: End - to - end is the core technological paradigm in the current autonomous driving field, and new architecture ideas and core capabilities, such as VLA and prediction capabilities based on the world model, are constantly emerging. Every major technological iteration requires enterprises to go through the painful period of R & D, testing, and switching, which is a process of iterative trial - and - error and spiral improvement.

The black - box problem does exist. The non - interpretability of end - to - end algorithms will restrict R & D debugging and model improvement, and also affect consumers' expectations of the boundaries and responsibility division. We need to do some work on interpretability. Some car companies have made the interpretability problem visibly displayed, but displaying it may bring the problem of information overload.

NBD: You also mentioned VLA just now. For consumers, what's the difference between a car using the VLA route and one using a general large model as the decision - making brain when buying a car?

Yan Jing: In fact, consumers don't need to care about the technical route. They should focus more on the actual effects. Technological R & D is constantly trying new paradigms, and the effects are the core.

Lv Peng: There have been too many marketing buzzwords in the past, creating many new terms. For consumers, the most important thing is the real feeling, experience, and effects.

A truly implementable L3 essentially requires L4 capabilities

NBD: Urban NOA is penetrating from high - end models to the mass market of 100,000 - 200,000 yuan, while L4 autonomous driving is more focused on commercialization in closed scenarios. How do you view the relationship between these two commercialization paths?

Yan Jing: I think they are more complementary. Urban NOA and true autonomous driving have significant differences in working conditions, user groups, and surrounding environments. Companies can apply their technologies to different scenarios as a supplement and balance to their revenue. There will be trade - offs at different stages, which is an economic complement.

Lv Peng: Urban NOA will gradually penetrate into lower - priced models. It is similar to automatic transmission in the past. At first, it was only available in high - end cars, but people need it to relieve fatigue and bring a comfortable and reassuring experience. Full - scenario urban assisted driving will definitely become a standard feature in the future, even in cars priced below 100,000 yuan. It will achieve large - scale mass production and accumulate a large amount of data, promoting the development towards true full - scale autonomous driving.

There is no competition here. It is more of a path - selection issue. Whether it is mass - producing passenger cars, developing Robotaxi, or operating in unmanned mining areas, the ultimate goal is to achieve true unmanned driving. We have two anchor points: the market anchor point is how to achieve commercial implementation, generate profits, and support further R & D, so we focus on the large - scale mass production of passenger cars; the technology anchor point is how high a peak you want to reach.

We haven't achieved full - scenario unmanned driving yet because the technology is not mature enough. There are two options: one is to continuously improve the technology to promote the unmanned driving of passenger cars; the other is to limit the scenario complexity due to the limited technology level and implement it in some parts first. The current Robotaxi, mining areas, and unmanned logistics follow the latter logic. It is not that the technology has reached a high enough level, but the scenarios are limited. However, everyone's technology anchor point is that higher level. Once reached, the technology capabilities can spill over to cover many vertical fields. So, it is more of a path - selection issue rather than a competitive relationship.

NBD: Some people think that "skipping L3 and directly attacking L4" is more efficient, while some companies insist on the popularization of L3 human - machine co - driving. How do you view these two views?

Yan Jing: This issue has been around for many years. There are different views on whether to gradually develop from L3 to L4 or directly jump to L4. In practice, it is more of a problem between ideal and reality. Theoretically, directly developing L4 can reduce the twists and turns in R & D, testing, and transformation during the gradual development process, aiming directly at the highest goal, and the responsibility division is clearer. However, in reality, we also need to consider the humanistic environment, social acceptance, policy maturity, road support, etc. These factors interact with each other to form a dynamic balance. Overall, it is difficult to achieve L4 on a large scale all at once. Currently, many L4 applications can only be tested in specific scenarios, limited road sections, and policy sandboxes.

In a broader space, the industry is still following the gradual development route. On one hand, the gradual approach is beneficial for the responsibility coordination of human - machine co - driving; on the other hand, from an economic perspective, companies can obtain cash flow and data earlier. Pushing the product to the market earlier and continuously collecting data during the human - vehicle coordination process can achieve higher scale and efficiency. Therefore, in the current larger market scope, the gradual approach is more practical. It should be emphasized that users must pay attention to driving safety at this stage.

Lv Peng: Personally, I don't like the L3 and L4 classification standards. We have now reached the stage of full - scenario urban NOA assisted driving. We believe that L3 and L4 are essentially the same. If you implement L3 in a limited scenario such as on the highway, it should have L4 - level capabilities and safety in this area. If not, there will be various conditional restrictions (such as interaction when the road markings are unclear or the light is poor), and it will be difficult to implement L3. Because consumers pay more money but have a worse experience than full - scenario NOA. So, a truly implementable L3 must reach the L4 level in the limited area.

Our development has always been centered around L4 capabilities. A key indicator is MPCI (the number of safety take - overs per certain kilometers). Currently, full - scenario urban NOA has a safety take - over every few hundred kilometers. To achieve unmanned driving, this indicator needs to be improved by several orders of magnitude. With the breakthrough of end - to - end capabilities, we are confident that it will increase tenfold every year, and it will reach 100,000 kilometers in about three years. The 100,000 - kilometer level is a good guarantee for the large - scale implementation of L4.

Therefore, if the achievement of L4 is delayed, L3 will have a transition period; if L4 is achieved within three years, L3 will be in an awkward position. Our view is that the mainstream full - scenario urban NOA assisted driving will move towards the implementation of true L4 passenger cars. There may be a short - term large - scale L3 transition in the middle, but how much it can be implemented depends on the technological capabilities. If there are too many restrictions, it will become a pilot project, and consumers won't buy it; if the capabilities are sufficient and the experience doesn't decline, it can be a reasonable transition period.

From this year on, vehicle - level intelligent agents will become more abundant

NBD: What is the core breakthrough point for the implementation of true unmanned driving? Is it technology, policy, or operation? Which one is the most lacking currently?

Lv Peng: All three are important. There are two global technological routes: one is represented by Waymo, which conducts small - scale operations in a single city (such as a few hundred vehicles); the other is Tesla, which does not limit the area. Neither has reached the real technological inflection point. Even the currently commercially implemented L4 cannot quickly expand the area or scale up. Operation itself is part of the product and needs continuous optimization. Policies usually lag behind. Because the technology is not mature enough, incidents are inevitable during the process, and policies will be tightened; but the general trend remains unchanged, and they will be relaxed later. These dimensions will fluctuate.

NBD: From the perspective of industry regulations, what standards are still lacking, or what ethical disputes do we face?

Yan Jing: Standards, regulations, and ethical disputes are essentially challenges brought about by immature technology. If the technology is 100% mature, many problems will cease to exist. The current main problems are unclear boundaries and difficult responsibility definition. Policies, insurance, and operation scopes all need to be improved. China's autonomous driving standards have advanced rapidly and are in a leading position. I have also participated in discussions on autonomous driving standards. During the discussions, there were many disputes among automakers, operators, traffic safety agencies, etc. due to different positions. The formulation of standards is a long - term process and will be continuously improved with technological development. Overall, China's standards and policies do not hinder or restrict technological development but rather encourage and support it cautiously.

NBD: What key changes will occur in the pattern and market status of urban NOA, end - to - end, and true unmanned driving in the next 2 - 3 years? How can enterprises seize the opportunities?

Lv Peng: With the breakthrough of the technological paradigm, it is a consensus in the AI field to provide higher intelligence with larger models and higher computing power. I think it is possible to achieve very good unmanned driving capabilities within 3 - 5 years. At the same time, from this year on, vehicle - level intelligent agents will become more abundant. The emergence of OpenClaw (AI intelligent agent) has transformed AI from a chat assistant into something that can actually do work. The vehicle is the largest personal computing center. This year, the vehicle - level intelligent agent will become a more personalized assistant that understands you better, connecting physical AI and interaction, learning skills, continuously updating, having memory, and even having a personality.

Horizon predicted this direction two years ago and made early arrangements for cabin - driving integration. The operating system and computing