Chen Yong: In the era of large models, how can intelligent driving follow the trend?

In November 2022, the emergence of ChatGPT changed human life forever. With the rise of large models, all walks of life are thinking about how to use large models to make products, how to use large models to improve energy efficiency and product experience.

How does intelligent driving take advantage of this trend? How to ride the waves?

According to Chen Yong, director of the Technology Planning Center of Geely Automobile Research Institute, the threshold in the field of large models will gradually become higher. Whether it is Baimo or Qianmo that we see today, it is still impossible to tell what will happen in three years. However, the ones that can really persist are the companies that find high-value user scenarios, because user scenarios determine the value of technology. .

In the field of intelligent driving, large models can solve problems and create value. For example, using large models to synthesize data, mining data value, etc., has great imagination.

The following is the transcript of Chen Yong’s speech.

I am very happy to share with you Geely’s application exploration in the large model era. Just now the two CEOs introduced how smart driving should be done in the fields of safety and data/how to make the smart driving experience better. Next, share Geely’s thoughts on intelligent driving.

For smart driving, whether it is users or products, the first thing that the market is more concerned about is safety, because in real smart driving, including assisted driving, safety is everyone and every user, including every person. What OEMs/partners care about most.

How can safety be ensured? The core of intelligent driving needs to solve many long-tail problems. The faster intelligent driving develops, the better it develops. The longer the last little bit of long-tail effect may take more time/cost/effort, because safety is 0 and 1. things. How to achieve this kind of corner case long-tail effect is very difficult and requires a lot of energy to think about and do it.

At present, most of the intelligent driving perception is still in the labeling stage, which means that it still lacks a cognitive ability. It is more about how many objects and traffic participants we have seen and how many targets and traffic participants we have labeled, but whether it really has cognition. ability? As mentioned earlier, if a stone suddenly slides down on the highway, does it have such cognition? Can such a thing be done through generalization? If it has not been annotated, can such a thing be done?

In addition, safety and security are two different things. How can intelligent driving be safe? When there is safety, there is a sense of security. This is not an equal sign. How to create a sense of security for users, security is a premise, and security is an experience.

Secondly, how to make the intelligent driving experience good? After all, it is an intelligent product and experience, but the intelligent experience is not whether it is there or not. Now many people are thinking about various functions, whether it is high-speed NOA, urban NOA, APA, RPA, etc., whether the functions are available or not does not mean whether the experience is good or not.

If the experience of a function is not good, it is better not to have it at all. During the autonomous driving experience, whether it is high-speed NOA or RPA, once something unexpected happens to you, I believe you will not use it for a long time, maybe even because of this You won't be able to use it. On complex roads, how to reduce the takeover rate? Secondly, most of the intelligent driving experience is incoherent.

The other thing is about cost, how to reduce the cost of intelligent driving. Now, in order to meet the needs of many functional scenarios, we have piled up a lot of sensors, hardware technologies, redundant designs, etc., making the cost of intelligent driving relatively high. How to bring it back to its business essence? How to make the cost lower, or how to make the experience better, are issues that intelligent driving should be concerned about.

In the past few years, intelligent driving has gradually shifted from a crude hardware-driven experience to a data algorithm-driven experience. There used to be 1V1R, 1V3R, 7V + millimeter wave radar, but now there are 10V, 11V + millimeter wave radar + lidar. Adding one is not enough, adding two, two is not enough, adding three. This is the case for most of them at present, this kind of roughness Types of hardware are stacked to meet various functional experience or security requirements.

If this matter returns to the nature of business itself or user needs, I believe that what users want is not sensors, but an intelligent experience. How to drive experience improvement through data and algorithms, rather than relying on the accumulation of hardware configurations.

I find it easy to do addition, but difficult to do subtraction. How to drive value creation through user experience and return intelligent design to rationality? This is something everyone who does intelligent driving needs to think about. Technology should be used to drive innovation here to improve the experience and cost-effectiveness, including various integrations, whether it is integrated transportation and parking, integrated cabin and driving, etc., "de-hardware". In addition, through various technologies, whether through large models or data closed loops, it is possible to "light map and remove hardware", which is worth thinking about.

What can large models do?

In the era of large models, traditional AI algorithms have been developed for decades, and large models should be available in the past few years. Since the release of ChatGPT in November 2022, everyone has paid more and more attention to it.

All walks of life are thinking about how to use large models to make products, how to use large models to improve energy efficiency and product experience. In the era of large models, how can we follow the trend in the field of intelligent driving? How to ride the waves? The current big model should be the Battle of Hundred Models and the Dance of Thousand Models.

Looking back at the new energy market five years ago, there were many new forces five years ago. As crazy as the current market is, it may be similar. There should have been more than 400 new energy companies at that time. How many are there now? Maybe that hundred is gone and only a fraction is left. How long can that fraction last? The same is true for large models. Compared with new energy, the threshold for large models may be higher than that of new energy.

Because large models have several core elements: (1) They require a lot of GPU computing power. As the number of parameters increases, the GPU computing power will increase. (2) A large amount of data is required. As the amount of parameters increases and the application scenarios expand, a large amount of data is required, including high-quality data. (3) The field of large artificial intelligence models requires a large number of talents.

The threshold in this field will gradually become higher. Whether it’s the Baimo or Qianmo you see today, what will it look like in three years? Whoever can really persevere and find the scene will win. User scenarios determine the value of technology.

If a technology does not find a suitable user scenario, I think this technology is not without value. It may have academic value, but it may not necessarily have commercial value. We should find a suitable application scenario to determine the value of this technology.

If the current large model is used for intelligent driving, I think it should be worth the price of intelligent driving. If the large model is used to steam steamed buns, it takes five minutes to make one bun, and a few minutes to make ten steamed buns. If the large model is used to make chickens and rabbits in the same cage, For something like this, the big model is worth the price. Who determines the value of technology? It is determined by the user scenario, not the technology itself.

What are large models used for? Obviously we need large models in the field of intelligent driving, but not all fields require large models. Every product has its market value and positioning. Today's high-speed rail and airplanes do not mean that we will no longer ride bicycles. It does not mean that large models are coming. Large models can be applied in all scenarios. Traditional things are no longer needed, not necessarily.

What can large models do in the field of intelligent driving? Large models are there to solve problems. If there are no problems, there is no need for new technologies, or they are there to create value. What problems can large models solve in the field of intelligent driving? Or what value can be created?

The first is data. The requirements for data volume or data quality are very high. Can large models solve the problem of insufficient data volume? The deep-level large model itself has strong generalization capabilities. Can data collection not rely on actual road collection? Can we use large models to generate data?

The second is data value creation. The cost of data collection and data annotation is very high. In the past, whether it was L2 or L2+, including city NOA, it cost at least several million frames, or even tens of millions, or hundreds of millions. Now it is done with data like BEV+Transformer. things. The cost of collecting and annotating a frame of data ranges from a few yuan to dozens of yuan. Most of the actual road collections should not be shared. Different models have differences. Can such a thing be done with a large model?

Third, we have data, but the value of the data has not been truly mined. We may collect a lot of data, but the value of the data depends on everyone's cognition.

My cognitive ability determines my ability to mine data value. If this picture is given to me, and you only mark three obstacles and three traffic participants, this is your cognition, which determines how much this picture is worth. Can the model help us deeply understand the semantics of the data and each frame of the picture and mine more value? Can it do this? I think it's possible to do something like this.