首页

Recently, the hottest topic in the science and technology circle is "ChatGPT". However, ChatGPT is only an external manifestation, and what deserves more attention is the development of AI technology behind it and its future application.

Some people even describe the changes brought by ChatGPT optimistically: Before ChatGPT, AI was only a module of existing scene products at most. Then, after ChatGPT, AI will redefine the product framework of existing scenarios.

Whether it is as optimists say remains to be seen, but whether autonomous driving, as one of the important scenes of AI landing, will have further development in this wave has still aroused many people’s discussion.

Some people think that autonomous driving needs more graphics, images and data processing ability, requires higher image algorithm, and has little correlation with natural language processing ability. It is not possible to realize autonomous driving with ChatGPT’s ability at present.

Of course, some people think that the appearance of ChatGPT shows us a possibility, that is, trained AI will make high-level autonomous driving expected to appear in a few years.

Why does the progress of AI technology make people pay attention to whether autonomous driving is affected?

Observing the development history of autonomous driving, it is not difficult to find that every major breakthrough of autonomous driving is synchronized with the development of AI technology.

We know that,AI is actually imitating the brain neural network and learning some very humanized skills by analyzing a large amount of data.In 1980s, the first practical application of neural network happened in the field of automatic driving.

In 1987, researchers at Carnegie Mellon Artificial Intelligence Laboratory tried to make a truck that could drive automatically. They manually write codes for all driving behaviors and write as detailed instructions as possible for various situations encountered by trucks on the road, so as to make the vehicles run automatically. But unfortunately, this way can only make the car achieve a speed of several inches per second.

Manual code writing failed, and another doctoral student named Dean pomerleau chose another way: neural network.

He named his system ALVINN. After adopting this system, trucks use the images taken by the roof camera to track what drivers are doing, so as to observe how to learn to drive on the road. In 1991, ALVINN drove from Pittsburgh to Erie, Pennsylvania at a speed of nearly 60 miles per hour.

However, a more direct and broader impact occurred in 2012.

Jeff Hinton, a professor at the University of Toronto, and two of his students, Alex Krzyzewski and Ilya Satsky, won the first prize in the ImageNet image recognition competition, and published a paper introducing the algorithm AlexNet. This paper is not only the turning point of artificial intelligence, but also the turning point of global technology industry.

As the key technology of autonomous driving, target detection and image recognition are highly benefited from the breakthrough of computer vision algorithm. Therefore, with the recognition accuracy of Li Feifei team, director of Stanford Artificial Intelligence Laboratory, surpassing humans for the first time on ImageNet open data set in 2015, autonomous driving, as one of the most important landing scenes of AI, has also entered the fast lane of development.

So, will the appearance of ChatGPT become the Milestone of autonomous driving again?

Generally speaking, AI can be divided into three parts: voice, vision and natural language understanding.The last wave of AI was mainly based on the breakthrough of visual image recognition technology, and this time ChatGPT is a natural language processing technology based on GPT-3 model, which can effectively simulate human language understanding ability, thus helping people better understand and analyze natural language text data.

When we want to discuss what impact ChatGPT will have on autopilot, we think that we should first find out whether autopilot here refers to mass-produced low-level autopilot (assisted driving) or high-level L4 autopilot. Secondly, does ChatGPT refer to a language model or a more generalized generation model?

From the perspective of natural language understanding, ChatGPT has a more direct impact on human-computer interaction in the assisted driving part, but it may not have a great impact on L4-level automatic driving.

Cui Dongshu, secretary-general of the Federation, also wrote in his WeChat WeChat official account that the innovation of human-computer interaction and intelligent cockpit system is very strong, especially the human-computer interaction ability of domestic car companies is very strong. Only China enterprises can understand Chinese more deeply. With the empowerment of the bottom layer in the future, the domestic automobile industry will have more good human-computer interaction effects at the application level.

For example, by using ChatGPT, the vehicle can interact with the driver by voice or text, and provide the driver with real-time feedback on vehicle status and driving information.

Before this, although a large number of in-vehicle interactive systems have appeared, the pain point of the industry mainly focuses on the "understanding" part, and most of the in-vehicle voice interactive systems are not intelligent in "understanding", resulting in a single function and command word of the whole system. ChatGPT’s explosion made the market see the hope of solution.

However, Cui Dongshu, secretary-general of the Federation, also said that,Electrification is the core of new energy vehicles, and intelligence is just icing on the cake. In the future, the core competitiveness of car companies will still be to build electric vehicles, and at the same time make full use of intelligence such as ChatGPT to empower the development of the automobile industry.

Of course, whether it is the core or not, it is not enough to have a technological breakthrough if you want to get on ChatGPT. An AI industry person told Titanium Media that "there are still cost issues, including the use cost, cloud service cost and targeted training cost."

However, from a broader generative model, the generative model with big data and large parameters will help to achieve a higher level of autonomous driving.

He Xiang, a data intelligence scientist at Mimo Zhixing, said in an interview with Titanium Media App that the vehicle-side capabilities mainly include two categories: perception and cognition. The perception ability really relies mainly on image technology, while the cognitive ability relies more on similar generation technology of ChatGPT.

That is to say,ChatGPT’s revolutionary significance lies in: letting AI model enter the era of knowledge and reasoning. At present, the biggest shortcoming of autonomous driving lies in the lack of sufficient intelligence in decision-making planning.

ChatGPT uses a training method called "Human Feedback Reinforcement Learning (RLHF)", and He Xiang, a data intelligence scientist at Mimo Zhixing, explained to Titanium Media APP.GPT is a large-scale universal pre-training language model. GPT1, GPT 2 and GPT 3 mainly improve the parameter scale, while ChatGPT mainly introduces human feedback data for reinforcement learning.

The introduction of this method can ensure the minimum output of useless, distorted or biased information according to human feedback in training.

It happens that there is also a kind of automatic driving decision algorithm called imitation learning, which is to let the machine learn how human drivers do it in different scenarios.

Generally speaking, every takeover by a human driver is an artificial feedback to the autonomous driving strategy; This takeover data can be simply used as a negative sample, which is a record that the autopilot decision is corrected. At the same time, it can also be used as a positive sample to improve cognitive decision-making

"The big model of big data and big parameters can learn more potential knowledge, including different environments and different scenarios, which is equivalent to learning a lot of common sense of autonomous driving. This common sense is very important for autonomous driving decisions." He Xiang, a data intelligence scientist at the end of the year, told Titanium Media App.

That is to say,In the process of autonomous driving research and development, the idea of human feedback reinforcement learning can be used to train a model to verify and evaluate the output of the machine model, so that it can make continuous progress and finally reach the driving level of human beings.

Therefore, it can be said that the improvement of basic ability has brought about the expansion of imagination and applicable scenarios. However, at this stage, we still can’t accurately judge how much change the big model represented by ChatGPT will bring to autonomous driving. An industry person told Titanium Media App that the excellent generalization ability trained by the big model may make there no corner case in the world.

Corner case refers to a small probability event that may occur during driving, but the frequency is extremely low. Although it is rarely encountered at ordinary times, it is likely to lead to a fatal traffic accident when encountering a corner case that cannot make a decision for an autonomous driving system.

The emergence of ChatGPT has made the industry realize that it is possible to gain a higher level of autonomous driving technology by constantly accumulating kilometers and running like this.

In fact, before this, both foreign Tesla and domestic Tucki, Baidu and Mimo Zhixing were already exploring the route of "big model".

In 2020, Tesla announced that it would introduce a large model based on deep neural network into its autonomous driving, and now it has realized a large-scale public beta of pure visual FSD Beta; Tucki expressed the viewpoint of using large models to get through the whole scene of XNGP on the 1024 Science and Technology Day in 2022. Baidu Apollo believes that the Wenxin model will be the core driving force of the elevator’s automatic driving ability.

As early as 2021, Mimo Zhixing announced that it would improve its data processing ability with the help of a large model. On February 17th this year, Mimo Zhixing officially upgraded the large model of human driving self-monitoring cognition to "DriveGPT", which will continue to introduce large-scale real takeover data, and continuously improve the evaluation effect through intensive learning of human driving data feedback. At the same time, it also uses DriveGPT as a cloud evaluation model to evaluate the driving effect of small models at the vehicle end.

However,The development of high-level self-driving cars is a complex multidisciplinary field, involving a wide range of technical and regulatory challenges. The progress of artificial intelligence technology can bring some impetus, but this is not a short-term problem.

It is reported that GPT3.0 involves 170 billion parameters, with more than 300 GB of memory, and the training process costs more than 12 million US dollars. The above-mentioned industry insiders said that the autopilot algorithm is to run in the car. Can such a large model be deployed to the car? How much computing power does it need to support? In addition, autonomous driving can not be completed by repetitive and simple road data stacking, so how to ensure a large amount of data is also a key issue.

The article is reproduced from the titanium media. The opinions in the article are only for sharing and communication, and do not represent the position of WeChat official account. If copyright issues are involved, please let us know and we will deal with them in time.

About the laundry list technology company

Laundry Technology Co., Ltd. is the world’s leading spatio-temporal intelligent big data service company and the leading location data and industry application solution provider in China. Relying on self-developed patented technology, based on high-precision maps and multi-source spatio-temporal intelligent big data, we will establish city-level and industry-level holographic spatio-temporal digital twin systems, and actively serve smart transportation (autonomous driving, smart highway, vehicle-road coordination), natural resource asset management (carbon neutral and environmental protection remote sensing data service), LBS smart industry applications (mobile Internet local business location service, smart travel, smart logistics, new infrastructure, smart city, emergency rescue) and other fields. For more information, please search for "laundry list technology" and visit the company website.

A year’s plan lies in spring, and the spirit of the start is seen.

After the Spring Festival in the Year of the Rabbit, Oaks has made great efforts to innovate, improve quality and empower industries, strengthen market expansion, and greatly increase orders to achieve a good start. From the roar of machines, from the busy work of assembly lines, and from the enthusiasm of employees, Oaks has shown great vitality.

On the one hand, Oaks actively supplements talents in all aspects through spring recruitment activities; on the other hand, it injects new kinetic energy into the high-quality development of the whole year with the attitude of improving quality and increasing efficiency in future factories and waiting for no time.

As a leader in the manufacturing industry, the high efficiency, high standard and high precision of Oaks products have always been recognized by the industry. The digital, intelligent and standardized "future factory" is the mystery of Oaks to improve quality and efficiency. With the accelerated integration of the new round of scientific and technological revolution and industrial transformation, the new generation of information technologies such as 5G communication, big data, cloud computing, Internet of Things and artificial intelligence are constantly changing the production mode of Oaks and accelerating its high-quality development.

In the future factory of Oaks, from the core motor to sheet metal, assembly and logistics, Oaks Air Conditioning Factory adopts intelligent manufacturing almost all the way. It only takes 38 seconds to get from a plastic pellet to an air conditioning panel, and the production efficiency of the factory is improved by 30%, and the proportion of intelligent products is 80%. This set of data is enough to illustrate the hard core speed of Oaks’ "intelligent manufacturing".

The intelligent future factory can not only improve the production efficiency, but also make the production process more refined and reasonable. Through the intelligent management of production, the production process and equipment are fully interconnected to realize real-time monitoring of production process and equipment operation, and can be pushed in real time through mobile APP to realize mobile management and mobile office, which can greatly improve efficiency.

In addition to the intelligence of the production process, the future factory of Oaks has also realized the intelligence of planned production scheduling, rate monitoring and comprehensive efficiency of equipment, and through the collection and utilization of big data, the efficiency of quality inspection is also very obvious, for example, through 5G+AI vision technology, it has changed the dilemma of low efficiency and difficult traceability of surface quality detection by manual naked eyes. Once the product surface has scratches, differences in color brightness and other defects, even if it is thinner than hair, the camera will not let go. The detection system will digitally calculate the defective image, compare it with the standard, and feed back the defects at the first time with the help of 5G communication technology, so as to take immediate improvement measures. This is equivalent to installing intelligent quality inspection eyes in the production process, which is a great contributor to product quality improvement.

The improvement of efficiency also means the reduction of cost. Oaks relies on the intelligent production of future factories. Traditionally, the assembly line needs more than 100 people. Now it is all completed by machines and conveyor belts, and only three workers are needed, which undoubtedly saves the cost greatly.

On the whole, with the help of digitalization and intelligent empowerment, Oaks has achieved remarkable results in optimizing production efficiency, improving product quality and reducing production costs, so that it has enough strength to cope with a more complex competitive environment and write a new chapter in development.

Knowing that there is a long way to go for development, john young is just breaking the waves. Oaks keeps pace with the times, making a good start, laying a good foundation and opening a good game for the whole year with an efficient future factory. In the future, Oaks will further accelerate the pace of upgrading intelligent manufacturing, keep up with the development speed of the times, and strive to run out of acceleration on the intelligent track.

Zhitong Finance APP was informed that Haitong Securities released a research report saying that AI big model training has less demand for temporary rapid expansion of computing power, and private clouds can also shoulder computing power requirements. At the same time, for continuous and large-scale AI training, private cloud has higher efficiency and lower cost. From a completely visible point of view, with the help of private cloud, enterprises can completely control and visualize their network security status, and can customize it to meet their specific needs. The bank believes that the high security of private cloud makes it a more suitable solution. Suggested attention: Pingao (688227.SH) and Qingyun Technology (688316.SH).

The main points of Haitong Securities are as follows:

AI big model training has less demand for temporary rapid expansion of computing power, and private cloud can also shoulder the computing power requirements.

In the past, China’s public cloud market has always occupied the mainstream position of cloud computing. According to China ICT Institute, in 2021, the scale of China’s public cloud market reached 218.1 billion yuan, while the private cloud market was only 104.8 billion yuan in the same period. In the past, the development of the public cloud market was mainly due to the rapid growth of the demand of Internet enterprises in China in recent years, and the process of traditional enterprises going to the cloud was accelerated, while the public cloud had the advantage of flexibility and easy expansion. The bank believes that in this context, enterprises’ demand for cloud is increasing, while public cloud is easier to expand and more suitable for high-speed growth enterprises. The bank believes that in the context of the current rapid development of AI, the demand for rapid expansion of cloud servers is not so strong; Moreover, from another point of view, before the large-scale model training, there is often a preliminary understanding of the training computing power required by the model, and the required computing power will be prepared and laid out in advance, so it is unlikely that a large amount of computing power will be temporarily expanded during the model training, which makes the advantage that there is a cloud that can rapidly expand no longer exist in the field of AI large models, and the private cloud can also support the computing power demand of model training.

For continuous and large-scale AI training, private cloud has higher efficiency and lower cost.

When building a private cloud, the limited budget can be used efficiently by carefully planning hardware, capacity, storage and network configuration. On the other hand, although the public cloud has some advantages in ease of use, these services have great "binding" characteristics. For example, if users use Microsoft’s pre-trained DNN for image processing, they can’t easily run the generated applications on their own servers, and users can’t use Google’s TPU and AuoML tools in non-Google public clouds. However, in the current training of large models, it is necessary to constantly use new data and functions to keep its "freshness". Because the private cloud is only used by a single organization, the enterprise can completely control its software and hardware selection. This high degree of control means that the owner of the private cloud can reconfigure or customize the cloud resources for the task, further improving efficiency. As the big model becomes more and more complex, private clouds can provide greater generality and finer specifications (such as plugging in specific applications and ensuring continuous availability and data speed independence). The bank believes that on the one hand, the cost of public cloud layout and private cloud layout is basically the same for the demand of fixed pre-trained AI big model. Because there is no demand for rapid expansion, the scale cost advantage of public cloud will be weakened; On the other hand, the training efficiency of public cloud will be lower than that of private cloud specially built for its own AI training. For continuous and large-scale deep learning, using local private cloud can save a lot of costs and improve training efficiency.

Al training needs massive and highly sensitive data, and the high security of private cloud makes it a more suitable solution.

For the public cloud, when companies store their data and information in the cloud, it is difficult to ensure that these data and information will be adequately protected. The huge scale of the public cloud and the diversity of companies covering users also make it a favorite target for hackers to attack. In addition, there is the problem of hardware sharing in public cloud. Through public cloud, the work between different companies will be carried out on the same server, and this sharing mode is likely to lead to the disclosure of confidential data and information. Unlike public cloud, private cloud is a cloud infrastructure specially built to provide "isolated access" in a "single tenant environment", that is, it can only be accessed by a single entity, which is usually an enterprise that uses and maintains the cloud. The only purpose of establishing a private cloud is to provide services for enterprises that own the cloud. From the control point of view, because the owner completely controls the physical computing, storage and network equipment, the data security is promoted to the highest level, and the internal administrator has greater flexibility in implementing and accessing security tools; From a completely visible point of view, with the help of private cloud, enterprises can completely control and visualize their network security status, and can customize it to meet their specific needs. Because AI training needs a lot of data, for example, the data volume of GPT-3 pre-training reaches 45TB; Moreover, the application of artificial intelligence faces great security risks. On the one hand, data is associated with user privacy information, on the other hand, the destruction of the model will lead to decision-making errors, data poisoning will affect the effectiveness of intelligent services, deep forgery will be used for extortion, and data leakage will lead to user privacy exposure. Events occur frequently.Therefore, there will be higher requirements for security. Therefore, the bank believes that a private cloud with high security features is a more suitable solution.