Robin Li talks about autonomous driving: After L2, the first commercial will be L4 instead of L3
This year, milestones in the field of autonomous driving in China have been continuously refreshed, and various related policies and management regulations have sprung up, providing an internationally leading policy environment for commercialization and large-scale expansion. As of July, the cumulative order volume of Baidu’s Radish Kuaipao self-driving travel service has exceeded 1 million, and its operations have covered more than ten cities including Beijing and Shanghai. Chongqing and Wuhan have respectively opened up fully unmanned commercial operations.
Li Yanhong judged that the first to enter commercial use after L2 is likely to be L4, not L3. Because the accident liability of L2 and L4 is clearly defined. The L2 responsibility lies with the driver, and the L4 operator is responsible for the accident. The L3 is different, the driver takes over when needed, which makes it difficult to define accident responsibility. Therefore, he believes that the popularity of L3 will take longer.
In addition, technological progress has enhanced the generalization ability of autonomous driving, and the scale effect has gradually emerged. “When we want to obtain the qualification for autonomous driving operation in a certain area of a city, technically it only takes about 20 days to prepare, because the generality of the technology is already very good. It is realized by over-fitting of the region.” Robin Li said. (Wen Meng)
Attached is the full text of Li Yanhong’s speech:
Distinguished leaders, distinguished guests, hello everyone!
It is a great pleasure to come to Shanghai again to participate in the 2022 WAIC World Artificial Intelligence Conference. WAIC has been held for four consecutive sessions, and its global influence and “gravitational field effect” are increasing day by day. The scale of Shanghai’s artificial intelligence industry has doubled, and the construction of world-class industrial clusters has taken solid steps. The holding of the new conference will boost the development of artificial intelligence in Shanghai to achieve a new leap forward.
In the past year, artificial intelligence has made tremendous progress, both in terms of technology and commercial applications, and some even have directional changes.
The AI painting you saw just now is a representative of the technological progress in the past year. The reason why we say there is a directional change here means that AI has moved from understanding language, understanding text, understanding pictures and videos to generating content. Xijiajia’s AI painting is to automatically generate various styles of picture works through text descriptions. , Baidu’s AI digital person Du Xiaoxiao, this year challenged to write the college entrance examination composition, wrote 40 essays in 40 seconds, and the score can be ranked in the top 25% of the total candidates. This is an example of automatically generating an article story from a title described in text. Some video content in Baidu APP today is the result of AI automatically converting the graphic content of Baijiahao into video. These are AIGC, or Artificial Intelligence Automatically Generated Content.
The technology behind AIGC is the so-called pre-training large model. Many of the people here are experts in artificial intelligence technology. I believe this technology will be covered many times in subsequent speeches. What I want to say is that AIGC will subvert the existing content production model, and can create content with unique value and independent perspective at one-tenth of the cost and a hundred times the production speed.
Of course, what is more exciting is the progress at the commercial application level. Artificial intelligence has been on fire for so many years, and business should always be one of the soft underbelly, and the lack of good business prospects will lead to stagnant growth of startups, huge losses, difficulties in financing and listing, and large companies will become more and more ungrounded, or Gradually become a pure research department, or gradually become a vassal of other businesses.
When it comes to commercial applications, the most obvious progress is in the field of autonomous driving. In June of this year, GM-backed Cruise launched a commercial operation of fully unmanned autonomous driving in San Francisco, USA. Although there were various bumps in the middle, they persevered and continued to expand the scope of operations. In China, Baidu’s Carrot Run has accumulated more than 1 million orders in July, and its operation covers more than 10 cities including Beijing and Shanghai. Earlier this month, Chongqing and Wuhan opened the fully unmanned commercial operation of Carrot Run. , providing an internationally leading policy environment for the commercialization and large-scale expansion of unmanned driving in my country.
It seems to me that there is also a directional change involved here. In the past, everyone believed that autonomous driving is still far away from us. Even Sfasky, the winner of the Turing Award, believes that it may take decades to achieve fully autonomous driving. Therefore, people place their hopes more on the progressive route such as L2+, and think that the technical route of autonomous driving is to realize L2 first, then L3, and finally L4 and L5. The policy support of the relevant national departments is also L2 first, then L3, and then L4. In fact, it is likely that L4, not L3, will be the first to enter commercial use after L2. Because the definition of accident responsibility for L2 and L4 is clear, if something goes wrong with L2, the responsibility lies with the driver. This is why OEMs will always say that the driver is still responsible for the accident, no matter how strong they think their autonomous driving ability is. The responsibility definition of L4 is also clear, that is, there is no driver, and the operator is responsible for the accident. The difference between L4 and L5 is that L4 is a limited range of unmanned driving, and L5 is an unlimited range of unmanned driving. L3 is different, the driver takes over when needed, which makes it difficult to define accident responsibility, so I think it will take longer for L3 to become popular.
In addition, from our practical point of view, the speed of technological progress of autonomous driving is beyond expectations. When we want to obtain the qualification for autonomous driving operation in a certain area of a city, technically it only takes about 20 days to prepare. Well, because the generality of the technology is already good, our autonomous driving is not achieved by transition fitting to a specific area.
Today, citizens of more than 10 cities can experience the autonomous driving service of Carrot Run, and autonomous driving is very close to us. Public trust and acceptance of autonomous driving is also growing. According to a survey, 83% of Chinese people accept autonomous driving technology, and Chinese consumers have relatively high demand for connected and intelligent vehicles, as well as their popularity and tolerance.
Of course, automakers are also actively embracing autonomous driving. Many auto OEMs realize that it is neither economical, inefficient nor competitive to start autonomous driving research and development from scratch, and they are more willing to cooperate with us. At present, there are more than 30 mainstream automakers at home and abroad that cooperate with Apollo. Jidu Auto, a subsidiary of Baidu, is also a partner of Apollo. In June this year, Jidu released its first robot concept car, the robo-01, and the mass-produced model will be launched in 2023. It is a smart car that can move freely, communicate naturally, and grow itself, reflecting the “smart awakening” of the car.
In addition to autonomous driving, the past year has seen progress in the commercialization of artificial intelligence in a number of areas. The most obvious is in the intelligent transformation of infrastructure.
The first is intelligent transportation. At present, China’s road traffic network cannot improve traffic efficiency and reduce accident rates through real-time signal light adjustment and vehicle-road coordination. Urban congestion has caused many people to waste a lot of time on the road. In order to alleviate traffic congestion, various localities have to implement policies to limit the purchase and travel of cars, which curbs the consumption demand that should have existed and cannot fundamentally solve the problem. According to our practice in various places, through the intelligent transformation of the transportation network, the traffic efficiency can be improved by 15%-30%, which means that the GDP will increase by about 2.4%-4.8% per year. At present, Baidu’s intelligent transportation solutions have been implemented in more than 50 cities across the country. Just a few days ago, the Ministry of Communications officially listed Baidu as a pilot unit for a transportation powerhouse to carry out pilot projects in high-precision maps, smart cars, smart roads, cloud platforms, and the ecological development of the smart transportation industry.
It is foreseeable that with the improvement of traffic efficiency, the policy of restricting the purchase and travel of automobiles will enter the history, injecting new vitality into the economic growth of cities after the epidemic.
The second is the intelligentization of energy and water infrastructure. China has established a strong physical infrastructure network in the fields of energy, water conservancy, water affairs, and heating. However, in the past construction, the emphasis was on hardware and less on software, and the level of intelligence was not high. This year, large areas of the country have been exposed to high temperatures, and the electricity load has hit record highs. At present, many provincial power grids in China have used Baidu Smart Cloud’s AI inspection, which can conduct 7×24 hours of uninterrupted inspection, and the inspection efficiency has been improved by 6-10 times, effectively ensuring the security of power supply. We believe that the next step should be to strengthen the top-level design of resource allocation of water conservancy and power systems, speed up the intelligent transformation of these infrastructures, and use AI to achieve efficient and real-time resource scheduling.
In addition, in the field of industrial Internet, relying on the unique advantages of cloud and intelligence integration, Baidu Intelligent Cloud has created an AI + industrial Internet platform “Kaiwu”, which was selected as the national “Double Cross-Platform”. Kaiwu is helping Chinese enterprises reduce costs and increase efficiency in major scenarios such as quality management, safety production, energy consumption optimization, and logistics scheduling, improve innovation capabilities, and help China transform from a “manufacturing power” to a “manufacturing power.” For example, in the quality management link, a car factory completed the quality inspection of 22 points of the lights in just one second; in the energy consumption optimization link, we used AI to help a thermal power plant optimize the energy consumption of the air-cooling island equipment, achieving 1 kWh of electricity. Reduce 1.55 grams of standard coal. If it is converted into 1,000 air-cooled units across the country, the carbon emission reduction potential in one year can reach 6 million tons, helping to achieve the national “double carbon” goal.
The commercial application of AI in these fields requires end-to-end technical tuning for each industry. Baidu has been in the field of artificial intelligence for 10 years. In the past 10 years, our cumulative R&D investment has exceeded 100 billion, and the annual R&D ratio has exceeded 15%. Last year, it reached 23%, which is very rare among the world’s large-scale technology Internet companies. This kind of pressure and marathon investment has enabled us to have leading self-developed technologies at all levels of artificial intelligence, from the lowest-level high-end chip Kunlun, to the deep learning framework of the flying paddle, to the pre-trained large model, (we recently launched In the end, the efficiency of the application field can be greatly improved.
Of course, we also realize that the digital transformation of many areas of the real economy has not yet been completed, and the digitalization itself has not brought about a significant improvement in efficiency. It will take time for the penetration of intelligence, and the huge effect of intelligence on the real economy will still be There is no broad consensus. Therefore, the commercialization of artificial intelligence still needs to grope in the dark for some time. But a new thing, from “no one is optimistic” to “no one can match”, the decisive victory often lies in the word “persistence”. Technological innovation, especially.
Scientific and technological innovation is inseparable from the support of institutional innovation. It is necessary to provide the best development environment for innovation with greater reform and innovation courage. For example, at present, the popularization of unmanned vehicles still faces the policy obstacles of “four problems and one difficulty”, that is, unmanned vehicles cannot enter the market, cannot be licensed, cannot remove safety guards, cannot operate and charge fees, and it is difficult to identify accident responsibility. my country’s autonomous driving technology is at the forefront of the world, but the opportunity is fleeting. It is necessary to promote institutional innovation and further break through policy bottlenecks. Only in this way can artificial intelligence and the real economy go both ways and promote great progress in society.
Finally, I wish this Shanghai Artificial Intelligence Conference a complete success! thank you all!