On May 18 this year, when Baidu AI & ldquo; key Mr. Lu Qi stepped down as Baidu President and COO, the outside world questioned Baidu's future AI investment and AI as the core strategy. On July 4th, the first day of the Baidu 2018 AI Developers Conference, this doubt was dispelled.
When Lu Qi retired, Baidu Chairman and CEO Li Yanhong said emptyly that he would “make the foundation of mobile and win the AI era”. What is different is that at the developer conference, Li Yanhong’s speech is “dry goods”: Baidu unmanned vehicle "Apolon" honoured a year ago's commitment to officially realize mass production off-line; Baidu's self-developed full-featured cloud AI chip "Kunlun" was unveiled, and Li Yanhong called it the design force so far. The largest chip.
At the same time, Baidu AI executive team released the 3.0 version of Baidu brain, intelligent voice operating system DuerOS and Apollo platform. Baidu also announced the official launch of the smart small program, the small program will fully access Baidu brain 3.0, developers can call only a few lines of code.
At the same time as "show muscles", Baidu once again expressed an open attitude. “Artificial intelligence is a technological change comparable to the industrial revolution. How to avoid creating a new smart divide? Our answer is open, “Li Yanhong said at the scene, “Open up Baidu’s accumulated AI capabilities over the years, and through the continuous iteration of data, computing power and algorithms, every developer can access the world’s most advanced AI technology. This makes it easy for every company and every company to use these state-of-the-art AI capabilities. ”
Unmanned commercialization speed
In the matter of mass production of unmanned vehicles, Li Yanhong did not have full control.
“Last November, at our Baidu World Congress, I blew a cow at the time, saying that in July 2018, we will have mass production of unmanned vehicles. “There is always a lot of risk in innovation, there are a lot of uncertainties, and the car is not the same as the PPT, and the car often has delayed delivery.” ”
So, when the timeline promised by Li Yanhong, did the Baidu unmanned vehicle mass production? “Our bulls have been realized, and the world's first L4-class self-driving car has been mass-produced. ” Li Yanhong said.
When connecting the Xiamen Jinlong Bus production workshop on site, an unmanned minibus “Apolon” slid slowly toward the camera. According to Xie Siyu, Chairman of Xiamen Jinlong Bus, this car is “Apolon”, the 100th production car. The car has no steering wheel, throttle and brake pedal. It is equipped with Apollo L4 automatic driving solution system. Applicable to scenes, factories and other scenarios.
At the same time of the connection, “Apolon” was also sent to Beijing, Guangzhou, Shenzhen, Pingtan, Xiong’an and Tokyo, Japan. In Tokyo, “Apolon” will be used for connection in the nuclear power plant park and for elderly people in the elderly community.
“Apolon does not have a steering wheel. It belongs to a specific area. It has a low speed and a fixed road condition. It is safer. Regarding the safety performance of the production car, a Baidu company told the 21st Century Business Herald that “the maximum design speed of the vehicle can reach 70 kilometers per hour, but the average speed is 20-30 kilometers per hour. ”
From the point of view of mass production, Baidu unmanned vehicles are indeed ahead. In fact, although technology giants, traditional car companies, travel companies or start-ups have entered since the birth of driverless technology, so far, no company has truly realized the commercialization of L4-class unmanned vehicles. According to the 21st Century Business Herald reporters, Google’s unmanned vehicle companies Waymo and Toyota Motor set the L4 production car at 2020. GM, Ford and Volvo said that they will mass production in 2021.
However, it should be noted that Baidu's current production of L4-class unmanned vehicles is a commercial vehicle, not a passenger car that is generally considered. The so-called commercial vehicle refers to the vehicle used to transport people and goods in design and technical characteristics, including all trucks and more than 9 passenger cars. Passenger cars are mainly used to carry people and luggage, and occasionally carry items, including the driver, up to 9 passenger cars.
“Operation in specific scenarios such as scenic spots, using fixed routes, this piece of unmanned technology is relatively mature, and the surrounding personnel and environment are not so complicated. "The company's strategic plan is to push AI to commercial deployment, the fastest way is to mass production of commercial vehicles," he said. The demonstration effect can also create certain benefits for Baidu. ”
In terms of mass production of passenger cars, Xiangyang is expected to wait until around 2020, “the current unmanned technology is still difficult to achieve full application on passenger cars. ”
Release full-featured AI chip
In addition to announcing the progress made in the commercialization of unmanned vehicles, Li Yanhong also announced Baidu's self-developed full-featured cloud AI chip "Kunlun", which includes the training chip Kunlun 818-300, the reasoning chip Kunlun 818-100 .
According to Baidu, “Kunlun” is by far the most powerful AI chip in the industry (providing 260Tops performance under 100+ watts of power consumption), using 14nm Samsung technology, with 512GB/s memory bandwidth, and tens of thousands of small cores. The composition can efficiently meet the needs of training and inference at the same time. “Kunlun” can be optimized for voice, NLP, image, etc., with a 10x reduction in cost under the same performance, supporting multiple deep learning frameworks such as paddle; high programming flexibility; flexible support for training and prediction.
However, there are also many people in the industry who bluntly say that there is a gap between the current information of the chip and the Cambrian chip. According to the data, as early as November last year, Cambrian released a new generation of AI chip 1M with a 7nm process and an energy consumption ratio of 5Top/W (that is, 500Tops performance at 100+ watts).
However, Xiangyang said that chip performance is on the one hand, and more importantly, the ecological creation of chips. According to Baidu, “Kunlun” has the “full functionality of the cloud” value, which is reflected in the cloud full coverage of the data center, public cloud and unmanned vehicle test and development scenarios, including the common open source In addition to the deep learning algorithm, the universal adaptation of specific scenes such as large-scale speech recognition, search sorting, natural language processing, automatic driving, and large-scale recommendation.
“The chip will only be released in 2019, which is a step slower than most domestic AI chip manufacturers. However, Baidu has the advantage of an algorithm framework. In the future, based on the PaddlePaddle platform, the ecosystem of chip + algorithm is the way out. & rdquo; Xiangyang said frankly, "This test is the promotion and operation capabilities of Baidu and the continuous iteration speed of the chip. ”
In addition to the AI chip, Baidu also upgraded Baidu's brain, DuerOS platform and Apollo platform at the developer conference. Among them, the core of Baidu Brain 3.0 is “Multimodal Deep Semantic Understanding”, which has opened more than 110 AI capabilities and incorporated chips into the technical system for the first time. The DuerOS 3.0 release introduces innovative technologies such as Kids and Geek for the first time and introduces more than 20 cross-context, cross-device solutions. In addition, this version will start the business share model, and in the first 6 months, DuerOS will give 100% of the skill income back to the developer.
The Apollo 3.0 version upgrades the original hardware reference design to a hardware development platform and the vehicle reference platform to an open vehicle certification platform. Based on this, hardware vendors can accelerate their own hardware development, and developers can also acquire hardware to build their own unmanned vehicles at a lower cost. In the field of car networking, Apollo launched the full-scale AI car networking solution “small car OS”, which allows developers to accelerate their development time from half a year to 30 days.
In addition, Baidu also released "smart applet". According to reports, Baidu smart applet will be fully open source in December this year, the developer's smart applet will not only run on Baidu mobile products, but also run on external App and DuerOS smart home, Apollo car life platform, once developed Multi-end operation is possible.
According to the data, from November to the present, Baidu's voice ability daily call times increased by 94%, visual ability daily call volume increased by 416%, and natural language processing day calls increased by 180%. The number of face recognition technology calls in vision has increased by nearly 8 times. “Open source and openness, AI is infiltrating into the capillaries of the economy and society,” said Li Yanhong. “All kinds of applications developed by developers can free humans from duplication, inefficiency and heavy mental judgment. . ”