Home > News content

The positive and negative of the Baidu AI Developer Conference

via:博客园     time:2018/7/6 16:36:01     readed:294


Baidu's 2018 AI Developer Conference is still beyond everyone's previous "expectation". Because in the speech of Li Yanhong who took the lead this morning, there was not too much “hanging appetite”, but a few heavy contents were released at once:

  • Xiti“The world's first L4-class production autopilot bus”;

  • Show how Baidu AI Assistant reminds participants by voice;

  • Finally, it was released "China's first cloud full-featured chip & mdash; — & lsquo; Kunlun & rsquo; & rdquo;.

In fact, these are not all. In the follow-up session of the conference, Baidu’s digital executives announced or displayed another bunch of new things:

  • Baidu Brain 3.0, which integrates more AI technology capabilities: covers more than 110 specific capabilities;

  • Baidu map upgraded by the voice assistant: the voice control capability is further improved, and the destination and the two routes can be set in one sentence and the driving route can be planned;

  • A new deep learning framework PaddlePaddle 3.0: More flexible and easier to use;

  • Baidu's small program: the core is natural, intelligent, open, and will also introduce all traffic development of Baidu mobile terminal;

  • The new DuerOS 3.0: not only provides voice recognition, but also helps to connect with product design power, supply chain power and paving the way for business models;

  • The new autopilot platform Apollo 3.0, the first is to access the various resources in the industry, set up the platform together, and then set up a self-driving safety report format, except for the “automatic driving mentioned by Li Yanhong mentioned by Li Yanhong”. In addition to driving the bus, Baidu also showed several examples of autonomous driving applications: including shipping, sweeping, and rural insurance.

With so many new points, the feedback obtained is naturally different, positive and negative. Today, we will simply analyze and judge the most concerned ones.

The world's first L4-class production autopilot bus: negative

According to the information at the press conference, this bus is a product of cooperation between Baidu and Fujian Jinlong Group. The official product name is “Apolon”, and the product focus is 3 points: the world's first, L4 grade, mass production.

“First” “Let's go to the end, first look at the “L4” level, which means that the car does not require human supervision at all and can travel to its destination with autopilot. However, the “A4 Long” rating of “Apolon” ​​is still discounted, because Baidu officially said: At present, the deployment of “Apolon” ​​is mainly in closed venues, such as scenic spots, parks, airports, and speed. Also limit the operation from 20 km to 40 km per hour.

Whether the unmanned bus in the field can be counted as L4, or even if it is “bus”, it seems to be a question.

Then there is the second key point. The official directly marked the words “mass production” in the release information, but when the communication was directly connected at the press conference, the 100th vehicle of the car was just off the assembly line. And at the production site of Jinlong, we did not see the common assembly line facilities, such as vehicle lifting and assembly line operation mechanism. Tiger Sniff then learned from Baidu "Apolon" after the possible production plan, but as of press time, did not get an answer.


Finally, let's talk about the first one. I also searched the world's autonomous vehicle products on the condition of “small bus + no driver's seat”. It was easy to find the above. The earliest launch time can be traced back to 2015. Before and after. Then why is Baidu only "the world's first”? It is estimated that Baidu is more generous when making adjectives.

It is worth mentioning that the above models are currently only used as L4 self-driving experimental vehicles on the site (in a specific closed field). The only long-term operation scenario is between the two terminals of an airport in the United Arab Emirates. ; shuttle bus & rdquo;.


In contrast, last year, the Shenzhen Public Transport Group and the startup Hailiang Technology tested “Alphaba” on the semi-closed road in Shenzhen, perhaps closer to the real expectations of the public on the self-driving bus.

Turned out, but extremely vague AI chip: negative


In today's entire conference, the most important and most forward-looking may be Baidu's AI chip "Kunlun". For this product, Baidu defines it as “China's first cloud full-featured AI chip”.

On specific models, “Kunlun” is divided into two types: the training type “Kunlun 818-300” and the inference type” Kunlun 818-100”. The chips feature a 14nm Samsung process with a memory bandwidth of 512GB/s and a computing performance of up to 260Tops, but consume only 100+ watts.


After listening to so many professional parameters, is it in the fog? Nothing, we simply compare these numbers with two other iconic AI chips, the Google TPU v3 and the NVIDIA V100 GPU:

  • Computational power: TPU v3 is 90TFLOPS, V100 GPU is 15.7TFLOPS (value in single floating point mode), Kunlun is 260TOPS (note that this unit is different from the previous two);

  • Power consumption: TPU v3 is about 200W, V100 GPU is 300W, Kunlun is less than 200W;

  • Process: TPU v3 uses 16/12nm, V100 GPU is 12nm, Kunlun officially says 14nm.

Looking through Baidu's official information, it does not explain exactly what the 260TOPS of the "Kunlun" chip refers to the performance of the operation, but one thing to confirm, this score is already numerically for Google and NVIDIA. The AI ​​pioneer-type company realized “Curve overtaking”.

It is also necessary to mention that the current value of the "Kunlun" chip is still the ideal number for the initial theoretical design. After the chip is officially mass-produced next year, the corresponding parameters may change accordingly.

Leave the customer to help make money "New Apple" & rdquo;: Neutral


Regardless of the "small program" in 2018, it is not too late. At least this time, Baidu did give all its efforts within the scope of its capabilities: opening its own 100 billion traffic portals and providing AI capabilities. Backstage framework.

Of course, from another point of view, Baidu’s idea of ​​this small program is completely different from WeChat. It can be described as “all the way” "

  1. WeChat does not open large ports and does not harass users. Baidu embeds all its own ports and compatible third-party ports, and actively promotes it to users after judging by AI technology;

  2. WeChat said that the small program will run out and leave no traces. Baidu not only has to take the initiative to introduce small programs, but also to precipitate users, to promote users, and ultimately to realize in users;

  3. WeChat's development framework focuses on functional and social attributes. What Baidu emphasizes is that the framework is full of AI capabilities.

Obviously, Baidu’s small program attack is not only to serve consumers, but the ultimate goal is to help developers and Baidu make money. Although it is still doubtful whether it can be realized, it is very real.

In order to push smart voice technology, even design and supply chain support are provided: positive


As a major beneficiary of the rise of A, the development of voice interaction is rapid, Baidu's DuerOS has also achieved a lot of results. According to Baidu's official figures, there are currently more than 200 partners in the entire ecosystem, more than 110 master control devices, and more than 16,000 developers.

It seems to encourage more people to participate in this ecology, but also paving the way for their own intelligent voice. Baidu also presented a two-module solution with no screen and screen at this developer conference. The modular solution, the hardware polishing of specific products will become very simple.

Baidu also found the innovative design group Rococo, directly using the latter's shared design platform to provide a complete set of technology-to-design solutions for the added ecological forces, and even provide the necessary assistance in the subsequent supply chain. To put it simply, as long as you have money, the key things can be handed over to Baidu.


In addition to hardware, Baidu also announced that it will give 100% of the revenue of the DuerOS skills (applications) to the developer for the first six months, and also provide another 100 million yuan to support the skills developers. It is not difficult to see that Baidu is very valued for the new way of interacting with voice. After all, the latter may also become a new cross-platform traffic entry.

Apollo 3.0, which goes straight to commercialization of autonomous driving: positive


As the soul of Baidu in the field of autonomous driving, the key to the success of Apollo's plan is how to enable all kinds of autonomous driving forces to be connected to the platform, so that they can truly achieve interoperability in technology, code and specifications, and ultimately commercial Turn.

This is why Baidu will release a new hardware development platform at this developer conference, and at the same time, it can dock more than a dozen sensor units in the market. On the basis of this ability to integrate resources in the industry, Baidu also put forward the concept of “30 days to complete delivery quickly”, and believe that the process of upgrading ordinary cars to smart cars in the future is only 30 days with the help of Apollo platform. . And depending on the existing code and hardware resources of the Apollo platform, different startups can also develop more autonomous driving scenarios.

At the press conference, Baidu showed several cases based on the Apollo platform, called “New Stoneware”, unmanned car, cool wow auto-cleaning car, HNA’s airport freight car, field spray pesticide picking fruit. Apo cattle and so on. The degree of completion looks good and a few commercial values ​​are beginning to show.


Last but not least, Baidu plans to deploy Mobileye's Responsible Sensitive Security (RSS) model in its Apollo open source project and the “Apollo Pilot” commercial project. The goal of this model is to formalize the common sense of safe driving, for example: always keep a safe distance from the vehicle in front, and do not use road rights in a rigid way.

It is a safety insurance for the autopilot project on the Apollo platform.

Written at the end


Compared to last year's AI Developer Conference, this year is still slightly inferior. One of the keys is that “Lack of Lu Qi”, in today's entire conference, Li Yanhong just gave a speech at the opening, and then stepped down when the final automatic driving platform cut the ribbon.

Judging from our judgment above, there are differences in the current progress of Baidu's various business lines and the leverage of external caliber. This also caused some of the new things to look very "love", while others seem to be fairly reliable.

Of course, this state is not the most important. How to find new market value from these lines and reflect it to Baidu's actual revenue figures. This should be the next problem that Li Yanhong needs to solve.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments