Volkswagen invested $180 million in this Chinese AI start-up company

For the past four months, Li Zhifei's Chumenwenwen or ‘Chumen’, has been one of the most-watched Chinese AI start-up companies. There is no doubt that the deep cooperation between Chumen and Volkswagen Group (China) has brought the AI company a lot of attention. On April 6, the two parties announced that they signed a cooperation agreement on March 30, that they will set up a joint venture enterprise, and that Volkswagen would invest a total of $180 million for this cooperation.

In the context of fierce competition between traditional car manufacturers and emerging intelligent car enterprises, the ‘binding cooperation’ between Chumen and Volkswagen is a milestone event. However, in this particular case, the ultimate question will be on the performance of the product itself rather than the category of cooperation or the amount of capital and resources that was invested.

At present, the world's most successful AI consumer products are Amazon’s Echo and Alexa. Chumen has undoubtedly worked on a similar ‘heavy’ model. But it is undeniable that the success of Amazon is achieved by a R & D team with thousands of members, 4 years of research and investments.

Of course, the founder of Chumen, Li Zhifei has his own logic concerning the next step. We recently interviewed this distinguished person who is known for his sharp personality. Our conversation covered various subjects from the cooperation with Volkswagen, the competition between traditional car manufactures and the "internet" car enterprises, to the future of automotive insurance industry, Baidu's Apollo plan, and deep learning in black box and so on.

Zhifei Li (李志飞) is the founder and chief executive officer of Chumen Wenwen, an information enquiry voice application for individuals traveling outside China. He is also the founder of voice recognition, Mobvoi Inc. a mobile voice search company with semantic analytics, and search technologies. Prior to that, he was a machine translation researcher at Google and a main producer at Google Telephone Offline Translation System. Li obtained a doctorate in computer science from Johns Hopkins University.

Here's what he had to say:

What is the logic behind the cooperation between Chumen and Volkswagen?

Li Zhifei: We have publicly stated technologies and telematics companies are open to collaboration with car factories. After all, only a very small number of strong enterprises can develop their own exclusive systems.

The current vehicle system is still in a competitive relation with smart phones. The reason for us to cooperate with Volkswagen in depth is that this cooperation will enable us to continue to update our products and users, to produce vehicle systems that can actually compete with smart phones.

At present, we are concerned about whether we can develop products with core competitiveness through this cooperation. And we are confident about the current situation, because there are few Internet AI companies which can reach a deep cooperation with car enterprises. Now we have the initiative, and we already began to upgrade our products.

In addition, through the openness of technology, it is more difficult for our competitors to catch up with us. After all, we have been in the leading position for several years.

Why did you choose Volkswagen, and not another company?

Firstly, we believe that Volkswagen has a high status in Chinese automobile industry, the market share is large enough (more than 20%). Even though a 5% market share is already quite good, we are still reluctant to give up the other 95%. So we chose Volkswagen.

Secondly, Volkswagen's long-term investment strategies, such as technology, are more in line with our vision. As for some of the other companies, either they have a relatively small strength, or the sizes of their market share is not large enough, thus it is difficult for them to do things for a prospect that can not be reliably predicted. That's why, Volkswagen's positioning is the most adequate for our goals.

What do you think about the competition game between traditional car manufactures and the emerging smart car companies?

Indeed, the game between traditional and emerging automobile companies is very obvious. After all, this is a very large market. From the perspective of macro-trend, automation, intellectualization, electrification and sharing are all parts of the irresistible trend.

Thus, some of the traditional car manufacturers will decline, but not all of the emerging car companies will be successful. Perhaps 20% of them will do well. This is different from the full retreat of traditional mobile phone companies in the smart phone field. After all, the time for mobile phone manufacturers to react is too short, like the war has already ended before it even started.

The automobile industry is actually a long-term industry. From the concept of a car to its verification and mass production, then to the rate of good products and supply chain support after mass production, very complex issues can arise. For example, the cost difference between selling 1000 cars and selling 1 million is twice or three times more expensive, so if the car companies adopt the 0% subsidy to customers method like smart phone or smart TV enterprises do to occupy the market, they would lose tens of thousands per vehicle. No enterprise can afford it in any case.

So, whether it is from the perspective of technology or that of the business model, the automobile industry could not adopt the same kind of strategy as an Internet related company.

My overall point of view is still relatively modest. In the future, the automobile industry will compete on user experience, product and marketing, in this case those who rely solely on the channel or prices will fail. At the same time, it does not mean that the traditional companies have lost their chances. After all, giants like Volkswagen, or BMW are also transforming into technology companies, and their reaction speed is very fast. The resulting imagination is even larger than some of the emerging Internet car companies.

In the past, most of the car manufactures were selling their products to the channels, and had no sense of the consumers. In fact, it is a B2B business model. After turning to B2C, based on powerful technology and a huge user base, then added with the intellectualized stuff, it is very possible that they could subvert the emerging Internet car companies.

What is the impact of intellectualization - the fact automobiles are becoming "smarter", on the auto insurance industry?

I think the most important thing is networking. After networking, the data can be collected and analyzed in order to develop personalized insurance products, rather than previous unified model of $ 3000 each in the United States. So, the changes in auto insurance industry in the future must rely on intellectualization or should be Internet-based. And this is why many manufactures are improving the intellectualization of their products.

At the same time, since the traditional insurance companies do not know the specific driving situations of consumers, the intellectualization or Internet-based is equivalent to enabling the insurance companies to directly face the consumers. Thus, it will be an opportunity to make this business bigger and better, and this is one of the considerations behind the cooperation we have reached with the Volkswagen.

We know that products like smart watches, or other intelligent hardware are in fact not particularly successful. So, what is the specific strategy of Chumen when you decided to cut in from the perspective of voice?

The wearable smart devices are indeed confronting challenges. But the smart watch is just a branch of the entire category, there are many other devices such as smart wireless headphones. We are generally optimistic about the wearable smart devices.

For most or even all of the artificial intelligence algorithms, the first step (and a very important step) is sensing, which means collecting the required data, such as personal physical status data, environmental data, or personal and environmental interactive data. Followed by this step there are big data analysis, control and action. And if the action is to change the state of the environment, then there is the need of sensing again. Thus the closed circle of intelligent requirements has been formed.

Mobile phones are able to collect data, but unfortunately, they are not the kind of devices that you can wear all day long. So it seems that the data collected by wearable smart devices must have a very unique value, but these devices are not yet popularized.

So, we are not really bothered by the fact that everyone is saying that the industry has recently changed from all sorts of perspectives. We want to look at its long-term development.

We hear that in the future vehicles will become the ‘third space’ of life in addition to the family and the workplace. What is the strategy of Chumen in the auto industry?

Currently speaking, if people want to buy a car, they only want to drive it. Just like before, if people wanted to buy a cell phone they only used it to make calls. But now, making phone calls is not the most important function of the smart phone. I think that this rule also applies to cars. The vehicle will become a new living space when the unmanned or autonomous car driving is popular. Thus, interactions, contents, services in this space such as entertainment, information query, etc. will become more and more important.

Therefore, our market entry point is not autonomous cars, this is not our strength. On the contrary, we focus on the intellectualization of equipment inside the vehicle .

The first thing to do is networking, through which the query, voice interaction, navigation, music listening and other functions can be achieved just like that of smart phones. So we are developing a set of intelligent interactive system. I think that this is even more imaginative than the unmanned driving, because it builds a brand new virtual world.

It sounds like the strategy of Amazon. Is there any difference?

We tend to combine the software and the hardware, developing both all by ourselves. Then we will choose an influential partner to cooperate, since products like cars require a very strong sales channel. Amazon hopes that all the cars will carry their Alexa platform. This is a magnificent vision, but also a difficult one, because the car ecosystem is not the same as the internet.

And today we have found a powerful partner, Volkswagen, which accounts for 20% of Chinese market. We will set up a joint venture company which is able to do all car-related products, and Volkswagen will open its offline sales channels and automotive products. It will allow us to carry out deep cooperation as two equal shareholders, both having certain control over our products, and the interests are shared.

This cooperation is far more successful than a platform at this stage, because you can not set up a joint venture company with all the manufactures, and it is impossible to be exclusive at the beginning. So, our strategy is a different style, we are considering the quick landing.

What do you think of Baidu's Apollo plan?

Recently there was a famous article that said Baidu's plan has made people cry and others laugh. Its main idea is that, those who are crying are companies focusing on long-term research and development, those who are laughing are companies focusing on the usage of the products which were developed by the former kind of companies.

In fact, the most fundamental problem is that, the platform and the deep integrated products are two very different ideas. Baidu is hoping to enter the automatic driving business through this platform, but it does not have its own vehicles, or even much experience in the business. Perhaps Baidu can produce a product with 90 points, but this product must be installed into vehicles, and might be related to a matter of life and death! Even a 99 point product can not be put into daily use.

And, with regard to those laughing companies the article mentioned, do they have the ability to make the 90 points product into 99 points? This is impossible, because it is more difficult to score if you have had high points, the hardest are the last few points. So, these laughing companies can not make a real first-class product. It requires a lot of inputs and deep integrated applications.

This plan might affect some start-up companies. They could have made some money through unmanned solutions or even certain pre-research funds. But now these will not happen because of Baidu's plan.

What is the business model of Chumen?

Generally speaking, we have the following business models:

The first part is the hardware itself. Our rearview mirrors or smart watches can generate money

The second one is data. Because we have built an interactive system, the accumulation of a large number of users will naturally produce a lot of data, and, like some internet companies' business models, we can create data value through advertising, insurance or other means. The premise is that your products have a large number of users.

In fact, this model is the same as the entire smart phone profit model. In the early stage you make money on the hardware, then slowly begin to make money on the software, and then make money on advertisers and insurance companies.

What do you think of Tesla? Do you think that it has already succeeded?

I think Tesla is very successful, it is no longer a start-up company. Although its current market share is still relatively small, it actually involves a difference between the software and the hardware industry.

Generally speaking, in the software market, the top two giants can occupy almost 80% of the market, but in the hardware market it is the opposite. There is no ‘winners take it all’ situation. Even in the field of smart phones, the company in the first position only has 15% market share, and this is simply determined by the consumer's personal preferences.

No matter how good Tesla is, it is impossible for them to occupy 50% of the market. We all know that the market share of Apple is only more than 10%. In sharp contrast, the searching engine business of Baidu accounted for more than 80% in Chinese market, and Google accounted for about 70% in the United States.

What do you think of the safety of unmanned driving / autonomous car technology?

In the short term, I am pessimistic about autonomous cars; in the long term, I think it will happen eventually. From the perspective of practical applications, a number of auxiliary unmanned technology might be the first to succeed, such as the short distance alarm or route offset alarm.

Completely unmanned driving is destined to be a long process, after all, it is not simply a technical problem, it also involves all sorts of legal, moral and business problems.

However, the entire AI business is probably like this. Because currently speaking, the most successful areas of AI practical application are not voice interaction or visual recognition, but some auxiliary functions. One of the most typical examples is Google translation, and the same reason applies for new media applications like TouTiao: it does not require 100% accuracy, as long as there are 80% accurately pushed notifications.

What are your prospects for unmanned driving technology?

We want to promote intelligent interaction before the unmanned driving is mature. My thinking has two aspects:

Firstly, in the current stage, how do we produce some auxiliaries which are able to realize some of the automotive functions? We have developed some products like our ‘intelligent magic mirror’.

Secondly, since the unmanned driving has not yet been fully realized, we have to find ways to simulate unmanned driving scenarios. The most typical example is the concept of ‘living space’.

In fact, we can now stimulate unmanned driving by modifying vehicles, for example, to completely close up the cab and communicate with the driver only through the APP. Through this kind of simulation, we can start to build our intelligent interaction in the absence of unmanned circumstances, to upgrade our early products and technical iterations.

What is the future development trend of this technology?

At present, many people are concerned about two problems of deep learning: understanding and controlling. Frankly speaking, this technology is currently a black box. How does it capture the semantics and structure? Could we adjust local parts of the model manually so that it can output results in accordance with the needs? These questions are still unanswered.

The former models trained out by the natural understanding or machine translation can be understood. These models are discrete, string-to-string mappings. So when it outputs a result, the result is explanable. For example, if we want to translate the word ‘China’, the model file can specifically tell you the probability of translating ‘China’ into the corresponding words in the Chinese language.

The second question is controlling. The probability of translating ‘China’ into the corresponding words in the Chinese language can be manually adjusted. This is actually an advantage of the previous non-deep learning technology in the machine translation.

So, from the perspectives of understanding and the field of machine translation which I am familiar with, the deep learning technology should make it possible for us to understand and control the model. And, our knowledge of the physical world or that of semantics should be able to enter the model in advance, without the ‘blind learning’ through data.

Deep learning is like a black box, people want to control the machine but they do not know the principles. Does this also imply certain security risks? In addition, ensuring safety may also be a huge business opportunity. What do you think of it?

I think that all of current machine learning methods are inseparable from certain data and objective functions, that is, engineers or scientists command the machine to optimize a objective function based on data. I think machines are still ‘loyal’ at the level of algorithm, since they do not have consciousness and conceptions of good and evil. They are more loyal than persons, unless there are bad guys controlling them.

So, whether the deep learning is safe depends on who is controlling it, and on whether if there are constraints on the legal level or relevant policies. Just like that not every country is allowed to possess nuclear weapons.

Original article here

#AI #technology #startup #China