[Introduction to NLG]

Natural language processing (NLP) is likely to be one of the most interesting AI directions in the coming year. Beginning in 2015, new NLP applications such as deep learning and intensive learning are seen. For NLP applications in particular, Dialogue system maturity can have a tremendous effect "assistant to" founding gold partner Wang Guanghuan and Hu Yichuan recently accepted the interview of Neo-Chi unit. Here, we will see "assistant to come from scratch", the two machines to learn sea turtles from the business, the acquisition of Baidu into Baidu, re-start the story.
From DVD online leasing, video personalized recommendation to the intelligent personal assistant and then to the smart industry assistant, Wang Guanchun, Hu Yi-chuan, two machine learning PhD entrepreneurial path, a complete presentation of the AI ​​entrepreneurial Vientiane and twists and turns: life and entrepreneurial direction of choice, Trial and error of business model and transformation, negotiations and cooperation with VC, and internal conflicts of opinion and unity ...

Wang Guochun and Hu Yichuan, a smart assistant company founded in 2015, are also completing a strategic upgrade from ToC to To B. Behind them is a deep insight into the current reality of AI applications: The industry-level solution will be breaking point.

"Assistant to also" on July 1, 2015 formally launched. In September of the same year, it received an amount of 25.46 million yuan from the angel wheel financing, led by the speed of China's lead, Sequoia Capital, real funds and surging capital with the vote. In December 2016, it completed a Series A financing of tens of millions of RMB, led by Microsoft, Light Speed ​​and Sequoia. The current round of financing B has been completed, soon the news will be announced. At present, the total number of "assistants" has accumulated to more than 3 million, which is not a small figure for new AI applications such as "smart assistants."

From 2011 to the date when two people tried their best to start their own business, it is the new wave of AI wave that has been going from ready to fully swept in a few years. Here, we will see "assistant to come from scratch", the two machines to learn sea turtles from the business, the acquisition of Baidu into Baidu, re-start the story.

Leave school, sounds beautiful: PhD from machine learning to the company founder

By the end of May 2011, Hu Yi-chuan, who was reading the essay and systems engineering department of the University of Pennsylvania at Philadelphia, received a phone call from his friend Wang Guanchun. The content of the call is short, but Hu Yi-chuan needs to make a big decision: whether to continue completing his Ph.D. or choose to leave school and start a business?

At that time, Wang Guanchun, an hour and a half from Philadelphia, a Ph.D. in machine learning at Princeton University, is about to graduate, has got the offer consulting consulting investment bank. But obviously, venture is the first choice of Wang Guan-chun.

Wang Guanchun (left) and Hu Yichuan during studying in Bo

Hu Yi-chuan recalled: "On the phone, Guan Chun asked me if I would like to return to attend a Chinese-student entrepreneurship forum organized by Xu Xiaoping, who at the time said that while he did not quite understand what personalized recommendation techniques were, He is willing to invest 1 million. "

Wang Guanchun was determined to find partners, he said: " Google by Larry Page and Sergey Brin this gold combination co-founded two are PhD Stanford University, have a technical background, already met at school, together Drop out of school business . At that time, I think, entrepreneurship must find such a partner.I first of all contact with the Princeton business plan competition with the younger brother, but he made it clear that you must first complete doctorate .I am depressed I suddenly thought of a more suitable Of course, he has to make choices more difficult than me, I just give up the offer of work, but for the Ichikawa, entrepreneurship means to leave or even drop out of school. "

When Wang Guanchun and Hu Yi-chuan, the "Chinese AI Partners" who have been working together for six years, sit in front of us and accept the interview with Neo-Chi, they can easily feel the difference in their personality: Wang Guanchun is more passionate and full of expression. Hu Yichuan relatively restrained, more in listening and thinking.
"What we all have in common is that we are optimistic and have the courage to embrace new trends in new things, even if it has certain uncertainties," said Wang Guan-chun. In his opinion, Hu Yichuan's entrepreneurial passion is in fact no lower than himself, but not easily revealed.

Wang Guanghuan's judgment is not wrong. Upon receiving his call, Hu Yi-chuan immediately purchased a ticket to Beijing and took part in the entrepreneurship forum. Later, in order to start a business, he really chose to leave school.
For this choice, Hu Yi-chuan explained: "When I was a graduate of Tsinghua University, I tossed a lot of entrepreneurship together with my classmates, and after the United States I did not completely give up my idea of ​​starting a business. , But I practice different internships each summer vacation and have internships at investment banks, research labs, and start-ups, where I spent some time in startups and experienced the startup environment in the United States, which buried me in my heart Entrepreneurship seeds inspired .I think it is still taking advantage of more to explore more young .Because if you find a relatively stable job, you may go along that road, it is difficult to come back ten years later In addition, we firmly believe that some emerging technologies - from the Internet to the mobile Internet to today's artificial intelligence will really profoundly affect the society in the coming five years, 10 years and 20 years. I think, choosing to start a business does indeed have Such a little idealistic feeling in it. "

Wang Guan-chun and Hu Yi-chuan first met each other because Hu Yichuan's girlfriend-and now Mrs. Hu-read a Ph.D. at Princeton. Hu Yi-chuan went to Princeton almost every week from Philadelphia.

At the time, Wang Guanchun was the chairman of the Chinese student union in Princeton and Hu Yi-chuan was the chairman of the Chinese student union at Pentagon. Both showed great organizational skills.

But when it comes to entrepreneurial experience, Hu Yichuan also toss some entrepreneurial projects, but actually Wang Guanchun and no real experience, the closest venture is only co-founded in the school open curriculum translation of public welfare organizations, in addition to the McKinsey, Goldman Sachs in such a business practice.

"The one that led me to start my own business was a course on the last year of my Ph.D., called 'Hi-Tech Startup.'" Wang Guan-chun said: "The teacher in that course is extremely infectious. Analysis.I found at the time that many of the daily headlines were mostly big business announcements of new product strategies or announce mergers and acquisitions and other information, and consulting firms and investment banks are often the trader behind these projects.I am fascinated that maybe that is Closest to the pulse of the world.But after this entrepreneurial class, I suddenly realized that, in fact, today to see the headlines are two or three years ago, or something earlier in the startup decision . If I If you want to go at the forefront of social development, I should start a business and make innovations instead of thinking of how to become a part of my headline.

In the final lesson, the very contagious teacher - his name is Ed Zschau - said he wanted to leave three words on the tombstone: father, entrepreneur, public servant. He was in his 60s and had done all three things, and I thought it was a wonderful life, so let's start with the entrepreneur. "

China Netflix - from the "rent" began

It was also at this time that the idea of ​​making "China's Netflix" began to emerge in Wang Guanchun's mind. He took this idea to participate in that year's Princeton Business Planning Competition, has been recognized by the judges, took the school first.

"At that time in Princeton, in the eastern United States, the entrepreneurial atmosphere was not as strong as it was in Stanford, Silicon Valley. Many of our brothers and sisters went to Wall Street after graduation and I decided to go to VC because I had the encouragement of classes and awards." .

"I was looking for VCs in New York at the time and also flew to California to talk to several start-up companies that did an online video at the moment, one of them called VUDU, acquired by Wal-Mart today and later met by friends Mr. Gong, the founder of Gong Yu and Douban, also got a lot of encouragement from entrepreneurs in the early days of video in China, and then contacted Mr. Xu Xiaoping through friends of his friends.

Wang Guochun told Xu Xiaoping the idea of ​​"making China's Netflix." This later became Wang, Hu two first venture - "Forrest network." Xu Xiaoping invested 1 million as an angel investor.

Forrest Gump ends cross-country run in Forrest Gump. Hu Yichuan also attracted the extras when the camera to achieve God's synchronization.

DVD rental to personalize the recommended transformation - the first disagreement of the gold partner
"Forrest Network" is first and foremost a DVD and Blu-ray disc rental company. At that time, Netflix was going to transition from Blu-ray and DVD rental businesses to online streaming media, but 7-8 percent of its revenue still came from the leasing business. Wang Guan-chun said that at the time they hoped to make the Blu-ray DVD rental business in China first and then do the transformation.

"This part of the business," said Hu Yi-chuan, "Mr. Xu understands that what he did not quite understand might be another business that Forgan wanted to do-personalized recommendations."

"Our Ph.D. was machine learning, and machine learning was far from being so hot today, but Mr. Xu knew it was a cutting-edge technology, and as a machine-learning method, he did not have a similar product at that time. When the industry's personalized recommendation technology itself is not as good as it is today in China, we are concerned about the field earlier and practice. "Wang Guan-chun said.

"We even thought that personalized recommendation engines might not be commercially viable at the moment, so we think we need to put it first into a Blu-ray DVD rental business," Hu said.

In "Forgan Networks", Hu Yi-Chuan CTO, the main commitment to technology and product development work; Wang Guan-chun, who served as the CEO's role in product strategy, team management and marketing assume more responsibility. Since then, until "assistants come" today, the two have always been such a division of labor.

"Maybe the background is similar, we have not had too intense conflicts and disputes." Wang Guanghuan said Hu Yi-chuan sat nodding. "Of course, our opinions differ in the face of some big, challenging and big decisions, and we're going to talk and work hard to reach a consensus and then make a choice, to date, a similar process has gone through twice and all involve To the company's strategy and business adjustments for the first time, is taking place in the Forrest Gump from DVD rental business to the online personalized recommended video platform transformation process.

"We found it very difficult to get through the rental business because it was also our first venture and did not know when to stop and when to continue." We've been doing this for more than six months and have begun to think of this The direction is more difficult to do.But should not be transformed, how to transition, in fact, not clear.Our disagreements mainly lies in the pace of transformation - is to immediately give up all the rental business, or temporarily retain some?

Wang Guanghui, Hu Yi-chuan into memory, he said: "There was indeed some disagreement.After all, we have just started to make the entire framework, although there is no outbreak, but also accumulated a number of paid users in the end or not to stop it at once , Switch to something else, or keep this part first while exploring other plans?

"We talked about many times at that time, but we are all more rational people and we can only persuade people to impose their opinions on each other by arguing. In particularly important decisions, we first agree and then proceed Our ultimate common understanding on this issue is:
1. We have limited resources and need to concentrate our firepower;
2. The issue of renting DVDs is something we can do before we return to China, but after six months we found that it really Not the future.If the future, we must do online streaming media, give full play to our recommended technical advantages. "Hu Yichuan said.

Now looking back, both think the rental business was "not reliable." "Forrest Gump" limited to the domestic logistics system was not perfect, distribution and recovery costs remain high, it is difficult to form an effective and sustainable development model. So the team eventually closed the offline service, focus on online movie search and personalized recommendations, the name changed from "Forrest Gump" to "tonight to see what."

By the end of 2012, "tonight to see what" was acquired by Baidu. This result undoubtedly confirms the correctness of the transition. Wang Guanghuan and Hu Yi Chuan, also because of this acquisition and become a Baidu, responsible for vertical search personalized recommendation, intelligent interaction and other product research and development work.

"Assistant to also" was born: see the new trend of three major AI technology

In the three years of Baidu, Wang Guanchun and Hu Yichuan helped and witnessed Baidu's mobile video service from zero to over 100 million users. Its personalized recommendation system has also been upgraded from the scale of one million users to the tens of millions of liveers on a daily basis. A year later, Wang Guanghuan turned to natural language processing department, responsible for intelligent interactive application team, made a series of product innovation attempts around the technology of semantic dialogue recommendation.

By 2015, Wang Guanchun and Hu Yichuan have noticed three new trends:

1. With the maturity of smart phones, instant messaging software has become more mainstream , such as the rise of WeChat. At the same time, WeChat made a huge innovation: it developed a public platform. Through natural dialogue, many services, as well as many users and brand communication, into such a product form;

2. The rise of O2O companies doing real-time services, while the traditional offline businesses such as finance, retail, education and so on are also going online . The offline trend of online service allows Wang and Hu to start thinking about whether they can make smart assistant products and directly connect humanized interactive services with services.

3. The depth of learning and reinforcement learning algorithm for NLP and other new applications, especially in mature dialogue system can have a tremendous effect . In vertical scenes such as air tickets and hotels, the user's reversion rate through voice conversations has been higher than traditional click interactions because natural language interaction can shorten the path of user actions.

It is felt in the above new trend, the two decided to leave Baidu, re-start, to be the most reliable smart assistant company. July 1, 2015, "Assistant to also" formally launched.

Assistant also anniversary

Hu Yi-chuan said: "In terms of technology, the most unique thing about us is finding the perfect way to combine AI with HI (Human Intelligence) . Let AI enable service providers to improve HI efficiency on the one hand and user experience on the other hand Collect high-quality user data.All the natural language processing methods are actually based on the data.How to get enough data from the actual application scenarios is the key.We have been stressed internally, to also the product, in fact, is to get through a Positive Cycles - Because we use AI + HI, users get a satisfied or even over-expected service experience so users will be active and will survive, so we have more high-quality data With the data, our AI is stronger and better able to assist HI. This cycle, I think we have already got through . "

Wang Guanchun added: "The involvement of HI can make the positive cycle acceleration, such as real assistant in the service process to do every action, whether clicks, annotations or error, you can optimize real-time model feedback, AI evolution faster. "

Talk about the AI ​​+ HI system architecture, there is such a picture:

Compared with the classical dialogue system, the main features of the dialogue system architecture of Llamasi are that it can seamlessly cooperate with HI and AI to provide a more stable and reliable user experience. Specifically, the cooperation between HI and AI manifests in two aspects: 1) AI's understanding of Context can be visualized in tools used by HI, and HI can correct errors when AI's confidence is not high ; 2) The action generated by the AI ​​to be executed is also displayed to the HI, allowing HI to intervene in a countdown manner. HI and AI through this cooperation, continue to accumulate high-quality dialogue corpus, to help AI continue to evolve. In addition, in conversation control, we use deep neural networks to learn and express the state of conversation. This has two advantages: 1) Characterizing the state of conversation using implicit vectors of neural networks, eliminating the need to manually define the state of dialogue, Strong universality, and can quickly migrate from one field to another; 2) The model can take as inputs the multi-dimensional features of context information, semantic understanding results and user characteristics, so as to make the dialogue system more personal and have different scenarios Under the adaptive ability.

The Transformation from "Almighty Assistant" to "Industry Assistant" - The Second "Disagreement" between Wang Guanchun and Hu Yi-chuan

However, despite its unique technological advantages, the number of users of C-end products continues to grow, but the problem of commercialization has always existed.

Wang Guanchun said: "We found that Almighty Assistant has many challenges, such as how to manage user expectations, how to define the boundaries that AI can solve the problem.We also do a lot of attempts in the business model, for example, But later found that head service providers rebate less and less, long-tailed partners even face operating difficulties. At present, the assistant to the model is through the VIP membership service to do the forward charge, but the user is The habit of paying for virtual services also needs to be nurtured, so what should companies do next and do they need to adjust their strategic priorities? This is something we have to think about. "

"One obvious option at this time is to verticalize and create AI by using ToB's AI technology and systems to power enterprise customers," Hu continued.

"Moving from 2C to 2B is definitely a big decision, more difficult than making the transition from Forrest Gump into tonight," said Wang Guan-chun. "There is a lot of communication between I and Ichikawa, the only point of conflict is how to adjust, C-terminal is stopped or narrowed? If you do 2B, how to choose the industry? These are not the conflict, but the issues that need to be discussed "

Wang Guanchun continued, "The most important reason for starting 2B is that we feel the demand for ancillary services in many traditional industries is on the rise because some companies are opening up some issues in the C-market and many enterprises are aware of this interaction Very good, but also hope to communicate with the user in this way.Some customers began to find us at the B side.For example, there are brands of maternal and child industries to ask whether the future AI interactivity can be applied to their products , To create a Service Marketing Consultant role - To do this, it now has hundreds of people using innovative channels such as Wechat and Community to create user connections and interact with users and brands Trust, creating more business opportunities to recommend more products and services.But in fact, mothers and new mothers everyday problems and needs are similar, artificial intelligence can improve service efficiency and user experience.This gave us a Implications: Maybe today, we have also accumulated the AI ​​technology and systems output to the industry, faster release price To achieve commercialization. "

"Over the past two years, too, we have accumulated over 3 million users who have the needs of online assistants, and the mother-baby itself is a very clear-cut scenario with at least tens of millions of potential users, working with enterprise customers to quickly reach more users , While constantly improving our experience while training our AI engine. "

"We are also exploring more AI solutions for vertical industries, such as next-generation car makers, to help them build products like smart car-stewards. Today, the 4S stores take control of automakers and consumers Between the car manufacturers hope that through the butler products, consumers can learn more and then the car's pre-sale, sale, after-sale full service in series, and even extends to the car virtual smart assistant system.In this mode , It is also faster to put a vertical industry products landing. "

Hu Wangchuan frequently nodded when Wang Guanchun explained his assistant to expand from 2C to 2B. He said: "The last transition, I did not Champion as firm, always feel that you can try more in the old model.Of course, I was later persuaded him, after all, because of time waiting for people ... ... although that time to burn Speed ​​is not too fast, neither of us are paid and live in the company, but the wrong strategy will consume the most valuable resource of startups, and that's the time. "

"The transformation is similar to this adjustment, though it looks alike, but the challenge is still not the same .First, the last to give up the DVD rental business, personalized recommendation is still from 2C to 2C, but the business model And products to make some adjustments. And this time by the 2C expansion to 2B, greater change; secondly, the first transition, our team is small, we both agreed, the basic OK. This time we have to do More is the idea passed to the entire team, so that everyone clearly understand why to do such a change, and what this change means for the company. "Hu Yichuan said.

"Between us, this time the adjustment is more like a process of intensification by both parties. We can not say who is more determined." Wang Guan-chun said: "Although the patterns of 2B and 2C are very different, the underlying technologies are actually interlinked . Is nothing more than using intelligent interactive AI technology to make direct consumer-facing C-end products, or by enabling end-B enterprises to reach consumers. "

"From my point of view," said Hu Yi-chuan, "While we certainly want to deepen the AI ​​technology on the one hand, on the other hand, we must make the value of the technology reflected in the C-end polished products, AI value can reflect the scene, Not so straightforward now that we believe we must find the right scenario and let us move faster into the positive cycle of value creation. "

"When we started an assistant two years ago, we only had to do our best to help the smart assistant product." Two years later, we went through an intensive study to clarify the specific path to achieve this goal. "Hu Yi-chuan continued: One of our important perceptions of today is that the company's mission is to get everyone with an assistant. Our C-end products are moving in this direction and 2B will not leave this mission. After the logic is clear, The two of us and within the company soon reached an agreement, and the whole team was very much in favor of the transition as well as being an assistant to an assistant in the field of industry, such as retail, travel, education , Auto, financial, medical and other industries, can do intelligent advisers and assistants. "

Participate in Microsoft Accelerator Demo Day

AI partners must be complementary? Not necessarily

This is indeed the case of two partners who have reached a consensus and solved the problem through full communication. The two have similar academic backgrounds, know each other very well, recognize each other and share sympathy. As they have described, there has not even been any fundamental disagreement since starting a business together. So, the two are not "complementary" enough problem?

Xu Xiaoping, the angel investor of the first two projects, once said that partners should be "complementary". "The most painful experience we've invested in is to cast a scientist without commercial awareness ... The perfect thing is that this technology is both commercially viable, but more often than not, technology may simply want to be technology, Next, we need to find a business expert, a management expert.I value a startup team, there are several 'Chinese partners', a tube technology, a tube market, a tube supply chain. "

In response, Wang Guan-chun said: "After the division of labor cooperation, in fact, gradually formed in the work of at least we are looking for partners that point in time, really did not consider the 'complementary' thing we see the success Businesses, such as Google's co-founders, have very consistent values ​​and academic backgrounds, and are aligned with Vision for the future. This is a solid foundation for cooperation . Our idea at the time was that we were young partners and the company started up Later, there will be more experienced and enthusiastic people to join, just as Google later attracted Eric Schmidt, in the management and commercial naturally better complemented. "

Hu Yi-chuan believes that more important is learning ability, he said: "We were very young at that time, with no experience, perhaps the biggest bright spot is that we all have a strong desire to learn.I approve Guanchun's learning ability, Reading is machine learning, but he can also go to Goldman Sachs, McKinsey internship, indicating that he has a strong understanding of business and learning ability; although I is also based on theory PhD, but also for a variety of new technologies and products are also very sensitive Interest, can quickly understand and master.I think this is very helpful for our first business.For the first time in Beijing, the team only two of us, but also while trying to determine the division of labor, as much as possible to maximize each other's strengths. "

Wang Guanchun sitting next to this view is very much agree, he added: "Even our own not good direction, but also through quick learning to improve their ability, or recruit people to fill vacancies. I think learning ability should be our One of the core competencies. "

In 2017, the two walk 100 kilometers together

Ideal partner in the eyes of AI partners

As two consecutive entrepreneurship partners, their ideal investor is like?

Wang Guanchun said: "First of all, he should have a strong belief in artificial intelligence and assistant products, because it seems that the maturity of AI technology, especially the cognitive AI technology represented by natural language processing, takes time. It takes a while for smart assistants to come out as AI products, which requires investors to have patience and foresight. "

"The second point, we hope investors can bring some resources.For example, we first introduced the speed of light, Sequoia, real and surging such investors, is hoping to quickly reach more service providers to help assistants to landing faster Various service skills. At the time of fierce competition, to also docking the major o2o platform, the speed is the fastest .This with our good investors endorsement, there are even investors matchmaking, there is a certain relationship.A round We are fortunate enough to get funding from Microsoft globally, and we chose Microsoft because we think it can communicate with us in terms of AI technologies such as deep learning and how to build AI products. In addition, Microsoft is doing business services Our experience and client resources come and go and can be learned and utilized, so attracting some experienced and resourceful investors through financing is our key consideration. "

AI entrepreneurs who can not find the application scenario will fall into the dilemma of not making deep technology and making money - advice to AI entrepreneurs

Describe the ideal portrait of the ideal investor, at the end of the interview, two PhD began to start a "veteran", the new AI entrepreneurs put forward some suggestions.

Wang Guan-chun said: "AI technology itself has three stages of development: perception, cognition and decision-making. The first to do AI companies may first cut into the perception level, such as face recognition and speech recognition. Knowledge and decision-making AI, involving language understanding, knowledge reasoning, many rounds of dialogue, etc. I think that if today's business, we must consider the cognitive or even decision-making AI scene, and strive to obtain the first mover advantage. The decision-making, business value from the creation is also very close.How to find the right scenario to solve the problem of the border, how to make your AI ability to quickly release the commercial value of the start-up company is crucial.If you can find that you Of companies will quickly rise; if you can not find it, it is likely to fall into the technical no data, not only do not deep technology, but also make money difficult situation.

Hu Yichuan said: "Today, what we call AI is more of a new technology, not an industry. Unlike the search, both technology and industry aside the scene, AI can not exist as a separate industry.Many entrepreneurs, In particular, a lot of technical background AI entrepreneurs, will want to make AI as a platform and my feeling is that even today you can make the platform, this event will eventually be dominated by large companies.If the start-up companies to To do it well, big companies are actually more likely to devote more resources to doing it better, because platformization is what big companies are best at. In combination with this, I think startups should do more Of course, the premise is that the traditional industries have data, are willing to make intelligent attempts based on these data. Today, large companies in the field of AI are often Internet companies, such as BAT, they It is not likely to go deep into a project that combines AI with traditional industries, where startups have both opportunities and advantages. "

At this point, the new intelligence unit of the two Chinese AI partners interview, which lasted nearly three hours. After the interview, the two plans to Hu Yi Chuan's residence, a rare look at a Premier League match. They appreciate the team's tactical style are even similar. The team not only pursue fast forward, pleasing play, but also emphasize the overall offensive and defensive, keep the formation; both emphasize coordinated operations, teamwork, but also fully encouraged talented individual performances - they are Manchester United fans. At the same time, Manchester United focus on youth training system and echelon training, which in their eyes is also building a corporate culture, inheriting the company's excellent gene model. Assistant to have their own football team, from time to time training and competition. Hu Yi-chuan kicked the back, claiming to be "calm-minded, have the overall situation concept, running active"; Wang Guan-chuan said that as they grow older, the role gradually by the "strikes forward" forward to the "front of the needle." After years of "cooperation", these two partners who have enjoyed each other's appreciation and trust from the beginning of the cooperation have already reached a tacit understanding and have become veritable gold partners in the Chinese AI venture.