黄峥在 2017 钛媒体 T-EDGE 科技生活节的演讲（技术改变生产力）

Speech at the 2017 TMTPost T-EDGE Festival: Technology Changes Productivity

> By Colin Huang · 2017-07-23 · [Original link](https://www.tmtpost.com/2702534.html) (Chinese)
> Our translation from the Chinese original.

Good morning, everyone. First, thanks to TMTPost for giving me this opportunity, and especially for putting me in the session on technology changing life, because ever since I was a student I've been a technical person.

Today's theme is AI. From university to my time at Google, what I did was always AI work, and I finally have a chance to talk about my own feelings about what I learned before. I may not be able to talk about many big industry events, nor much about what will happen in the next 10 years.

As an old AI technical person, I'll just talk about how I see AI.

Before talking about AI, I'll still spend a little bit of time introducing what Pinduoduo is, because for many people Pinduoduo is still a new name.

Pinduoduo is a third-party, all-category social e-commerce platform. The previous generation of e-commerce was going to search when you wanted to buy some specific thing; Pinduoduo is more often non-purposeful shopping, like strolling in a Shopping mall: when you go you have no clear purpose, you walk over there, feel it's very interesting, and buy it.

As my third, or fourth, startup project, I also feel especially lucky about Pinduoduo. Pinduoduo's whole time hasn't been long, only a little over two years now. In one year we already had over 100 million paying users, and now we have over 200 million paying users. The speed of user growth and the momentum of continued growth have exceeded many people's expectations, including the founding team's expectations.

Looking back at what drove growth like ours: from one latitude [Translator's note: the original writes 纬度, "latitude", a common miswriting of 维度, "dimension"; the sense is "from one dimension"], it's the business model, and I think it's mainly two points. First, doing a relatively right thing at a good time. When we started Pinduoduo, social networks were already very developed; everyone was using WeChat, Weibo and all kinds of social means, and social media already accounted for more than 40% of all mobile internet usage time, which you can see clearly from your phone's power use.

But the commercialization of social media is very poor: the transaction volume converted from 40—50% of people's eyeballs may account for only 8% of total social retail sales. Compared with the contraction-style e-commerce of purposeful shopping [Translator's note: the original writes 收缩式电商, "contraction-style e-commerce"; in context it is probably a mishearing or slip for 搜索式电商, "search-style e-commerce"], this is a huge mismatch. At a time when the market capacity was this huge and mobile payment was exploding, we used a model to unleash this thing. In the past two-plus years, what I'm proud of is that under this model our team showed differentiated thinking and different genes.

Before Pinduoduo we also did e-commerce outsourced operations, and also did games. You could say that in this industry it's very rare for the same group of people to have both successfully built an e-commerce operations company and also built a games company, because an e-commerce operations company is oriented toward strong operations and strong execution; it's Alibaba's thinking. But when you go into games, you're learning product thinking, learning Tencent's thinking. This angle of business model and team, which most media and investors have overlooked, was crucial in our growth.

Past technology, including today's AI, emphasizes efficiency, emphasizes how we use technical means to raise efficiency: making cars go a bit faster, making buying things a bit more efficient, going from producing 5000 pieces in 1 hour to producing ten thousand. At present eighty or ninety percent of AI projects put the data accumulated in human history into one place, return it to the same place, and use machines to learn humans. This is centralized AI. Centralized AI, to put it figuratively, is Google trying to make a super-powerful brain. This super brain understands everything; you ask it, and what it tells you is what it thinks is most correct. It's like you go and ask God what you'll eat today, and God says you'll eat beef today.

But in life you rarely go and ask God; most of the time you go and ask friends. You don't know what school your child should go to, so you don't go to the temple to ask the priest, but to ask people who have been on the same journey: which school around here is good, which teacher is good, what the application route is. Then, if you don't get the right answer, you'll go on to ask the next person. After going around, no one's answer is complete, and at that point you'll stop. At this point the structure of AI is very different from the centralized kind before. In this structure there's no emphasis on complete search; what it emphasizes is equal, spreading exploration using social relationships. In the structure there's no centralized God; in an equal network there is a host, and there's no centralized decision-maker either. [Translator's note: the original reads 在平等的网络里是有主持人, "in an equal network there is a host"; the context ("no centralized God ... no centralized decision-maker either") suggests 没有主持人, "there is no host", was likely meant. Translated as written.] These are two different forms of social organization.

What are the advantages of this form? The first especially obvious advantage is that it can help machines, help algorithms get out of their cage. From face recognition to text recommendation, the algorithms used are nothing more than the multilayer neural algorithms talked about back then. They predict what you want next based on your historical knowledge. What kind of situation arises then? For example, today in a shopping scenario, you looked at a pipe today, maybe you looked at a pipe again tomorrow, so it thinks what you want to see is pipes.

You want to read articles; today you read about cars, tomorrow you read about cars again; when you've finished buying a car, it still thinks what you like is cars. And the more data you accumulate, the more it thinks this is what you like. This is like when I used to do robots at school: in path selection, the machine learning would walk into a corner and couldn't get out, walk into the cage of its preferences and couldn't extricate itself. If you face only your own inner self, alone, sometimes it's easy to walk into an obsessive place.

People are social animals. A very important point is that you have friends, and through friends you experience things you haven't experienced. Take having a child as an example: having a child is a matter of one stage of life. Before having your first child you haven't read the books on before having a child, and you don't have knowledge in this area. But many friends have been through it. If you use a God-style AI algorithm, you may never get knowledge in this area; only after you've had the child will it recommend related things to you. Only by using the distributed way of friends will other people's knowledge be absorbed, letting you explore a new world with your friends' eyes.

The second advantage is that it can reflect human emotions. When we talk about artificial intelligence it's ice-cold, and even when it describes heartwarming scenes it's about how to raise our efficiency. But as people, where we differ from dead people is that I have warmth, I can have emotions, I quarrel, cry and laugh. This kind of emotion has a very good place where it shows in economic life, and that's the stock market.

The stock market is irrational. Sometimes, when a bunch of people hype it, this stock will soar; sometimes it will crash. This emotion itself is sometimes not good, but a lot of the time it's also very interesting. Bringing it into our shopping scenario: for many girls shopping itself is a kind of entertainment. You go to a Shopping Mall, you try on clothes in all kinds of stores, and your happiest instant is the instant you swipe the card; in this process you have emotion in it. Many people are looking at this piece of clothing; the feeling the surrounding scene gives you is different. This kind of distributed artificial intelligence has the chance to build human emotion into machine learning and into the network.

Suppose that suddenly 100 friends come and ask this robot about this piece of clothing, or this product; the robot will say this is very hot, his choice is very right, and he'll be happier. Another situation is that nobody comes to ask you about this piece of clothing, or everyone else is mocking it. Linked together, this emotion is group emotion, and it also shows in shopping, like the stock market's chasing rises and dumping on falls; this is how some hot spot, some fashion, shows up in the shopping scenario.

Third, I think it's fundamentally different from centralized AI. Centralized AI tries to use AI to replace the work of the brain: people only need to ask the machine; it calculates faster than you, and it feels like people are useless. Distributed AI has no central neuron. It puts people into the loop; each person is part of the neurons, influences the whole neural network, and is part of the artificial intelligence network. You can show your personality, you can express your attitude, and you can influence your friends and also influence your network.

This is very interesting. It isn't just using technology to raise efficiency; it's using technology to raise the flourishing of each person's personality, letting people connect more deeply and have better means of expression, and you can also influence more things through your own personality.

In the next twenty years, Pinduoduo, as an e-commerce company and a technology company, will keep working hard to merge technology with life, merge technology with human emotion, and try to use technology to make people happier, rather than simply making people more efficient.

Thank you all; I'll stop here.

Copyright: © LookMinds. Our own edited text and translation, not a substitute for the original; rights in the original belong to the author and the platform. Please see the original link.