Interview Wunda: I don't care if AI turns evil.

Interview Wunda: I don't care if AI turns evil.

It can be said that these universities have the best computer science courses: Carnegie Mellon University, MIT, Berkeley, California and Stanford. These are the places where Wunda received a bachelor's degree, a master's degree, a doctorate, and 12 years of coaching.

Wunda is a benchmark in AI, and he is not even 40 years old. In 2011, he founded the Google Brain, a deep-learning project that has been brilliantly powered by Google's strong computing power and massive data. When the computer analyzed a lot of YouTube screenshots and was able to identify a cat, Google's brain rushed to usher in one of its most important achievements. (The New York Times "How many computers do you need to recognize a cat?") 16,000 units. "For the title to be reported. Wunda himself explained: "It is noteworthy that the system has discovered the concept of" cat ". No one accused it of what a cat is. This is a milestone in machine learning. 」

There was a Wunda of joy in the placid stream. He talked about all the mistakes and setbacks in his career and the papers he didn't understand. He wears the same blue Oxford shirt every day. He was both shy and proud when his colleague brought up his lovely robot-themed engagement photo with his current wife. His wife is a surgical robot expert Carol Reiley. (Pay attention to his blue shirt in the photo!) ）

He has a softer voice than anyone, but that doesn't prevent him from becoming a popular orator. In 2011, when he hooked up his Stanford University machine learning video to the Internet, 100,000 people enrolled in the class. In less than a year, he co-founded Coursera, the largest online course platform to date, with partners including Princeton, Yale and top schools in China and Europe. Although all courses are free to open, this is a profitable business. Wunda once said: "Charging for content will be a tragedy." 」

Then, last spring, something terrible happened. Wunda announced that he would leave Google and gradually fade away from Coursera's daily operations. China's technology giant, Baidu, has aggressively spent $300 million to build a research lab, and Wunda is the head of the project, along the path of Google's Silicon Valley headquarters to engage in artificial intelligence.

As before, in Baidu Wunda also tried to let the computer with amazing accuracy to identify audio and images, but also real-time oh. Wunda believes that 99% speech recognition accuracy will revolutionize human-machine interaction and the design of the operating system. At the same time, Wunda will also help Baidu service millions of search users, digital life for them is still new. "(in China) you will receive some problems that will not occur in the United States. "Wunda explained:" For example, we will receive the question "Baidu Hello, I last week in the corner of the shop inside eat noodles, super delicious." Do you think they'll get a discount this weekend? "This kind of problem. "We are trying to answer these questions," Wunda added. 」

Elon Musk and Stephen Hawking have sounded a wake-up call to the potential threat to human beings from advanced AI, but Wu En does not. "I don't study how to prevent artificial intelligence from becoming evil, just as I don't study how to manage the Martian overpopulation. Wunda says Ai has to evolve into something similar to consciousness for several decades (or longer). At the same time, there are more pressing issues. Machine learning brings about a rise in computer power that is replacing some of the work that human beings have been doing for a long time. This trend will only accelerate, and Wunda calls on policymakers to be prepared to deal with the socio-economic consequences.

At the newly established laboratory in Sunnyvale, Calif., we discussed with Wunda a post-delivery project to sophia--A life experience lesson from a diverse group of people. He explained to us why he thought "following your passion" was a terrible career proposition, and shared his own strategies in teaching innovation; Wunda also talked about his failure experience and some useful habits, the most influential book he has ever read and his latest view of the AI frontier.

Q: Recently, you said: "I've seen people become more creative by learning." "Can you explain it?"

A: The question is, how does a person create a new idea? Is it an isolated act of genius that is unpredictable, such as Steve Jobs's special kind of people? Or is it that innovation can be taught and systematically?

I believe that innovation and creativity can be taught. We have a way to make people more systematic or creative in a systematic way. At Baidu, I have been working on a workshop on innovation strategies, in which innovation is not a random and unpredictable act of genius, but rather a behavior that can be very systematic in creating unprecedented things.

In my own life experience, whenever I'm not sure what to do next, I do a lot of studying and reading and talking to professionals. I don't know how the human brain works, but I find it almost magical: when reading enough or talking to the pros, that means when the input is long enough, the new ideas will start to emerge. A lot of people I know seem to have had this experience.

When you get the latest level in your chest, you stop randomly selecting ideas. In the selection process, you become thoughtful and know how to combine these ideas. You'll be careful when you come up with new ideas or eliminate the idea of modifying them.

Now there is a challenge-how will you deal with these new ideas and how to strategically turn ideas into useful things? This is quite another thing.

Q: Can you talk about your information acquisition habits (information diet) and how do you learn?

A: Read a lot and talk to people a lot. In my opinion, the most effective way to learn and access information is two: reading and talking to professionals. I spend a lot of time on these aspects. There are only 1000 books in my Kindle, and I probably have read two-thirds of them.

In Baidu, we have a reading interest group, about half a week reading a book. In fact, I joined two reading groups in Baidu, each of which reads half a book a week. I'm the only one who's involved in all the reading groups (laughter). My favorite activity is to study at home in Saturday.

Q: Can you talk about the impact of early education on you? Have parents ever done something that affects you far beyond what most parents would not do?

A: At the age of six, my father bought me a computer to help me with my programming. Many computer scientists have been learning programming from a very young age, so it's nothing special, but I'm glad that I'm one of them, and I can have a computer learning program at a very small time.

Unlike the old-fashioned Asian parents, my parents are relaxed. Whenever I get good grades at school, my parents don't make a fuss, their reaction really makes me feel a little awkward, so I'm used to hiding good grades (laughter). I don't like to show my report cards to my parents, not because of bad grades, but because of their reactions.

In addition, very fortunate, I can grow and work in many different places. I was born in England, grew up in Hong Kong and Singapore, and got to college in America. For personal academic research, he received a degree from Carnegie Mellon University, MIT, Berkeley and then to Stanford.

I am very glad to have walked through these places and met some top talent. At T-Bell Labs, the world's leading laboratory has just been set up, where I interned, and then to Microsoft Research, the opportunity to see a lot of different points of view.

Q: Are there any things in your education or early career that you would have had a different choice? and lessons to be learnt from others?

A: I hope our society can give young people some better career advice. I think that the phrase "Follow your passion (Follow your passion)" is not a good career advice. is actually a lousy career-planning proposition. your passion for driving does not necessarily mean that you should aspire to be a racer. In real life, "following your passions (follow your passion)" is actually revised to "follow things that are not only your profession, but you are also full of enthusiasm." " Usually, you have to be good at something first, and then you are passionate about it. I think that almost everyone can become good at anything.

So, when I think about how to spend my life and what I want to do, I refer to the two criteria. First, is there a chance to learn something. Does this project allow me to learn something new and interesting and useful at the same time? Second, the potential impact. There are countless interesting problems in the world, and there are countless important problems in the Earth. I prefer people who focus on the latter kind of problem.

Fortunately, we can always find a valuable opportunity to engage in work that has both great potential and excellent learning opportunities. If young people are able to maximize these two things, they usually have an excellent career.

Q: Will you define the importance of this impact event primarily based on the number of people affected?

A: No, I don't think quantity is the only important factor. Dramatically changing the hundreds of millions of human lifestyles, I think this is a degree of influence that we are very eager to reach. It's also a way to make sure that what we do is not only interesting but also influential.

Q: You've talked about some of the failed projects before. How do you deal with failure?

A: Well, this kind of failure often happens, so it's going to be a long story (laughter). A few years ago, I had a list on Evernote that tried to record all the items that started but didn't work. Sometimes luckily, in a completely unexpected direction, the project was conquered by me, of course, more luck than skill.

However, I made a list of all the projects that had been researched but did not have any results, or projects that were not successful, or the ones that were totally out of proportion to our upfront effort, and then I tried to classify them according to the reasons for failure and try to do a rigorous post-mortem. One of the failures was at Stanford when it was inspired by a V-shaped wild goose, and for some time we tried to get the aircraft to fly in this formation to understand the energy savings. The reasons for aerodynamics are very reliable. So it took us about a year to get these planes to fly automatically, and then let them fly in formation.

But a year later, we realized that it was impossible to control the aircraft fully and accurately to understand energy efficiency. Now, if we take the position into consideration at the outset, we will realize that these small planes simply cannot achieve our research purposes. The wind can scrape the plane far away from the precise position of the formation when it flies. This is my past often made a mistake, in the project, the first step, then complete the second step, followed by the third step, then you realize that the fourth step from the beginning of the end there is no way to achieve, fortunately, this error is now a lot less. I've also talked about this example in the workshop on Strategy innovation, which teaches me to try to eliminate risk early in the project.

I am now better able to identify and assess risks at an early stage. Now, every time I say "we should reduce the risk of the project early", everyone nods and agrees, because it's obviously possible. But the problem is that when you really need to conquer some new projects in this situation, it's hard to apply to the project you're working on.

Because this is a strategic decision-making skill. In our education system, we are good at teaching facts and procedures, like recipes. How do you cook a traditional spaghetti sauce? Of course, according to the recipe. We are good at teaching facts and recipes.

But innovation or creativity is a vital skill, and you wake up every morning in a unique situation that no one else has, and you need to make good decisions in a completely unique environment. In my opinion, the only way to teach this strategic skill is through cases where you need to look at hundreds of cases. When we have seen enough cases, the human brain will continue to learn and assimilate the rules and guidelines that make people make good strategic decisions.

In general, I find that it takes years for a researcher to penetrate those cases and absorb the rules. So, I'm doing some experimental exploration here, trying to build a "flight simulator" for an innovative strategy. Here, you don't have to spend five years trying to fathom a case, and we'll give you a lot of cases within a much shorter timeframe.

Just like a flight simulator, if you want to learn to drive a Boeing 747, in reality, you'll need to drive for years, or just a few decades to meet a sudden emergency. However, in the Flight Simulator, we can simulate various emergencies for you in a short time, allowing you to learn faster. This kind of project is what we are now exploring.

Q: When this lab was first established, you mentioned that you had not been aware of the importance of team culture for a long time in your career, and then gradually began to recognize its value. What do you realize in the past few months about how to build the right team culture?

A: Many organizations have a clear team culture slogan, such as "we give each other strength" and so on. When you shout out slogans, everyone nods and agrees, who doesn't want to give their team mates the strength? But when they returned to their desks in five minutes, did they really do it? It is difficult for people to translate abstract slogans into concrete actions.

In Baidu, we have done a few business for team culture, at least I have not heard of other organizations to do this attempt. We did a set of tests to describe a particular situation to the employee in the title, and the question was, "What do you do when you encounter something in a situation?" A,b,c or d? 」

No one can get a full mark on the first answer. I think that the interaction of a test question, which gives the team members a concrete response to the hypothetical situation, is the way we connect the abstract culture and concrete actions, what would you do if your team partner did something to you?

Q: What books have a significant impact on your knowledge development?

A: Recently I was thinking about what books I would recommend to those who want to innovate and create new things. The first is Peter Thiel's "from 0 to 1: Opening the secrets of business and the Future," a book that outlines entrepreneurship and innovation well.

We often divide business into business ("Enterprise-to-enterprise", that is, corporate customers are other enterprises) and business-to-consumer ("Enterprise to consumers"). On the business side, I recommend lean entrepreneurship (the Lean startup). In terms of consumer-to-consumer aspect, my favorite one is "crossing the chasm (Crossing the Chasm", although the angle of view is narrow, but gives the concrete strategy of rapid innovation. Although its scope is somewhat narrow, it is very well written in the parts it deals with.

A further breakdown of consumer-to-consumer aspects, my favorite is two. One is "talking to humans", the book is very short, teach you how to communicate with users to develop and user empathy. Another is the Miaoshouhuichun: Website Usability Testing and Optimization Guide (Rocket surgery Made easy). If you want to create important, customer-related products, this book will teach you about the different means of users, whether it's user research or user surveys.

The last is the hard (the hard Thing on hard things), although it is a bit dark, but it involves a lot of useful knowledge about the construction of the organization. For those who are trying to make a career choice, there is an interesting book called So good they Can ' t Ignore, which gives a valuable perspective on how to choose a career path.

Q: Do you have some helpful daily habits?

A: I wear blue shirts every day, I don't know if you have found them. (laughter) Yes. One of the most important levers in life is your ability to develop useful habits.

When I talk to researchers and people who are interested in entrepreneurship, I tell them that if you insist on reading papers, such as studying several articles in earnest every week and sticking with them for two years, you have learned a lot after two years (Option2: You will reap a lot in two years). This is an excellent investment in your long-term development.

But this kind of investment, for example, you spend a whole Saturday studying instead of watching TV, no one will pat your back and tell you that you are doing well. It is likely that your Saturday full-day study will not make your work in the next Monday more productive. The short-term rewards of such efforts are minimal or even rudimentary, but it is really a great long-term investment. This is really the way to become an excellent researcher and you have to read a lot.

People who want to rely on willpower to persist (learning or other long-term investments) almost never succeed, because the will is depressed. Instead, I think people who like to create habits – you know, study every week, work hard every week – these are the most important things. These people are the most likely to succeed in a group of people.

For me, one of my habits is to follow an app every morning for seven minutes of fitness. I find it much easier to do the same thing every morning because you need to make a few decisions. My wardrobe is full of blue shirts for the same reason. I used to have two shades of shirts, blue and magenta. I think there are too many decisions that need to be made. (laughter) So now I only wear blue shirts.

Q: You urge policymakers to take the time to think about a future scenario in which a large part of human work is being computed and robots cut. Do you have any way out?

A: This is a very difficult question. The computer excels at the repetitive work of machinery. So far, machines have been adept at automating the task of repeating tasks every day.

Now this may be a few points on the spectrum. The workers on the assembly line repeated the same procedure for months, and now the robot can do the work. Intermediate challenges such as driving a truck. The truckers do similar things every day, so the computer is trying to do the same. This is more difficult than most people think, but the possibility of autonomous driving may be what happens next decade. Then, more advanced tasks, such as radiologists, need to look at the same type of X-rays every day. In the same way, computers can show their own skill in such fields.

But I think, for a long time, humans are better at doing non-mechanical, non-repetitive, socially powerful tasks than robots. A lot of our work needs to do something different every day. Get in touch with different people, arrange different things, and solve problems in a different way. At present these tasks are difficult for the computer to complete.

The challenge is that when the US shifts from an agrarian economy to an industrial and service economy, people shift from a programmatic effort, such as farming, to another programmatic job, such as manufacturing or working in a service center. The majority of the population had done so, and they had found other jobs. However, most of their work is usual and repetitive.

The challenge for us is to find a way to impart (in size) a change to non-routine, non-repetitive methods of work. Historically, our education system is not familiar with this way. Top universities can only offer this kind of education to a few. However, the vast majority of people are still engaged in important and routine duplication of work. This is the problem that our education system faces.

I think the problem can be solved. That's why I've been thinking about educational innovation strategies, that is, teaching innovation strategies. We need to allow a lot of people to engage in non-routine, non-repetitive work. The strategy of teaching innovation and creativity, as well as the Flight Simulator, will be the way to achieve this goal. We haven't come up with a concrete solution yet, but I'm optimistic.

Q: You once said, you said, "At work, engineers in China are more hardworking than Silicon Valley engineers." The engineers at the Silicon Valley start-ups are working really hard. In a mature company, I did not see the start-up companies and Baidu as the intensity of the work. "Why do you think?"

A: I don't know. The engineers in both places are very good individuals. The difference may lie in the company. Baidu's team of engineers is very flexible.

Not to mention the Chinese Internet economic situation, I think, the challenge of various ideas, all up for grabs, will have greater significance. China's Internet ecosystem is very dynamic. Everyone sees great opportunities and everyone sees great competition. The situation has been changing all the time. New inventions in the birth, big companies will suddenly one day into a new field.

I'll give you a clue. In the United States, if FB intends to develop a new web search engine, it will feel a bit strange. Why does FB want to do search engine? This is really hard. But in China, this kind of thing will not be so incredible, where there are more assumptions that it will be a new creative business model.

If the video does not play, click here to try

Q: It seems to be a different management culture, where you can quickly make important decisions and keep these decisions smart and efficient without confusion. is Baidu operating in a unique way, and do you think this is particularly helpful in its growth?

A: Gosh, good question. Baidu pushes the decision-making process down to a far-off place. Employees have a lot of autonomy and they are also very strategic. I really appreciate the company, especially the management, that they have a very clear vision of the world and competition.

When management meets, we don't pretend to talk to the whole company. "We did a great thing at that point. We're not happy with those things. This is progressing well. The progress is not going well. These are the things that we should highlight. Let's look at the mistakes "Obviously there is no bluff here, and I think it provides a great environment for companies that need innovation and focus."

Q: Relative to other issues, you will be more inclined to focus on the issue of speech recognition. In the problem you are facing, which problem can be solved to lead to a huge leap in the accuracy of speech recognition technology?

A: We are building a machine learning system for speech recognition. Some of the machine learning systems we're using are in fact decades old. But only in recent years has the real results begun.

What is this for? I usually make comparisons by building rocket ships. The rocket ship itself is a monster, and it needs to carry a ton of fuel, both of which require very large space. More fuel, small rockets, not flying, big engines, less fuel, can take off, but can't get on track. It can be seen that giant engines and a lot of fuel are necessary.

Machine learning is now on the rebound because we finally have the tools to build giant engines-giant computers, and this is our rocket engine. And the fuel is the data, and we finally get the huge amount of data we need.

Social Digitalization has created the possibility of large amounts of data, and we have been making the data for some time. But only in recent years have we been able to build enough rocket engines to digest these data fuels. So building bigger engines and getting more data is part of the way we implement speech recognition, but not all of them.

A little technical about what we've done. Where do I get a lot of data for speech recognition? One of the things we do is extract voice data. Other teams may use thousands of hours of data, while we use 100,000 hours of data. We need a lot more data fuel than the rocket fuel that can be read in academic work.

Then we'll do another thing, assuming we have a piece of your audio clip, we'll add some background noise to your audio, as if the audio was recorded in the café. We synthesized an audio clip that sounded like you were talking in a coffee shop. By synthesizing your voice in a large number of backgrounds, this has doubled the amount of data we get. We use this strategy to create more data for our machines, our rocket engines. There is also an important fact about language recognition: Most people don't understand the difference between 95% accuracy and 99% accuracy. 95% accuracy means that there is a typo in every 20 words, and every time you use speech recognition on your phone, you have to go back to correcting typos, which is painful.

And 99% is completely different, if you can achieve 99% accuracy, speech recognition will become reliable, once it works, you can always use it. So, the 4% difference is not just a gradual improvement of 4%, but a far cry from being used and almost unused.

Q: So what are the difficulties that now hinder us from achieving 99% accuracy?

A: We need bigger rocket engines and more rocket fuel. These two are still restricting the development of speech recognition technology, we need to develop both. We are still trying to break through the limits.

This article was selected from the Huffington Post, the machine heart compiled and produced, participating members: Shodan, Chao Kelsie, Zheng Lao, 20e, Chen Xiaoyong vitarte, chubby.

Interview Wunda: I don't care if AI turns evil.