AI expert Li Mu's "Reflections on Five Years of Work" has become popular, netizens: I was impressed again

Jin Lei reposted and compiled from Zhihu
Quantum Bit Report | Public Account QbitAI

When entering the workplace from the ivory tower, many people feel lost or confused:

How should we walk this long road? And how can we walk it well?

Recently, Amazon's senior chief scientist and AI expert Li Mu published an article titled "Reflections on Five Years of Work" on Zhihu based on his five years of work experience .

In just 2 days, it received more than 5,600 “likes”.

After reading this article, many netizens expressed their feelings, "I have gained a lot" and "I am impressed again."

Next, let’s take a look at this article by Teacher Li Mu.

(The following content is reproduced and organized with permission)

Five years ago today, I flew to Seattle for an interview at Amazon. After the interview, I took a red-eye flight to Boston overnight to attend my wife's doctoral defense at MIT. Halfway through the defense, the phone rang. The person on the other end said congratulations on passing the interview and wanted to discuss salary. I said that I only interviewed with you and I would just give you the salary, and then hung up.

The next day after the defense, I went to the Municipal Bureau with my wife to register for marriage. While queuing at the door of the oath-taking hall, the boss called and said excitedly that I could do this and that after I came. I said that it was great, but I had to get married first. The boss was stunned, congratulated him, and continued to talk. I had to interrupt: We have to go first, it's our turn to go in and take the oath.

Five years have passed in the blink of an eye. From the outside, the biggest change is that I have two more children. But the biggest change comes from cognition, which is the change in my outlook on life, the world, and values. When I graduated with a doctorate, I wrote about my experience - "Five Years as a Doctor" (see the link at the end of the article) .

Many students left messages saying they were inspired. Now I want to share with you my experience and insights from the past five years. More precisely, I want to share with you the lessons of failure, because each point is a lesson learned after paying the tuition fee. I hope these can also be helpful and inspiring to you.

The value of things is their value to society

When you are studying, you have clear goals, such as test scores, further education, or finding a good job. The biggest difference after work is that you have too many goals to pursue. The change brought about by this is that you need to decide which things to do now, which to do later, and which things you can leave alone.

Prioritization should be based on the value of things. I now evaluate the value of a thing based on its value to society, which can be expressed in the following formula:

Number of beneficiaries x Average time per person x Value difference per unit time

The number of people who can benefit from an event and the average time they benefit are the attributes of the event itself. The third item depends on how well you complete the event, that is, you do it better than others do similar things, so that the beneficiaries benefit more from you than from others.

This formula can be used for a variety of different things, and we will continue to use it. Here are a few examples. For example, great products generally have extremely high value. Take WeChat, for example. It is a mobile communication software that is used by billions of mobile phone users, and each person uses it for several hours a day, so the first two items of its value are very large. Because WeChat has a good user experience, the value it brings to users compared to other alternatives is the value difference. So WeChat is a very valuable product.

Let's take a small example. For example, you lead someone to write a paper. The number of people affected by the paper is the size of the research field; the time of impact is the time it takes for someone to do a research related to your work, which may be only a few months in a lifetime; the value difference is the contribution of your research over the work of predecessors. In this way, you need to write popular and epoch-making papers to achieve high value. But we know that a paper generally does not contribute much and only a few people will read it, so it is basically worthless. Why do people still actively "flood" it?

Here we need to look at two values ​​in detail: First, you become familiar with a new field or new method through this research, which is valuable for you personally. Second, you can improve others' ability to think of methods, do experiments, and write papers when you lead them to do research, which is of great value to them. Therefore, even if you are a well-known researcher, your name will appear on many papers that are like exercises by novices.

Let me give you another smaller example. In the past four years, I have spent more time taking care of my children than working. For quite a while, I felt that my career was delayed by my children. Until later, I used this formula to calculate: Although there are only two people, the benefit time is quite high, more than 50 hours a week. And there is only one father, and there is a huge difference for the children whether I am with them or not (my own feeling), so the last two items of the value formula are very high. In addition, taking care of children is also valuable to me personally, including how to understand others with completely different ways of thinking, and constantly fighting against my impulse to be angry, and finally reaching the state of Buddha nature. This calculation makes me feel at ease.

Serving the society is ultimately serving yourself

The value formula above emphasizes the contribution of things to others. Before using it, my value formula focused more on myself. For example, I often use the fun of a thing, or the technical content in it to divide the priority. The problem is that although I enjoy the process of doing things, the sense of accomplishment afterwards is not high. It is a bit like the emptiness after playing a game. Because after finishing it, I often find that although it is cool, it is useless and not many people pay attention to it. The reason is that things that are of direct high value to individuals may not be of great value to others. It is very likely that this thing itself is only meaningful to a small group of people, and each person may benefit for a short time, or it is actually reinventing the wheel, and there are already similar substitutes on the market.

If you optimize the value to society, you will get delayed rewards for yourself. This reward includes the higher level of inner satisfaction when you know that what you do is useful to others, and the gifts given by others when they benefit from you (like you, or your boss gives you a promotion and a raise). Of course, these two do not necessarily appear at the same time. Many times, the value you create may not be noticed by others (decades of maintaining open source toolkits that everyone is accustomed to using), and sometimes people will exaggerate and tout your contributions. You should actively seek recognition from others, which will give you more resources to do bigger things. But you should care more about inner satisfaction, because it is more controllable and not easy to be misled by others. More importantly, it will give you the internal motivation to do things better, which is the foundation for your continuous growth.

Technology ultimately serves products

Students majoring in technology usually continue to work in technology when they first join a company. When I just graduated, I thought that joining a big company meant working in technology and becoming one of the best technical experts in the world. And I didn’t want to work on products, because if I work on products, why don’t I start my own business? The money I earn is still my own. There is nothing wrong with the second half of the sentence, but the first half ignores the fact that technology ultimately serves products. Although the expected time for technology to enter products varies from company to company, it is usually between six months and five years. There are only a handful of companies that expect more than five years, and most of them have passed away. So even if you are in the company's research department, you should know the company's expectations for the time it takes for technology to land. Otherwise, when the time comes, you will face the company's reduction of investment in technologies that do not meet expectations. The worst case scenario is that you are on the news: the director of a certain company's research institute resigns, and department members go their separate ways.

So what kind of technology can be used in products? Usually you will have a general idea based on the company's existing products. Next, you need to know what the main value of this product is (applying the previous formula). Then you need to figure out the value of your technology to this product. If your technology can improve the core functions of the product, even a little bit, you will get resources to implement the technology. For example, improve WeChat's video compression technology, Toutiao's recommendation algorithm, and Apple's shell materials. On the other hand, if you don't grasp the main trunk, such as WeChat skinning, Toutiao web version acceleration, and Apple operating system compatibility with other hardware. Even if you can do a lot better than the existing technology, the product team may not be motivated to help you, or even tell you not to do this from the beginning.

So whether you are working in a product team or in a company research institute, you should have some understanding of the value of the product. For example, you should have a deep understanding of the product manager's mantra: market, rigid demand, pain points, and high frequency. At the same time, you should also know the value of the technology you have to the product and use it to guide your planning of the technical route.

A soldier who doesn't want to be a general is not a good soldier

People's satisfaction comes from comparison, whether it is comparing with others or comparing with their past. This desire drives you to pursue things with greater value. This means you need more resources to become bigger and stronger. At the very least, you need a team. You may be the manager, leader, or both of these teams.

Maybe you prefer to work on your own, at least that's what I thought at first. But as your abilities grow, others expect more of you, and you'll inevitably have to lead a team. Otherwise, you'll have to find satisfaction elsewhere. Rather than being pushed into this position, it's better to be prepared from the beginning.

There are a lot of workplace books for reference. My own experience is very simple: leaders are the ones who lead the way and need to have a good vision. Managers are the logistics officers who make the team perform efficiently. The following explains these two points respectively.

Looking ahead three years

The most important thing for a leader is to find the right direction when leading the team to explore unknown areas. In other words, make sure that the product or technology you are making is valuable. Because it takes time to do something, you have to predict the value of things in the future. If you make a wrong judgment, everyone will work hard for a long time, but after the work is completed, the results are mediocre, then the team morale will be low, and various problems will ensue.

When you think about the future value of something, the time should not be too short or too long. Three years is more appropriate. Assuming you want to continue in the current direction, you need to consider whether the user base and usage time you focus on will change in three years. More is a good thing. If it does not change, you can't do too much. But if it will become less, you have to consider changing direction. You also need to be wary of the emergence of new technologies. It is very likely that new technologies will completely overturn old technologies (deep learning, smartphones, electric vehicles) in just a few years. Analyzing those failed examples, the parties actually noticed the new technology very early, but underestimated its power. They only saw the shortcomings of the new technology compared to the existing mature technology, and then applied the development speed of mature technology to the new technology, underestimating the height that the new technology can reach in three years, and the degree to which users like new things and hate old things.

If you want to go in a new direction, you may no longer have the advantage of accumulated technology, that is, the value you bring to users may not be significantly different from others, or even lower. Then you need to find a good track (investor mantra), usually disruptive new technology, and the changes in products and users that follow. Only in a fast-changing track can new entrants easily overtake by more accurately predicting future value. In other words, heroes emerge in troubled times.

Good vision requires long-term training. You need to constantly think deeply and gain your own unique perspectives, rather than relying on the opinions of your friends. So you need to put down your work from time to time and give yourself time to think deeply. For example, I will walk for a few hours on the Bay Trail near my home from time to time, thinking while walking.

The core of management is to treat people with sincerity

If you have a clear team goal and a high-quality team, efficient execution is a natural thing. Therefore, managers have three core things to do: recruit people, retain strong members, and help those who are lagging behind. The best way to recruit people is to recruit people who are better than yourself. Another way is to recruit people who are better than half of the people at the same level each time, so as to ensure that the quality of the team can be continuously improved when the team expands.

Members with outstanding abilities will be welcomed everywhere. One of your tasks is to keep them in the team as long as possible (although they will eventually leave). One way is to put yourself in their position and imagine how you want your leader to treat you. For example, what I hope most is to continue to do things with greater value (sense of accomplishment) and learn new things from them (personal improvement). My boss can support me when I am in trouble (which happens often). Everything else can be converted into current treatment, such as how much time you can spend doing things you don’t like (disagree with the value of something, but fail to convince others not to do it), the traffic on the way to and from get off work, and the lack of lunch in the cafeteria. So the general direction is to create a relaxed environment, establish new and valuable projects every year, and try to get the best treatment for everyone.

For team members with unsatisfactory performance, you need to frequently point out problems and give suggestions. If there is no improvement after a period of time, you need to discuss whether the current project is not suitable. If there is still no progress, you can only help them change groups or ask them to leave. Similarly, you need to put yourself in the other person's position and understand what kind of help and respect they want. Most of the time, it's not that they are not good enough, but that you two are not suitable. A happy breakup can help former members find more suitable positions faster (thus avoiding them sending you razor blades).

Focus! Focus!

Some people say that a common cause of death for startups is blind expansion after achieving certain results. This is true everywhere. Whether you are alone, leading a team, or leading a company, resources are always limited. Concentrating resources on the most valuable things can ensure success. For example, Apple has hundreds of thousands of people, but it is very cautious about expanding its product lines. This ensures that each product has enough resources to subvert the market. In the early stages, you may be able to cast a wide net to catch more fish, but once the value of things becomes clear, we need to gradually concentrate resources. Because it is better to do only one thing to the extreme and make it the best on the market than to do several similar things at the same time. In this way, you can always get a positive value difference. One first is worse than ten second, and the third usually doesn't live long. The same principle can be applied to life, socializing, and learning.

As long as you put in the effort, your weaknesses can become your strengths

Every time a new feature of MXNet was released, my Zhihu classmates would complain: Go back and write the documentation first. Everyone knows that programmers don't like to write documentation. I have been hovering around the passing line in Chinese and English since I was a child, and I was even more resistant. In 2017, I made up my mind to write documentation. I spent all the time I had left for technology on it. Finally, I wrote the textbook "Hands-on Deep Learning" with everyone, which is now used as a textbook by nearly 200 universities around the world. So, your shortcomings can become your opportunities. As long as you face it, make up your mind and spend effort, and keep improving, your shortcomings will become advantages.

Strengths and weaknesses

All the gifts of fate are secretly marked with a price. When I spent all my energy on the documentation, I ignored MXNet itself. Failure to organize and invest a lot of resources to continuously improve its performance and ease of use resulted in it not being able to make it to the top two. In terms of value, "Hands-on Deep Learning" and MXNet are similar in terms of the number of users and user value, but users spend more time using the deep learning framework than reading textbooks, so MXNet is more valuable. It is hard to say that it is worthwhile to spend a lot of effort to win a battle in a field you are not good at, but lose a more valuable one in the field of advantage. Therefore, when it comes to leveraging strengths and making up for weaknesses, we must judge based on value, not face.

Communication overhead is the biggest cost in distributed systems

When the computing power of a machine is not enough, we use multiple machines to work together to complete the task. Although distributed systems are the infrastructure of many Internet companies, it is still difficult to improve performance. Because the actual computing power of each machine is different, and it will go on strike from time to time. And when one machine completes its own small task, it often needs to wait for the results of other machines' tasks, resulting in frequent data communication and waiting. So everyone knows that the key to improving performance is to reduce communication overhead.

When you need a large team to work together on something, communication overhead is also a big part, and it is more difficult to optimize than a distributed system, because people are more diverse and unstable than machines. People have different abilities and work efficiencies. Different divisions of labor among people lead to different ways of doing things, and even different optimization goals. Most people only care about their own affairs and don't want to or are unwilling to worry about other people's affairs. If you can't effectively bring everyone together, it will be a mess and nothing big can be accomplished. If you have just entered the workplace, the most important point you need to realize is that you need to reserve enough time and energy to communicate. Don't complain that this is a system problem in your company, this is an inherent phenomenon when a large team is working.

Promotion

I was promoted to a relatively high position because of my good luck, so I had the opportunity to participate in the company's promotion evaluation from senior engineer, scientist to director. Although different companies, positions and levels bring differences, in general, whether a person can be successfully promoted depends on whether the value of her largest project to the company meets the requirements of this position. There are three key points here: First, the value of the project to the company. This means that the target group and the value difference are what the company cares about, not you or the society. The value here is usually how much money you make for the company, or how much money you may make in 3-5 years. Second, it depends on whether your largest project is "high-end". If you have accumulated many projects, it may not work to get promoted by looking at hard work instead of merit. Third, your contribution to the project, such as how much you are responsible for, whether you contribute code, team coordination, publicity, plan making, or apply for resources. A common misunderstanding is that working with others will reduce my contribution. If you and your collaborators do not cooperate well, resulting in 1+1 being much less than 2, then your contribution is indeed reduced. But if you increase the value of the project through cooperation, then your contribution will not be less. Especially if the project value is taken to the next level, that would be even better.

There is an often overlooked "unspoken rule" for promotion, which is influence. As your position rises, the company's expectations of your influence also increase. From being able to influence a small team, including formulating technical routes, helping team members get started, answering questions, and even helping others complete their work, to influencing the next group (manager), influencing the next department (senior manager), influencing the next group (director), and finally influencing the entire company strategy (vice president).

Unless you are a natural leader, you have to work hard to cultivate your influence. In simple terms, it means actively helping others in addition to taking care of your own affairs. When others trust you, consult you, and are willing to cooperate with you, then you have influence over them. You may feel that helping others will delay your own work. But from the company's perspective, this kind of dedication needs to be encouraged and rewarded. In addition, the trust you win will bring you reputation and connections, which will be very useful in the long run.

Salary increase

Salaries in large companies are relatively transparent, and the corresponding salaries for each level can usually be found online. Salaries within a level fluctuate, and generally have a minimum and maximum. 80% are near the median, and 10% at each end, respectively, are for people who have just been promoted to this level and are about to reach the next level. You can simply think that as your ability improves, your salary will gradually jump from the minimum to the maximum, and then be promoted to the range corresponding to the next level.

The median salary of different levels is usually a geometric sequence, not an arithmetic progression. For example, someone one level higher than you may make half as much as you, but someone three levels higher than you makes 238% more than you, not 150%. In this model, you need to optimize the height you can reach in five years or ten years. So when comparing offers, you should not care too much about the numerical difference between them, but care about your future development after you go there (you say that if I take more money to buy coins now, maybe I will be financially free soon, that is also an idea).

Summary: Focus on the most valuable things

If I were to condense my insights from the past five years into one sentence, it would be a simple one: focus on the most valuable things. First, you need to have a clear understanding of value. Then, for a thing, it is not only necessary to understand the current value, but more importantly, to predict the future value. Secondly, when you have a grasp of the future through continuous rapid trial and error, you need to gradually focus the resources you can mobilize on the most valuable thing and do your best.

If you can do a few things of great value in your life, then it will be worth it.

When writing this article, I was shocked to hear that Professor Yuan Longping had passed away. I would like to use this article to commemorate his great life: focusing on hybrid rice, creating one of the greatest values ​​in human history. He is a role model for our generation.

"Reflections on Five Years of Work" Address:

https://zhuanlan.zhihu.com/p/374777591

Address of "The Five Years of Doctoral Degree":

https://zhuanlan.zhihu.com/p/25099638