I’ve been shopping around Zhihu for a long time, and I’ve read all kinds of doctoral questions. Some say that a doctor’s job is not worrying about getting a job, and some say that studying for a doctorate is a waste of time.
The current social situation is very serious, especially age anxiety. Most doctors are no longer young after graduation. Most units with establishments seem to have a doctoral age requirement of 40 years old or 45 years old? Can not remember.
Some time ago, I had a brief chat with a friend who was studying for a Ph. D. He said that it is very difficult to get a job after the age of 35.
How is society now? ? What about people over 35? There are also quite a few seniors who are studying in graduate school. Don’t they have financial concerns? Is it true that unmarried and childless economic conditions are not suitable for "advanced" graduate school?
I believe that everyone has the above similar problems. Today I saw a post shared by Dr. Li Bojie on Zhihu. It is worth reading and recommending to everyone~
Link: https://www.zhihu.com/question/559157484/answer/3149816238
Author: Li Bojie
Editor: Deep Learning Natural Language Processing Official AccountEnter the NLP group —> join the NLP exchange group
I am a Ph.D. student from the University of Science and Technology of China. When I graduated, I basically got top offers from major manufacturers, including Huawei Genius Boys, Alibaba Star, etc. In the end, I chose Huawei. The following is a sharing of my job search experience.
Graduated with a Ph.D., academia or industry, entrepreneurship or a large factory?
When I graduate with a Ph.D., it involves the choice of work. I roughly divide it into 5 types: academia, freelance, entrepreneurship, large foreign companies, and large domestic companies.
The first is academia. Although I like scientific research very much, just like my concerns when I started my Ph.D., I don't like theory and like practical research. I feel that research in academia, even in the field of network systems, is often divorced from actual needs. For example, during my doctoral period, the whole team spent several years researching hard work. Although they published good papers, they only explored some possibilities and have not been commercialized in any production system. Today, I still hope to summarize the successful practices in the industry into papers and publish them, just like I like to write a blog to share my life, but I don't want to "speak for new words", and do some short-term work for writing papers. Things that cannot be used commercially.
The second is freelance work. Although the hacker spirit encourages freelance work in essence, the current knowledge-based benefit sharing mechanism is not perfect. Except for a few big Vs, they only rely on consulting services, knowledge payments, and personal website advertising revenue to support themselves. Very difficult. For most freelancers, freelancing can make far less money than working for a big factory. As a result, my idea of freelancing gradually faded. Envious of LUG's friends who choose freelance work, they use their lives to practice the spirit of idealism. For now, we still need to ensure that we do not starve to death before talking about the hacker spirit of freedom, sharing, and openness.
Then there is entrepreneurship. The advantage of start-up companies is that they are efficient and can drill into market segments like catfish. Analyzing the size of the target market and the current player situation, you will find that it is not easy for the company to make money. For a company of Microsoft's size, a market worth less than $1 billion cannot be seen at all. Because a large company needs a complex process to set up a project and needs a multi-functional team support, a small team of three or five people can get things done, and a large company may need a team of dozens of people. The development process of a large company is complicated, and the code productivity is 300-500 lines/person-month, and the output code is also strictly inspected and tested; the code productivity of a small company may be as high as 3000-5000 lines/person-month, but the code quality may not be so high. Therefore, large companies are more suitable for the to B market with strict quality requirements, while startup companies are more suitable for the market with small steps and fast iterations.
It is difficult to start a business in the field of data center network research where I work, because data centers are only owned by large companies. Large companies either develop their own smart network cards or purchase smart network cards from other large companies. Who would buy a little-known company? What about the network card of a small company? Therefore, although there are many start-up companies in this typical to B field, they are all founded by successful and famous manufacturers, which is not suitable for young people to start a business. If you want to start a business, you may have to do relatively "emerging" fields such as blockchain, metaverse, privacy computing, artificial intelligence, etc., but I have no accumulation in these fields, and I don't want to give up the existing accumulation in the data center field .
I think the field of data centers is more suitable for working in large factories. Big companies have more rules than freelancers or startups. These rules and regulations are reflected in the rules and regulations, project management and development process on the one hand, and in the selection of topics and projects on the other hand. In my mentor's words, it is to follow the company's commercial facts and choices, respect the company's position in the industry and the status quo of its products; research problems must have a way to be implemented in products, rather than solving "human problems" ".
Working in China or abroad is no different for me, but my girlfriend just graduated with a Ph.D. this year. If I go abroad in 19 years, it will change from a different place to a foreign country. When we are in a different place, we can meet once a month, but if we are in a different country, we can’t meet so frequently, and there will even be jet lag, which is even more troublesome. So I quickly ruled out the option of working abroad. At the beginning of 2018, on a whim, I signed up for the naked TOEFL test. I also scored 103 points in the test. Among them, speaking and writing were the worst, with 22 and 24 points respectively. After some preparation, I may be able to improve. But then I didn't plan to go abroad to work, so I didn't study English deliberately.
Dachang interviews are mainly about doing questions, talking about research results and chatting
From such a comprehensive look, only large domestic companies or unicorns are left. There are generally several channels for large companies to recruit people, such as invitations from big bosses, internal referrals, or overseas resumes. Because we are in the circle of MSRA, many big shots from MSRA naturally become our important choices, and they will also take the initiative to contact us. This way of inviting big shots has the highest probability of getting top-level Offers; internal referrals are to find Recommended internally by my friends, my friend Cong, who is a doctor of joint training, helped me recommend several companies; Haitou’s resume is generally not recommended. Departments are not necessarily appropriate either. The method of invitation or referral by the boss is equivalent to using the endorsement of the recommender to give the interviewer a better first impression before the interview. However, you must find a reliable friend for internal referrals. Someone once recommended a candidate to me and said bad things.
I interviewed a total of 12 companies, divided into three categories:
Major domestic manufacturers, including Ali, Tencent, ByteDance, Meituan, Huawei
Domestic unicorns, including Pony.ai, 4Paradigm, Cambrian, Horizon
Foreign companies, including MSRA, Xilinx, VMWare
The interview format of most companies is similar. The first two rounds are technical. Most companies will talk about the research they have done and do some algorithm questions. Behind it is the director's face, the more senior president's face, and some HR faces. These high-level interviews are mainly based on chatting, mainly to see the influence of research results, future career planning, and the degree of matching of company values. Senior leaders will also take this opportunity to promote the company and attract candidates. Some senior leaders will also ask to write code during the interview.
For example, Alibaba Star includes two rounds of technical interviews, one round of supervisor interviews, one round of cross interviews, and one round of Alibaba Star final interviews. At the end of Alixing, the P11 president of the department, two big bosses from other departments, and the senior HR director interviewed together. First, let them give a 30-minute academic report, and the big bosses will ask questions for another 30 minutes. At that time, I talked about the database project that integrated batch and stream processing that I was doing. I didn't realize that the P11 boss was a senior expert in the database, and I was sprayed. However, many of my immature projects use the opportunity of reporting to the boss to receive feedback in this way. When I was an intern at MSRA, once (not an interview) Turing Award winner Butler Lampson came to visit MSRA, I told him about the idea of total order message transmission, and he gave very valuable feedback. This job was rejected Finally published in SIGCOMM '21 after reviewing it 4 times.
I was also impressed by Ali's cross interview. The interviewer for the cross-interview is from another department. He is an expert who jumped from Intel. He asked me a lot about the CPU architecture. He said that I know a lot among my peers, but I still know little about the details. . I said that it is difficult to know the details of CPU microarchitecture in public channels. He said rather profoundly, this is the value of work experience. This remark made me realize that I must join a team with core technology and have as broad a technical vision as possible in order to learn as much know-how as possible in my short youth.
During the interview with Mr. Guo Chuanxiong of ByteDance, Dr. Guo asked me a math problem. That question is given a recursive formula to find the upper and lower bounds, which is derived from his DCell paper (the number of servers increases exponentially with the DCell level), and I have not read this paper seriously. In fact, the upper and lower bounds in mathematics are not difficult to find, but I did not make it at the time. Dr. Guo told me that it suddenly became clear. Dr. Guo said that the research he likes is the type with beautiful mathematical theory and practical value at the same time, but such research is rare. During the interview with Byte CTO Yang Zhenyuan, he asked me to write the code on the spot and talked a lot about the details of my research work. He said that it has been a long time since he talked about such low-level things. He did this kind of low-level optimization when he was in Baidu.
During the interview with Pony.ai, I met the legendary Lou Jiaozhu as I wished. Before I met Lou Jiaozhu, there were two rounds of code interviews. Each interview had two or three not-so-difficult algorithm questions, which was very similar to Google's interviews. Lou Jiaozhu gave me several puzzles, which I have never seen before:
Use a coin with the probability p of heads every time it is thrown, and allow multiple throws to realize an equal probability 01 random variable, requiring the expectation of reducing the number of throws as much as possible. This is indeed a very beautiful question. I figured out the optimal strategy on the spot but failed to calculate the expectation. Later I found that the expectation is 1/H(p), where H is the entropy function.
The number of vertices and faces of the n-dimensional cube.
Given any starting point and end point on the six faces of the three-dimensional cube, find the shortest distance for the bug to crawl along the face.
Divide one side of a rectangular paper into thirds and fifths. The leader of the building gave me a piece of A4 paper to fold. I struggled for a long time and gave a method. The leader thought it was right, but I found it wrong; then I changed it and found a third-class paper that he had never seen before. The division method is relatively complicated, but the leader's method is very simple and can be easily extended to any equal division.
At that time, I asked Mr. Lou what he thought of the end-to-end automatic driving pipeline. He said that he was worried that the life-critical things were completely handed over to the black box of the neural network; the interpretability and debuggability of deep learning were not strong. Even if the effect is good, it will be difficult to explain to the public in case something goes wrong, and it may not necessarily meet the requirements of regulations.
MSRA may have the largest number of interview rounds, including two rounds of code interviews, collective interviews with researchers and supervisors, cross interviews, deputy dean interviews, HR interviews, etc. Each round of code interviews consists of two algorithm questions plus one system design question. It is called "one-vote veto system", because those who fail to pass the code cannot be hired no matter what the supervisor says. MSRA's group interviews with researchers and supervisors are similar to Huawei's collective interviews with researchers and supervisors, and Alistar's final interview. They all talk about academic reports first, followed by questions and exchanges. The cross-section is mainly to examine the breadth of knowledge and the openness of thinking. Once a candidate attacked the research field of the cross-interviewer during the interview, which is not appropriate.
During the interview with Huawei, Tan Bo asked me what I thought of the application of FPGA in the data center. If you just talk about the advantages of FPGA from the perspective of defending your own research results, this answer will be narrow. Different companies have different choices, including reasons for technical path dependence and commercial considerations. Companies prefer candidates who can objectively analyze the pros and cons of different SmartNIC architectures. When I was an interviewer myself, I also hope that candidates can jump out of their own research work and have new independent thinking. Say it again.
In fact, I didn't prepare much for these interviews, because I think the algorithm questions of Dachang are not difficult (unless it is a question about mathematics or brain teasers); the research results have been explained many times, and there is no need for PPT materials. You can speak out; you don’t need to prepare for chatting. My career plan is to become a system architect. I also like to struggle, which matches the values of major domestic companies. At the same time, I also maintain a strong interest in computers and hope to continue to explore Frontiers of Pasteur's Quadrant. Some interviewers will still ask some questions related to basic computer knowledge, I just know what I know, and I don’t know what I don’t know. For example, I have never used C++ and Java, so I said no; although I have done a little kernel development, they are all the simplest kernel modules, and I have not studied the network protocol stack. When I am an interviewer myself, I also like to examine the candidate's basic computer knowledge, such as operating system, network, distributed system, database, etc., as well as the syntax and semantics of the candidate's usual programming language and the knowledge of compiling and linking runtime (such as garbage collection). ), it is often found that many candidates answer well when they ask theoretical questions, but they can't answer the first level of questioning, because they have not really used these systems, but just memorized some face-to-face scriptures. Some candidates even try to deceive the interviewer, which is easy to spot if the interviewer is an expert in the field.
It wasn't until today that I became an interviewer myself that I realized that the closer the candidate's level and experience are to the interviewer, the more comfortable the interview will be, just like when I was a candidate back then. Meeting a master is like chatting with friends. For example, if you meet someone who is playing ACM, talk about which data structures and algorithms you are good at, which questions are tiankengs, and how to cooperate during competitions; , How many pitfalls are there in CSS and JS, and what is the history of blood and tears in the process of operation and maintenance; when you meet Zhihui Jun who tosses embedded systems like Mr. Zhihui, talk about which single-chip microcomputers you have played (burned out), and how to adjust the parameters of PID control; If you meet someone engaged in blockchain, talk about the consensus mechanism and smart contracts. By the way, ask him to give me the currency price that I haven’t seen for many years; The "glorious history" of which loopholes have been identified; even if you encounter a field that you don't understand, you can ask for advice with an open mind, and you can learn a lot from listening to the candidate tell his story.
Selection of Big Factory Offers
Thanks to the great love of the interviewers and leaders, I got good offers from these companies. Offers from major domestic manufacturers will match each other. For example, after I get an offer from Alistar, several other major manufacturers will refer to this package. Therefore, if juniors and juniors want to find a job in a big factory, they must not only apply for one, or they will be slaughtered.
In order to choose an offer, I wrote a Summary (Summary Summary) for each company, listed Strengths (advantages) and Weaknesses (disadvantages), and put the information collected during the interview process and various aspects into Comments to the Author (comments). At the same time, I will also record the big bosses, direct supervisors, new excellent employees and interns in the company. Just like the review comments for each paper in Dinghui are very substantial, I also wrote thousands of words of comments for each company. Based on this, I created a spreadsheet to rate each company on several dimensions, including:
entry salary
long-term expected salary
Offer's relative rank among peers
Business matching
growing space
Industry influence
Academic influence
Work-life balance
leadership familiarity
job stability
Company Culture
Company brand
sector outlook
field prospect
Technology accumulation
Daniel Boss
Daniu classmate
Work place/account
Each item is scored from 1 to 5 like a paper review, and then calculates statistical indicators such as the arithmetic mean, geometric mean, and variance. Based on these statistical indicators, the companies are then sorted. Just like reviewing papers, the failure of a paper is not just based on the score, but in the first round of screening, the low score can be filtered first. There are 7 companies with an average score of 3 or more, that is, 5 companies were filtered out in the first round. Because only one company can be selected in the end, the "passed draft rate" is only 8%, which is indeed a cruel screening. For the companies that were screened out in the first round, in order to avoid recruiting HR and wasting the sincerity of the supervisor, I will speak directly earlier. Just like telling the author early after the paper is rejected in the first round, the author can transfer to other conferences earlier. I am not good at rejection, and I have to think long and hard when rejecting every company.
The lowest average score of these seven companies is 3.78, and the highest is 4.11 points. Among them, the average score of the first six companies is at least 3.94 points, so it is really close to each other. I'm also having trouble weighting these metrics above. In the paper submission, the final acceptance is not determined based on the score, but after discussion at the PC (Program Committee) meeting, if there is a champion (support) and no one objects, the article will probably be accepted. The same is true in the context of choosing a company. If there are obvious advantages and no obvious disadvantages, it is a better choice.
In the end, because my doctoral mentor was in Huawei, and Huawei has the advantage of combining software and hardware design, I made a difficult choice. It is even more difficult to reject the other 6 of these 7 companies, because each company has broad prospects, favorable treatment, ample room for growth, and batches of great bosses and classmates. At the time, maybe it didn't make a huge difference which company to choose.
Not long before I graduated, I encountered the "516" US sanctions incident, and I was panicked for a while. Many people said that Huawei was about to die, and advised me not to go to Huawei. A teacher said to me that the most precious thing people have is time. If you want to support Huawei, you can spend money and buy a mobile phone to support Huawei. But if you choose to join a company and devote most of your time, it is different. Great scientists such as Qian Xuesen can give up their generous remuneration and return to China to do unknown work, but they can no longer do world-leading research like Yang Zhenning, win a Nobel Prize, and their living conditions will not be very good for the rest of their lives. However, after field research and interviews, I confirmed that Huawei is not as vulnerable as rumored outside. The entire company is still operating normally, and what projects to do and what projects to do. When eating in the cafeteria, the TV broadcasts and sanctions Relevant news, and everyone has become accustomed to it. Therefore, I tried my best to reject all opinions and insisted that joining Huawei and Qian Xuesen's return to China were not in the same nature at all, and I maintained my choice. It's also the major choice I've been opposed to the most in my life.
During the induction training for new employees, we had a discussion with a president of the human resources department, and I asked whether it is possible to follow the example of "Alistar" and implement a systematic project to attract outstanding talents. In August 2019, the company just launched the "Genius Boys" program, and I was lucky to be a part of it. Of course, the "Genius Boys" program at the company level has nothing to do with my random suggestions, but I can predict that the company will pay more and more attention to the recruitment of excellent talents such as Ph. Tighten the recruitment of excellent talents. In the past 3 years, under the care of the leaders and colleagues, I have taken on more and more important project responsibilities and played more and more value, which also shows that this choice back then was not wrong.
Although I only have very limited work experience, what I want to tell my juniors is that in the above "scoring items", the importance of leaders and departments is higher than the overall situation of the company and research field. Whether a person lives comfortably or not, and whether he can be valued at work depends largely on the leader, especially the direct supervisor. Consistent with the leader's vision (vision), cooperation will be smoother. It's like in the process of PhD reading, the importance of the mentor is far greater than the reputation of the school and college. Second, the importance of the department mainly depends on the business and atmosphere of the department, that is, whether the basic business is stable, whether there is room for imagination, and whether the team atmosphere is harmonious. In large companies, a drop in the company's stock price may affect revenue, but the short-term impact on individual work is not significant. Whether the research field is popular or not is a slow process of change. If one day your research field becomes old, it is not too late to change it. Many bigwigs have experienced switching research fields.
The above is the latest answer from the boss, the latest news: the boss has left his job to start a business
Enter the NLP group —> join the NLP exchange group