I, an AI engineer born in 1995, created an "AI virtual teacher"

This article is authorized to be reproduced from Ranciyuan (ID: chaintruth);

Author: Lu Jingzhi and Ma Shuye

Editor: Cao Yang

AI painting is popular.

Previously, Douyin, Weibo, and Xiaohongshu were all flooded with AI paintings, and the most common comments about AI paintings were that they were funny and outrageous. "The photo of me holding my dog ​​was painted by AI into a monster with a humanoid fox face." Ayu, who is addicted to AI painting, recalled to Ranjiyuan. At that time, Ayu thought it was very funny and recommended the mini program to his friends. I made a WeChat avatar using a monster.

Douyin is also full of "outrageous" AI paintings shared by users. Some painted a foot as a cartoon girl wearing a pink skirt, and others painted a bicycle as a cartoon boy with "weapons." .

As AI painting becomes increasingly popular, the keyword Artificial Intelligence (AI) has once again entered everyone’s field of vision.

In fact, since the AI ​​field became popular in 2018, questions such as "Which industry is the most popular this year" and "Which majors are the best to find a job" have appeared on Weibo, Zhihu, and Xiaohong every year during the graduation season. Hot searches on books and other platforms. And AI is elected as the most popular employment major almost every year.

AI is increasingly being paid attention to behind the scenes, or with the development of technology, there are more and more application fields, including network security, medical diagnostic assistants, vision and imaging, advanced manufacturing, data mining and marketing, etc. According to IT Juzi statistics, in recent years, industries where AI technology is widely used, such as medical health, advanced manufacturing, and intelligent hardware, have also been highly sought after by the primary market.

Screenshot of Ran Diyuan

Many relevant practitioners who have extensive research on AI told Ranciyuan that after years of observation and personal experience, AI entered the peak period of the first wave of capital and talent influx probably in 2016. At that time, companies were recruiting algorithm engineers and education lecturers with salaries that far exceeded the average salary in the industry. This also directly caused some start-up companies with small cash flows to be unable to support them.

With the wider application of AI, the AI ​​industry is becoming more and more "popular", and the "involution" of the industry has become more obvious, and it has become more and more intense as capital enthusiasm retreats and technology continues to improve.

Li Hao, who works in the Game Industry Department of NetEase Intelligent Enterprise, has already been engaged in AI industry-related work as early as 2018. Li Hao, who is now an AI robot algorithm engineer, said that the year he just graduated, fresh graduates only need to master computer languages. With knowledge and relevant work experience, you can basically be selected by big companies. "But in recent years, especially last year and this year, almost all fresh graduates without very vertical project experience were eliminated in the preliminary screening stage."

In this regard, Ruoxuan, who also works at NetEase, said that the AI ​​industry requires practitioners to have strict self-discipline and continuous self-learning, and studying papers has long become a daily routine at work and even during rest.

Lu Wei, who is engaged in computer vision-related work at Baidu, even bluntly said that in addition to the increasing demand for work skills, his department is more like Party B within the company, and daily communication and friction with project managers are inevitable.

The ever-increasing competition for talent, unstoppable self-learning, and friction with colleagues in other departments... What is the daily life of practitioners in the glamorous AI industry like? How do they evaluate the changes in the industry?

"After the emergence of AI education, 'replacement' was the subconscious reaction of teachers." Gu Qian, who entered AI-related industries a year earlier than Lu Wei, told Ranciyuan. In 2017, Gu Qian broke into the AI ​​education track. As a co-founder, Gu Qian, born in 1995, led the team to develop an AI virtual teacher that has caused controversy. Behind this controversy is the possibility of "AI replacing real teachers."

Regarding this, Gu Qian was a little bit amused, but she also strengthened her confidence in AI entrepreneurship. As a technology enthusiast who dreams of "changing the world with AI technology," it was the powerful logical reasoning and computing capabilities of AI that fascinated Gu Qian, and she resolutely transitioned from a media person to the AI ​​industry. In fact, the seeds of my fascination with AI were quietly planted while I was studying econometrics at Guqian University.

During college, the tutor who guided Gu Qian's studies was researching AI. Based on this, through her mentor Gu Qian, she first learned about basic concepts such as metacognition and metalanguage, and became familiar with new things such as Bitcoin. Through this, she also met like-minded entrepreneurial partners.

After graduating from university, Gu Qian had one year of experience in the media industry. In 2017, under the craze of "entrepreneurship", she firmly believed that "AI will become the trend of the future", and she hit it off with her partners and became a member of the AI ​​entrepreneurial army.

"AI is still too far away from our daily lives." For Gu Qian, AI is very suitable for application in the learning process of highly logical subjects such as calculus and mathematics. "If AI can be implemented in the educational scenarios of these subjects, perhaps it can drive education itself through technology and take another step forward."

Soon, Gu Qian and her partners formed a team of more than a dozen people and set their sights on the more segmented "AI mathematics education." Except for Gu Qian, the team is full of technicians who specialize in algorithms. In order to promote the product to be accepted by the market faster, Gu Qian always "mixed" in the algorithm team to "catch up" with knowledge, and gradually became familiar with various "jargons" of mastery.

Subsequently, Gu Qian actively cooperated with the school, and the "AI virtual teacher" she promoted was gradually accepted by the school as a "teaching assistant".

In 2020, affected by the epidemic, the online education industry ushered in a capital boom, and Gu Qian also reaped the dividends. "At that time, almost all the leading online education companies on the market had contact with my team, and many of them reached cooperation." Gu Qian said that it was also at that time that the popularity of online education made AI education become a One of the hottest "concept stocks".

According to data from China Research Network, 2020 will be an explosive period for the number of company registrations in the AI ​​teaching industry. There were 172,000 new registered companies throughout the year, a year-on-year increase of 292.8%. In addition to online education giants New Oriental, Xueersi, Yuanfudao and other institutions, major Internet companies such as Bytedance, Tencent, Baidu, and NetEase are also vying to enter the game, such as Bytedance’s Guagualong series and Tencent’s Happy Mouse English. "The reason behind this is that there are too few teachers, and there are even fewer teachers who can be used for personalized teaching." Gu Qian said.

In addition, on the one hand, a teacher in online education can be responsible for up to 200 students. In school, a real teacher can only be responsible for about 50 students at most. "For companies and schools, the introduction of AI virtual teachers can better control labor costs."

On the other hand, considering the time and opportunity costs of real teachers, AI virtual teachers can share some of the repetitive work, freeing up real teachers to do more targeted and difficult work. On the other hand, AI virtual teachers can perform intelligent analysis based on students' words, better sense students' knowledge mastery, and then provide improvement plans.

"Behind this, we need to rely on powerful algorithms and huge amounts of data input. This is also one of the ways that AI technology can improve the quality of education." Gu Qian explained. In fact, compared to the abstract concept of "AI virtual teacher", Gu Qian prefers to call it an "AlphaGo" that can be used for education. Its main function is to allow the AI ​​system to respond to every sentence of the students through interactive dialogue. Process the information behind the words.

However, although AI is highly sought after in the capital market, not all AI practitioners can live a prosperous life. On the one hand, the wave of employment brought about by hot capital money has led to a mixed bag of entrants. Gu Qian told Ranciyuan that some teams only focus on conceptual marketing, under the banner of AI, but actually record and broadcast courses given by real teachers. As a result, the content and format of the courses are the same, and consumers gradually lose confidence and motivation to choose.

On the other side, there is the crazy involution of the industry caused by companies trying to dig out truly outstanding talents. In October 2020, ByteDance’s education business Dali Education announced its independence and provided an annual education business budget of 10 billion yuan. Large companies that were not short of money simply left the industry and poached people with wages higher than the industry standard, causing the average wage in the entire industry to continue to rise. According to Gu Qian, at that time, the monthly salary of teachers of AI courses in first-tier cities even soared to six figures.

Along with generous salaries, some companies have also launched price wars. The ubiquitous "9.9 yuan limited time courses" have made some start-up companies unable to bear the huge imbalance between revenue and costs, and have announced their withdrawal. Gu Qian, who had already fallen into "involution" at that time, chose to "lay down". Gu Qian said that instead of spending a lot of money on advertising, it would be better to follow the technical route honestly. "I always maintain the staff of our company's technical team." The proportion is no less than 30%. To me, technology is our core competitiveness.”

With the implementation of the double reduction policy, while online education pressed the pause button, the trend of AI education seemed to have quietly stopped. Last year, Gu Qian adjusted the company's strategy and chose to go overseas to North America. "The U.S. market will undoubtedly provide sufficient data for AI companies, allowing us to calm down and continue to polish our products."

Today, Gu Qian and her team are making overseas arrangements step by step. Regarding the future, she said optimistically, "Many people ask whether AI is a bubble, but to me, it is also good to let the bubble hang in the middle and attract more talents to explore together."

Ruoxuan also had the same idea as Gu Qian. In the increasingly "volume" AI industry, Ruoxuan also chose to "go up". "My husband and I are both algorithm engineers, so our biggest interest after work is to discuss the latest papers." Ruoxuan, currently a computer vision algorithm engineer at NetEase, told Ran Ciyuan half-jokingly.

In 2011, Ruoxuan entered the School of Electronic Information and Electrical Engineering of Shanghai Jiao Tong University (hereinafter referred to as the "Electronic School"). At that time, Ruoxuan had not determined her future academic and career direction.

"When I was a freshman, the School of Electrical Engineering was more of a broad subject, not very subdivided. At that time, we studied some basic courses in general electricity. When I was a sophomore, after more detailed majors were divided, I chose Later, because of my own passion for computers, as well as the popular environment of computer science and some of my mentor’s research projects, I gradually transitioned to the AI ​​industry,” Ruoxuan said.

Although before graduation, Ruoxuan was involved in the automatic analysis and diagnosis of medical images by AI algorithms, the main application scenario was that after inputting medical CT images, the algorithm would provide an automatic diagnosis and treatment opinion through calculation. However, this is a big leap from the application of "network content risk control" that Ruoxuan has been engaged in since graduation.

Ruoxuan explained that the amount of data sets in the medical field is relatively small. Due to problems such as medical instrument collection, the data set for a certain disease will not be very large and distributed relatively evenly. But in comparison, there are thousands of forms of unsafe content on the Internet.

In addition, online content risk control will cause confrontation at any time. "There are constantly black and gray industries who want to take advantage of the loopholes and launch attacks on you. In the field of medical diagnosis and treatment, we will not encounter attacks." Ruoxuan said that more importantly, AI technology can only play an auxiliary role in medical diagnosis, and most doctors still rely on their own professional judgment. However, online content risk control is more dependent on technology, so the technical requirements for staff are also higher.

Ruoxuan also felt the upper limit of her abilities due to changes in work application scenarios. "I did not major in computer science at university, so my knowledge of computer languages ​​and computer hardware has always been my shortcoming. In addition, the algorithm scenarios I studied are also quite different from the actual application scenarios, which requires me to continue to learn. ." Ruoxuan analyzed herself.

Ruoxuan recalled that the most "crazy" self-learning stage was a task she received half a year ago. "At that time, in order to control labor costs more effectively, I started to learn 'semi-supervised' and 'self-supervised' algorithms."

"Before, there was a joke in our industry, 'How much artificial intelligence can be achieved' because a large amount of data needs to be labeled to allow the machine to learn better, and this process requires a lot of manual work. Later, there was It adopts the so-called 'semi-supervised' form, that is, there is no need to manually label each set of data, and some data do not even need to be labeled. This allows the machine learning task to be better realized while controlling labor costs. Ruoxuan explained that this was also an opportunity for her to come into contact with semi-supervision.

"I followed many big Vs in the industry on Weibo, WeChat and other platforms. I continued to extract keywords from the articles they published, and then searched them in top academic journals. I first learned the theory of the paper and then put it into practice at work. Go." Ruoxuan recalled, "Every algorithm engineer will basically have 2-3 tasks at the same time. You have to conduct theoretical exploration while completing the task. Therefore, my time schedule at that time was to complete the task during the day. I started reading the paper during the rest and evening hours.”

“I was a bit obsessed at the time.” Ruoxuan recalled with a smile that she would discuss the latest papers she saw with her husband almost every night. “Just like an artistic creator looking for inspiration, we were researching and solving problems. This kind of inspiration is also needed when planning.”

Later, Ruoxuan gradually applied the "semi-supervised" method to more of her projects and demonstrated it to the entire group. When asked when this stage of self-learning will stop, Ruoxuan said, "It's hard for me to say that this stage will stop. In fact, AI is an industry with a huge amount of information and rapid evolution, and practitioners need to constantly learn on their own. , exposure to the latest solutions and algorithm trends to get things done.”

Of course, in the process of self-learning and improvement, Ruoxuan also felt the changes in the entire AI industry.

“Although I have only worked in this industry for 3-4 years, I have already felt that industry recruitment is becoming more and more strict.” Ruoxuan said, “When I graduated as an undergraduate, fresh graduates basically only needed to understand some basic computer language theory. I participated in some projects during school. Even if I just understand the open source code on the Internet and actually run it and debug it, I can basically get a good offer from a major Internet company. But now, fresh graduates want to enter a major company. To publish a top-notch paper, even to participate in projects, the job application scenario must be very vertical.”

As for the future development of AI, Ruo Xuan said, "AI sounds grand, but if you 'take off' the 'outer coat', it is just the inevitable product of human technological progress to a certain stage. How technology evolves depends on how people use and restrict its development."

Compared with Ruoxuan who gradually transitioned from the Electrical Engineering Institute to the AI ​​industry, Lu Wei, who now works at Baidu as a senior engineer and is a core member of the Fuchun Mountain Residence Project, is an out-and-out computer major. However, he initially chose the computer major, but It is also a "accidental hit".

Lu Wei said frankly that he entered the university campus in 2011, when the computer major had cooled down from the crazy craze in 2009. “At that time, the admission scores for computer majors in many schools were not high. I only regarded the computer major as an alternative. , I didn’t expect to win.” What’s more coincidental is that as time goes by, the computer major has ushered in a “second spring.” Between 2016 and 2018, it gradually became one of the hottest industries, and at that time, Lu Wei was graduating from graduate school.

Lu Wei took advantage of the situation and joined the computer vision industry. His main work content is to edit videos and images according to the needs of different project groups and finally generate and present the visuals. "Among them, style transfer is one of the directions I am good at." Lu Wei further explained, "Style transfer refers to combining a classic painting and a modern photo or image through AI technology."

Like Ruoxuan, Lu Wei also needs to constantly learn on his own at work.

"Reading the latest papers in top journals is one of my daily tasks. Because the requirements I receive are rarely one model that can solve all problems from end to end after being applied. Therefore, I need to dismantle these requirements myself, Then it is transformed into an algorithm problem and solved step by step. In addition, everyone has different tasks, so it is difficult to find others to ask for advice. More often, I need to disassemble the problem myself and find the answer from the paper. "

But unlike Ruoxuan, who can solve most problems through learning, many of Lu Wei's needs require communication with other colleagues in the project team. "My job is somewhat similar to that of the visual design department. They both belong to Party B within the company. We receive tasks from the product manager, which is the role of Party A within the company. After dismantling and completing the tasks, we finally deliver them to the project manager for review. And modification. However, there is less space for art and creation, and the essence of solving problems through technology is greater," Lu Wei said.

When talking about the communication with the product manager (Party A), Lu Wei said with a smile, "There is a problem of reasonable quantification in all the communication between technical personnel and Party A." He further explained, "Especially for visual quantification, it is more subtle. Some product managers cannot accurately describe their expectations for vision. Sometimes they only have some rough imaginations. The most important part of my task is to quantify these imaginations."

“Our solution to this is to collect some data sets and let product managers rate them. This can better help the team understand the project manager’s preferences more intuitively.”

Of course, even so, Lu Wei will still have ideological conflicts with the project manager during the work process. At this time, Lu Wei is more inclined to adjust according to the other party's requirements than to stick to his own ideas. "We need to recognize that the essence of our work is not secondary creation of art, but a full understanding of needs and problem solving." Lu Wei So said.

Lu Wei, who officially started working in 2018, has the same feelings as Ruoxuan about the increasingly strict recruitment evolution in the AI ​​industry. “The first is whether the journals in which fresh graduates publish papers are top-notch; secondly, the companies’ response to internships and on-campus participation projects and work projects for fresh graduates. There are also stricter requirements for closeness.”

At the same time, Lu Wei also observed that as the AI ​​industry has become more popular in recent years, many science and engineering students have made the transition. Regarding the joining of these transforming young people, Lu Wei said, "Basic programming is the basic technology for engaging in AI-related work. Therefore, some people who study automation and communication majors will have computer theory courses themselves, and the transition to AI will be easier. Relatively speaking, "It is relatively more difficult to transform traditional engineering disciplines such as materials and chemical engineering that have little to do with computers."

In Lu Wei's view, whether one is from a major or is transitioning across industries, continuous learning and maintaining a passion for knowledge are indispensable qualities in this industry.

*Luwei is a pseudonym in the article.

This article is reprinted with permission from Ignition Dimension (ID: chaintruth)

Latest articlesabout