Alibaba’s thorny road to artificial intelligence: What does it feel like to sit on a mountain of gold and eat steamed buns?

In the past five years, Alibaba has quietly entered the first echelon in the fierce competition of global artificial intelligence, and a group of scientists behind it, who are known as the DAMO Academy, are becoming increasingly mysterious and attractive.

This is a group of ambitious scientists who personally broke through the comfort zone that others envied and stepped into the unknown land full of taboos, just to find a broader stage. They have devoted their lives to "seeking proof", either in academic theories or in the value of technology, and are willing to take the risks and embarrassing situations that come with it.

In the past five years, Alibaba has experienced the beginning and breakthrough of the long march of artificial intelligence technology, and they have also experienced a huge adventure and self-affirmation.

When you first meet them, you may be surprised that they keep talking about "paying the bill", "landing", "scaling" and other business-oriented words, without any academic air at all, as if they are entrepreneurs on the front line who are obsessed with...

Until you hear about those difficult yet fortunate times, everything becomes clear - the more you understand the truth, the more pragmatic you will become.

This is also the wonderful thing about life. What changed them was not their glorious academic path in their early years, nor their current material status, but the desperate "self-proof" after being questioned and misunderstood again and again. Different from the public reasoning and simulation verification in the ivory tower and in front of the test bench, "self-proof" in the business field requires real business indicators and actual results, just like the slogan that can be seen everywhere in Alibaba's office area - "NO DATA, NO BB".

The atmosphere of "self-proof" under high pressure has created a technological value system among scientists that is almost harshly realistic.

After the heavy work is unloaded, they regain their innocent hearts and immerse themselves in the fun of algorithm formulas and theoretical reasoning. But life is not always free and romantic. What is longer is the loneliness and torment of overcoming difficulties.

Through conversations and exchanges with nearly ten Alibaba artificial intelligence scientists, this article attempts to restore Alibaba's thorny road to building an artificial intelligence technology system, as well as the scientists' inner journey along the way.

1. Both difficult and lucky
Many years later, when encountering smog in Beijing, Qi Yuan might still think of that Spring Festival when he worked overtime - in the Alimama conference room on Xida Wang Road, a table of people were having a meeting wearing masks, the air purifiers at their workstations were whirring, and outside the window there was no view of the opposite building due to the smog.

Before returning to China, Qi Yuan had already made a series of adaptation lists in his mind in preparation for the various challenges after returning home, and the "fog crisis" was probably the most important one that was omitted. It not only came from the weather in Beijing, but also included that bold decision. 

Qi Yuan has walked a smooth path that countless science and engineering men dream of: he graduated with a Ph.D. from MIT at the age of 31, became a tenured professor at a top university at the age of 39, settled in the United States, owned a spacious and bright laboratory and a single-family villa, a beautiful wife and two lovely children, and two fixed long vacations a year were enough for him to strike a perfect balance between work and life.

Until a bold decision, Qi Yuan personally broke this balance. After rejecting offers from Google, Facebook, Baidu and other companies, Qi Yuan decided to return to China to work in Hangzhou, and chose Alibaba, which seemed to be the "least technical" at the time.

On September 19, 2014, Wang Jian, then a partner of Alibaba, brought Qi Yuan to the New York Stock Exchange to witness Alibaba's listing. The seven bell ringers were all buyers and sellers of Alibaba's e-commerce, but none of them was a scientist.

Facing Wall Street, this company with a market value of 240 billion US dollars at the time was not satisfied with being defined as an "e-commerce company". He was planning a new and grand "imagination story" to give the company a stronger growth momentum in the future.

He understood that cutting-edge technology would be the protagonist of this story, and soon after that, iDST (Institute of Data Science & Technologies) was announced to be established, with Qi Yuan and Jin Rong becoming early founders.

But at that time and in that place, the story sounded a bit adventurous - the Chinese Internet market was still in the early stages of model innovation. How could a company based on commercial interests build a technology-driven research institute? There was almost no successful operating model in China to refer to.

The unknown challenges first fall on early founders and research mechanisms.

Qi Yuan, Director of the Financial Intelligence Laboratory at DAMO Academy

After joining Alibaba, Qi Yuan's first task was to build a large-scale machine learning platform and implement it on Taobao's advertising platform Alimama. He clearly remembers that he had 20 million features and 14 days of data at the time. Since the amount of data was too small, he planned to start after accumulating half a year of data.

At the same time, he also submitted a plan to apply for thousands of servers. This is not a small amount of money. Even now, an Nvidia computing card still costs $2,000. The plan was rejected after discussion.

After waiting for half a year, the server still hadn't been allocated. Even a good cook could not cook without rice. "At that time, there was no  GPU  cluster, and the entire group didn't have one," Qi Yuan said. "My colleagues thought, 'You professors here basically don't understand the business or the technology.'"

When a scholar meets a soldier, he can’t explain his reasoning clearly. Qi Yuan’s team was caught in a paradox of being unable to “prove themselves”: without a GPU cluster, how could they prove that their algorithms and technologies were efficient? Without being able to prove that their algorithms and technologies were efficient, how could they get a GPU cluster?

Qi Yuan once tried to argue with reason, and the stalemate lasted for nearly half a year. "The team was once on the verge of disbanding."

Jin Rong is another founder of iDST, a tenured professor at Michigan State University, and has won the National Science Foundation Award (Qi Yuan has also won this award) - more than 200 Nobel Prize winners have received funding from this award.

Jin Rong, currently the head of machine intelligence research at DAMO Academy

Jin Rong led the team to provide technical optimization for traffic distribution for Juhuasuan. The team naturally placed low-priced goods at the top of search and recommendation results to increase transaction volume, but neglected to understand the essence of the business - although low prices can stimulate purchases, they shift the target user group from second- and third-tier cities to third- and fourth-tier cities.

"Although GMV has increased, the value of the products has changed," said Jin Rong.

Business paved the way for technology, but scientists inadvertently tampered with the core of the product, "the shortcomings were exposed very obviously," Jin Rong said. At that time, most of the team members had a research background and were proficient in basic theories, but lacked business understanding and engineering practice experience, so they could not see the huge gap between technology and products.

"If you want to play your true value in Alibaba, you must overcome these difficulties," said Zhang Yong, Alibaba's COO at the time, to Jin Rong and several key members.

Since you are not familiar with the business, go to the first scene!

Therefore, iDST's early scientists split up into several groups, focusing on the two core businesses of e-commerce and finance, and delved into products and engineering. Jin Rong led the team to the search division of Taobao and Tmall, Qi Yuan and several colleagues went to Ant Financial, and the voice team stayed in Alibaba Cloud. Later, this was called the "going to the mountains and the countryside" movement of Alibaba scientists.

The awakening of technology is not limited to the top management. Earlier, a bottom-up innovation force emerged and the culture of internal innovation began to become popular - "as long as your boss doesn't object, it is the greatest support for you."

Li Hao was particularly impressed by the National Day in 2013. He had no plans to travel, nor did he have to go back to his hometown to visit the streets. Instead, he was alone in the office tinkering. For seven consecutive days, he was typing away at his keyboard, as if possessed.

Finally, before the end of the holiday, he breathed a sigh of relief - the demo ran smoothly, with a GPU and a host computer, and a purely handwritten deep learning algorithm for image search. It was a very simple device, but the result was gratifying - the accuracy was significantly improved compared to traditional algorithms.

As the third wave of artificial intelligence was gaining momentum, the ImageNet competition entered its third year. AlexNet, designed by the godfather of deep learning Geoff Hinton and his students, shined in the competition. The idea of ​​deep neural networks solved the thorny problem of image classification in one fell swoop, thus opening up a grand occasion of a hundred schools of thought in neural networks, with deeper and wider networks emerging one after another.

Li Hao was greatly encouraged when he read the news on foreign websites. He graduated with a Ph.D. from the Institute of Optoelectronics Technology of the Chinese Academy of Sciences and has been in Alibaba for less than a year. He is eager to put his knowledge into practice.