What kind of path did Liang Wenfeng follow when his 80,000 capital went to DeepSeek to shake Silicon Valley?

Author: Beating

Original title: Liang Wenfeng's Youth Past: 80,000 Capital, One Fiat, and One Man's Long March

Monday, January 27, 2025. Liang Wenfeng is playing with several middle school students on a court in Wu Chuan.

Seven days ago, Deepseek-R1 was released. Seven days later, the news crossed the Pacific Ocean, causing turmoil in the US capital market. Nvidia's stock price fell by more than 16% in a single day, and the market value evaporated by nearly 600 billion US dollars. The US media picked up old words from the Cold War era and called this day a new “Sputnik moment.”

DeepSeek is already at the top of the free app store list in both China and the US. People in Silicon Valley learned to read an unfamiliar Chinese name overnight. Interview emails flocked to Hangzhou from all over the world, and reporters went around asking where Liang Wenfeng had gone.

No one found him.

That afternoon, he wore a jersey and ran back and forth at a stadium in a county-level city in western Guangdong. The ball falls to his foot, he catches it, turns around, and passes it out.The world outside the field is shaking for him, and the game on the court continues.

The next day was New Year's Eve. When Liang Wenfeng returned to Mililing Village, he had already pulled up a banner with white letters on the red background, “Warmly welcome Wen Feng, the pride and hope of his hometown.”

Visitors came one by one, standing under the banner and taking pictures. Some people bent down, grabbed a handful of dirt from the side of the road, and carefully collected it, as if they wanted to take some good luck from this village. Liang Wenfeng's high school classmate, Mr. Chen, told the reporter that he promised to go back to Wu Chuan for the New Year, but after returning, he would find a place to hide.

Children of Mirei Ling Village

In 1985, Liang Wenfeng was born in Mililing Village, Qinba Town, Wuchuan City.

Wu Chuan is in western Guangdong. It is surrounded by rivers on three sides and faces the sea on the south side. The Jianjiang River came all the way down from the north, then suddenly turned a corner and flowed south into the sea. In the past, merchant ships running along the waterways were often anchored in the river bay, and the warehouses on the shore were connected, and it was very noisy. As a result, this county town received a very impressive name, “Little Buddha Mountain.”

Wu Chuan also believes in reading. Xiajie Village, the hometown of Qing champion Lin Zhaotang, is not far from Mililing Village. The county register records that there was a champion here, 20 contestants, and 165 people. For hundreds of years, generations of young people have set out from here and followed the same narrow road to counties, provincial towns, and Beijing. Fame has long been a thing of the past, and the idea that studying can change one's life has always remained in this land.

Mililing Village also has its own story. When talking about their ancestors, they talk about anti-Japanese general Leung Wah Shing, and Leung Tai-hee, who graduated from Sun Yat-sen University in 1936. According to village records, after liberation, nearly 100 students have been admitted to national key universities from here.

However, Liang Wenfeng grew up in the 90s. It was the age of money in Guangdong. Factories were recruiting people, business was booming, and coastal cities were the same day by day. A young man who is willing to suffer hardships and dares to break through may not need to sit in a classroom for many years; he also has the opportunity to earn money he never dared to think of before. In contrast, studying is too slow. You have to work year after year, and no one can say for sure what you can get in return.

Many parents in the village decided that studying was useless, and even made a special trip to Leung's family to persuade his father not to let his children continue reading. At that time, the imagination of fame that Mililing Village had accumulated over hundreds of years was being overshadowed by fresh and hot stories of wealth in Guangdong.

Many years later, Liang Wenfeng talked about this period of the past in an interview:

“I grew up in a fifth-tier city in Guangdong in the 80s. My father is an elementary school teacher. In the 90s, there were many opportunities to earn money in Guangdong. At the time, quite a few parents came to my house; basically, my parents felt that studying was useless. But now that I look back, my ideas have all changed. Because the money wasn't easy to make, I probably didn't even have the chance to drive a taxi. Time has changed in one generation.”

“Time has changed in one generation.”

This phrase refers to the 90s, and it also refers to 2024. When he said this, he was doing another thing that was just as slow and just as difficult to explain.

Liang Wenfeng's father taught at Meiyi Elementary School and later became the Deputy Teaching Director. Her mother was also a teacher and taught at Bianling Elementary School. This family doesn't have much to do with scientific research, but books are piled up in the attic all year round. According to local media reports, the largest collection of books was in the thousands.

My father set two rules for the family. At the dinner table, we don't talk about studying; we only talk about family life; after dinner, the whole family finds a book to read on their own. Whether Liang Wenfeng did well or not, his father rarely questioned him. What he says more often is that solving problems is more important than scores.

The Leung Family's most valuable appliance is a Feiyue radio. A black case, a metal knob, and an antenna that can be pulled out one section at a time. When the switch is turned, a rustling current sound comes out of the speaker. After the frequency is adjusted, the sound from a faraway place slowly comes out of the noise.

When Liang Wenfeng was in fourth grade, he took this radio apart. The screws are removed, the case is opened, and the dense threads and parts are exposed on the table. After disassembling it, putting it back and forth, he put it back and forth, and it is said that he went back and forth 37 times.

His father didn't teach him the circuit, and didn't stop him; he just handed him a screwdriver next to him and ordered, “Remember to put it back when you've disassembled it.”

Later, when the family had a TV, Liang Wenfeng began to disassemble the TV. One time, Liang Wenfeng tore up the TV. His father didn't scold him or confiscate his tools. He just watched the TV that never turned on and asked him, “Next time, how can I not tear my TV apart?”

His preschool class is at Meiling Elementary School, and his elementary school is at Mei Ling Elementary School. Teacher Li, the primary school class teacher, remembers that he was rarely distracted during class. He was particularly focused when he encountered difficult problems, and was good at summarizing methods. In the sixth grade, he won the second prize in the National Olympic Numbers Award. Also in that year, he was admitted to Wu Chuan No. 1 Middle School.

In his third year of middle school, his overall score was around 100 in the whole year. Not a top pick, but he often took first place in math competitions. In the 1999 middle school entrance examination, he scored 819 points, which was only more than 20 points higher than the admission line for Wuchuan No. 1 Middle School.

In his first year of high school, he jumped from the top 100 in his grade to the top ten. Yang Yahong, a senior year class teacher and math teacher, said that this kid is not assertive; he never has to worry about his homework; he is particularly excited in the math group. He takes part in the provincial math competition every year, winning at least the third prize. Teacher Rong, the middle school class teacher, also said that he finished high school math on his own in middle school and started college math. Wen is quiet, but he is not a nerd; it seems like it doesn't take much time to learn every subject well. In his middle school years, he fell in love with computers, disassembled and installed them, and helped the school repair computers.

In the summer of 2002, the college entrance examination list was released. Liang Wenfeng scored 806 points, out of 900, and became the winner of the Zhanjiang college entrance examination back then.

When he applied, he only wrote about Zhejiang University because the information and electronic engineering majors here are good.

Yang Yahong asked him if he was really unsure; he said yes.

It was later reported that with his score, he could have gone to Tsinghua.

People who don't attend class

In the fall of 2002, Liang Wenfeng went to Hangzhou to study electronic information engineering at Zhejiang University.

He seldom attends classes. Many years later, a student from a different class in the same class on Zhihu came back and recalled that Liang Wenfeng spent most of his time studying on his own, and felt that the teacher was slow and a waste of time. In his sophomore year, his classmate was catching up with his homework. He already drew his own circuit board, wrote MCU programs, made an interface, and made a software similar to MiniPlayer.

He also changed a regular guitar to an electric guitar, and the sound can be adjusted by connecting to a computer. After finishing it, he didn't think this thing was amazing; he just thought the guitar's pitch wasn't good enough; he said it would be better if it could tune automatically.

During the summer vacation of his junior year, he teamed up with two students to participate in the National College Student Electronic Design Competition. None of the three are usually in the top few in terms of professional results. During the training session at Zhejiang University, most of the design topics were handled by Liang Wenfeng alone. In the end, they won the first place in Zhejiang Province and the first prize in the country.

The awards were not announced until October, and the insurance window for that year was missed. As a result, Liang Wenfeng was one year late in his studies. In his free year, he continued to work on electronic sensing systems, which are said to be related to ocean navigation, and the hardware, software, and algorithms were still handled by him alone. An undergraduate student later commented that any of the electronic systems he had worked on in college was more than enough for a master's thesis in the Department of Electronics.

He also rode a poor bike. While in college, he rode a bicycle through several provinces in East China. He drove the road during the day and paving the floor in the field at night. It didn't cost much money to get around. This story was told by his teammate in the Zhejiang University campus forum. The title of the post also wrote “Liang Wenfeng” as “Liang Wenfeng.”

In 2007, he began studying for a master's degree in information and communication engineering at Zhejiang University. His research direction was machine vision, and his mentor was Xiang Zhiyu. Xiang Zhiyu's undergraduate, master's, and doctorate are all at Zhejiang University. Later, he went to Aveiro University in Portugal and Ohio State University in the US to do post-doctoral research, and stayed in the field of computer vision after returning home.

Liang Wenfeng's graduation thesis was “Research on Target Tracking Algorithms Based on Low-Cost PTZ Cameras”, submitted in May 2010.

The problem he wanted to solve was for an inexpensive, rotatable camera to keep an eye on a moving target in an environment where light, occlusion, and background were constantly changing. Equipment is limited and the environment is complicated. Thanks to Li for writing the paper, Mr. Xiang led him into the door of machine vision, prepared study plans for him, and often guided him line by line.

In 2011, this paper was revised and published in the “Journal of Zhejiang University (Engineering Edition)”. The author's brief at the end of the article has only one line:

Liang Wenfeng, born in 1985, is a master's student in Zhanjiang, Guangdong, engaged in machine vision research.

Up to this point, the file is neat and clean, like a standard good student.

Looking back many years later, there was already a problem in that paper that he needed to deal with over and over again later:How to identify the really important goals and keep doing them under limited conditions.

80,000 principal

2008, financial crisis. A-shares fell all the way from a high of 6,124 points to 1,664 points. The trading floor was bleak, and the backbone was cut down. Many people had only zero numbers left in their accounts.

Liang Wenfeng, who was in his second year of study, saw another incident in the midst of chaos. He thought that in this turbulent market, there might be a breakthrough hidden. He brought in like-minded students to study the use of machine learning in transactions.

According to later media reports, around 2008, in the laboratory at the Yuquan campus of Zhejiang University, several students studying machine vision began to think about a kind of “plug-in stock trading software.” Their initial idea was not ambitious; they just wanted to make back the money they had lost before. In the beginning, the programs they wrote still earned less money in the stock market. You can only start making money after repeatedly revising the optimization strategy. That thing they call a “plug-in” has a more formal name in the financial industry, quantitative trading.

Market data back then was not as easy to obtain as it is now. I took a screenshot of the data from the market software, and wrote my own program to try to connect to the trading software. His ability to research cameras was used by him for staring at the camera. He also collected data from all over the place, took relationships from financial institutions, then modeled and adjusted parameters. He wrote code day and night, often staying up late until the early hours of the morning.

Liang Wenfeng's initial capital was only 80,000 yuan. 80,000 yuan is not enough to buy a top-end video card today. In 2008, it was Liang Wenfeng's total net worth. He didn't explain this big gamble to anyone. At an age where others were busy brushing up resumes and looking for internships, he was busy facing the screen every day. He seldom explains to people what he is actually doing. Back then, quantitative trading wasn't a decent professional name in China.

In 2009, he went to Shanghai for an internship in Aiqi Information. The company was founded by Zhejiang University alumnus Zhou Chaoen. Before he graduated, he was appointed as the new technology department manager with a monthly salary of 16,000 and responsible for AI video imaging technology. When he left his job, Zhou Chaoen gave a suggestion to find a high-margin business. Liang Wenfeng has memorized this sentence for many years. The two met again in 2023, and he also took the initiative to mention it.

However, Ai Qi's message is just a branch line in his life. In 2009, while leading a high salary, he prepared for quantitative transactions. In his own life plan, that is the main line.

In June 2010, three students in the lab graduated with a master's degree. One went to a large factory, one chose to start a business, and the other went to Chengdu and continued to work on things that others couldn't understand.

Liang Wenfeng went to Chengdu.

Rental house in Chengdu, Fiat bound for Tibet

Liang Wenfeng's rental house in Chengdu is less than ten square meters, dark and humid. It only has a dilapidated desk and a single bed inside. Eat instant noodles when you're hungry, and lie down at the table to take a break when you're tired. The room is too small to fit anything other than a bed and a desk, but everyone will be crammed into this room in the future.

Graduate students graduated from famous universities, did not go to large factories, did not enter the system, and faced the screen every day. In the context of 2010, this is the opposite of the standard textbook. No one knows what he's doing against the screen. Dark Surge reports say that after graduation, he “hid in a cheap rental house in Chengdu and constantly accepted the frustration of trying in many scenes.”

On April 16, 2010, the Shanghai and Shenzhen 300 stock index futures were listed, and the main contract opened at 3,450 points. For the first time, China's capital market has mature shorting and hedging tools, and it has finally opened the door for quantitative trading.

It was this door that Liang Wenfeng was waiting for.

According to later reports, during those two years, he spent most of his time in a rental house in Chengdu, testing his ideas over and over again. These prices may seem chaotic, but they aren't necessarily irregular behind them. As long as there is enough data and the model is accurate enough, it may be possible to find a set of repeatable explanations from those ups and downs.

This is also the belief Simmons left to his latecomers. Originally a mathematician, Simmons later founded Renaissance Technology to trade using mathematical models and computer programs. As a result, his Medallion Fund became one of the most famous funds in the history of quantitative investment. He believes that market prices may seem cluttered, and there are still rules that can be identified and calculated. There must be a way to model prices; the only difficulty is whether that method can be found.

At the same time, Liang Wenfeng was also surrounded by a group of young people who were equally uneasy. Some people live in an urban village in Shenzhen to research unreliable aircraft, and they invited him to join them. Liang Wenfeng didn't go. Later, those people formed DJI.

In 2011, he didn't know the future of himself and his friends. Back then, he only had an unfinished model, a rented cottage, and a car.

In October of that year, Liang Wenfeng was 26 years old. Weibo had just become popular that year. People were basking their lives and loneliness on it. Just that fall, he drove a Fiat to Tibet.

It's not a car suitable for expeditions. Fiat is an established Italian car. Back then, it wasn't expensive in China; you could buy one for 560,000 yuan. It's an ordinary car that can be found everywhere on the street. Later, the brand withdrew from the Chinese market, and those cars slowly disappeared from the road.

Liang Wenfeng called it a “mount.”

At the time of departure, the car had almost fallen apart.

He has been tweeting along the way. If you spread out those microblogs one by one, it's just a page by page diary.

“October 19, Lhasa. Back to Lhasa. I'm moved ~”

Back in Lhasa, it's also the name of Zheng Jun's old song. People who drove into Tibet before probably had this melody swirled around in their minds.

“October 24, Yangzhuoyongcuo. The ancient road is westerly, thin horses, and the sun sets.”

Yangzhuoyongcuo is 4,441 meters above sea level. The lake surface is more than 600 square kilometers, and the shoreline is more than 200 kilometers. He shot the lake, the snowy mountains, and the Fiat.

“October 25th. View the mountains and small mountains at a glance.”

“October 26th, by the Brahmaputra River. Ride all over the world.”

“October 27th, Himalayas. Fiat visited here. Tibet, the Himalayas, and the other side of the snowy mountains is India.”

“October 28, lakeside. That lake is actually a few kilometers wide, and the snowy mountains are tens of kilometers away.”

Liang Wenfeng continued west.

He passed by Yung Bulacan. The palace stands at the top of a mountain. It is the earliest palace in the history of Tibet, and is even older than the Potala Palace. Afterwards, he walked along the northern foothills of the Himalayas in the direction of Mount Everest, photographed the snowy mountains from the north and west, then passed through the Shishabangma Nature Reserve and entered Ali.

The average elevation of Ali is over 4,500 meters; people call it the “roof of the roof of the world.” There is only about half of the air left over the plains. The road passes through the wasteland, and there are often hundreds of kilometers without people.

Liang Wenfeng drove the Fiat in.

He then disappeared on Weibo for a week.

When he reappeared, he began posting photos along the way. Entering 200 kilometers into an uninhabited area, he took pictures of Tibetan antelopes; at 400 kilometers, eagles hovered above the snow. Below in the photo, he wrote:

“Wind, Xiao Xi, Yi Shui Han ~”

Along the way, he still used those old verses to speak for himself until November 7, 2011, when he posted his last tweet:

“Stuck in no-man's land for a week. Luckily I came out. But my Fiat didn't come out. It stayed forever in that endless alpine grassland. That's the end.”

Since then, Liang Wenfeng has never updated his Weibo account. This account was stolen on August 19, 2014. The young man who used to photograph lakes and snowy mountains, and also laughed with strangers in the comments section slowly disappeared from the internet. The latter half of his life began the day he lost his car.

The last thing is to go or stay

After his trip to Tibet, Liang Wenfeng stopped updating his Weibo, but he did not completely disappear from the timeline of that group of Zhejiang University friends.

In those years, they would still ask each other questions in the comments section, discuss math, games, and ask each other where they finally decided where to go. Seo Jin and Jung Da-wei are the two names that appear the most. In a few years, they will work with Liang Wenfeng to establish Jacobi and become the earliest members of the Magic Square team. However, on Weibo in 2012, they were still just a few young people standing at a crossroads in their lives.

On August 23, 2012, Zheng Da-wei posted a post on Weibo: An ant can only crawl along the edge of a cube, starting from one vertex to the opposite vertex. Every time it reaches a vertex, it randomly selects the next edge to find the mathematical expectation of the final journey.

Zheng Da-wei featured Leung Man-fung and Xu Jin. Liang Wenfeng replied, “Seems like neither Xu Jin nor I came up with that question; it's definitely better than 4.”

Later, many people recalculated this question, and the answer was 10. He gave no proof at the time; he only determined that the answer must have been greater than 4. The direction is right.

Underneath this question, he asked Cheng Tat-wai:

“Is the last thing to go or stay?”

This statement has no context, and no further explanation. Back in the day, it was probably just an ordinary question between friends. Looking back many years later, it's like an early invitation, a math problem that hasn't been solved yet, a group of young people who are still choosing where to go, and a team with no name yet.

Cheng Da-wei later joined Magic Square and became a core member of DeepSeek. Xu Jin also walked into Liang Wenfeng's entrepreneurial story since then.

Zheng Da-wei left more than 4,000 tweets. If you look through them one by one, you can see a young man moving from his student days to society, and you can also see how that generation of skilled young people used to talk about their lives.

He said that he had experienced delivery four times before and after, saved four years, and didn't waste time taking the exam. He portrayed life as a triangle. The three horns are love, effort, and play. He wants to make “moving products,” and he also wants to promote new technology so that one day his parents can enjoy the changes brought about by technology.

In between these ideals, there are plenty of younger, more crude things sandwiched between them. He played Dota and wrote “I'd rather have one less person than a teammate like a pig”; he also retweeted a phrase that has been widely spread among programmers, to the effect that the smartest minds of our generation are studying how to get people to advertise one more time.

Sublime and joke are now on the same timeline. People back then didn't think this was a contradiction. They talk about math and games; if they want to change the world, they also get mad at a bad gaming teammate. Weibo hasn't become a carefully manicured personal window; what one person says is often left there in such a cluttered manner.

Xu Jin's Weibo has almost been deleted. In the remaining post, someone asked about Bao Song, and he answered very directly: “Because I don't need to go to my senior year of high school.”

His resume was later taken up again: a mixed class at Zhejiang University's Zhu Kezhen College, where he studied robot navigation at the doctoral level, and also participated in a project related to visual navigation of the “Yutu” lunar rover. The person who thought his senior year of high school was a waste of time on Weibo later actually sent the machine to a place where there was no way.

This young man's past is not just about Liang Wenfeng. Some of the people who later formed Magic Square and joined DeepSeek had already met in the same loose network of relationships at the time. The fascination with technology, impatience with inefficient rules, and an almost strict focus on smart peers all appeared earlier than the company.

Magic Square doesn't have a name yet; its initial personality grew better in the Weibo era.

In 2013, Liang Wenfeng returned to Hangzhou from Chengdu. He and Xu Jin and Zheng Dawei decided to continue to do “external stock trading” and start a business together on the grounds that “they can make more money than going to work.”

On September 10, 2013, Hangzhou Yakbi Investment Management Co., Ltd. was registered with a registered capital of 100,000 yuan. Liang Wenfeng held 50.5% of the shares, while Xu Jin, Yao Fengfeng, and Zheng Dawei each held 16.5%. The company's name comes from German mathematician Carl Jacoby.

Two years later, the market opened another door for them.

In April 2015, China Securities 500 stock index futures were listed, and quantitative traders had more new hedging tools in their hands. Two months later, Hangzhou Magic Square Technology was established. The name comes from the ancient Chinese Luoshu Jiugong map, a kind of matrix where the horizontal, vertical, and oblique additions are equal. In the same year, they moved into the Huijin International Building on Huancheng North Road in Hangzhou. At the time, there were few new high-end office buildings in Hangzhou. Rents were expensive, and most of them were private equity institutions.

In December 2015, a recruitment notice appeared in the Mizuki community. The notice did not directly write Liang Wenfeng's name; it only told the story of “Mr. L” in the third person:

In 2008, with a principal of 80,000 yuan, Mr. L began his own independent quantitative trading path. In 2015, Mr. L, who had experienced a seven-year round bear market, joined the billionaire team with a compound yield of over 100% each year.

The notice also stated that China's quantitative trading is about to bid farewell to the era of individual rangers and move into an era of private equity funds where geeks gather. Mr. L's ideal is to one day stand at the same table as Renaissance Technology, founded by Simmons.

A person who rarely talks about himself tells his past to a stranger for the first time, but hides himself behind a letter. He writes about himself in the third person, and his tone is as calm as if he is sorting out another person's resume.

80,000 yuan, seven years, over 100 million yuan.

This is the most complete diploma in his youth, and it is also his first confession to the world.

The Tibetan Weibo diary was written for myself, and the Mizuki community's recruitment notice was written for future colleagues. Between the two texts, four years apart, there is also a Fiat that has never left the uninhabited land of Ali.

Firefly

On October 21, 2016, Magic Square's first stock position generated by a deep learning model officially went live. Also from this day on, GPUs entered the Magic Square trading system.

Eight years ago, Liang Wenfeng also had to take data from images of market software, find an interface himself, and revise the rules one by one. Today, what stocks are bought, what are sold, and how many positions are placed is being left to the model to calculate. People retreat farther away, prepare data, design systems, and wait for machines to give answers.

In 2019, Liang Wenfeng explained the difference between the two at the Golden Bull Awards ceremony. When people make investment decisions, they rely on experience and feeling, and are closer to an art; when a program makes decisions, they are faced with a problem that can be solved, and there should be a better answer behind it.

From 2008 to 2016, he spent eight years trying to remove himself from every specific judgment. The model doesn't need to be excited, or panic, let alone change her temper because of the previous day's loss. It just takes data, calculates probabilities, and then executes them.

Quantifiers often start by trusting their own judgments, and finally establish a system that no longer requires their own judgment. Having arrived here, Liang Wenfeng is already able to manage more capital and continue to polish this machine more accurately.

But he didn't stop at the size of his capital.

To keep getting smarter, the system requires more data, more complex models, and more computing power. As a result, the number of GPUs in the Magic Square grew in a steep upward curve. Liang Wenfeng later recalled that at first, there was only one card; by 2015, it had increased to 100 cards, reached 1,000 in 2019, and since then it has reached 10,000 cards.

In 2019, Magic Square invested nearly 200 million yuan to build “Firefly 1”. This computing cluster has about 1,100 GPUs and covers an area close to a basketball court. Two years later, “Firefly 2” was completed, with an investment of 1 billion yuan to install about 10,000 Nvidia A100. The size of a computer room alone is close to ten basketball courts.

After quantifying private placement to make money, it is common to scale up, recruit traders, or find more strategies. Liang Wenfeng exchanged a large portion of that money for video cards, cabinets, and computing equipment that continues to roar. At the time, China's big model industry had not actually started, and not many people knew what such huge computing power could eventually be used for.

Someone asked him why he bought so many video cards.

Liang Wenfeng answered that an exciting thing probably cannot be measured by money alone. It's like buying a piano at home. One is that you can afford it, and the other is that there really are people who want to play it.

In 2018, a financial media went to Magic Square for research. Liang Wenfeng was holding a thermos mug and wearing a dark blue workwear cotton shirt. He was thin and had a slightly reserved look. He looked like an engineer from the 90s of the last century.

The interview was originally scheduled for half an hour, and we ended up talking for more than two hours. As soon as he asked about technology, he almost always answered; when he came across a place that was difficult to describe accurately, he stopped and laughed for a while.

The reporter asked how does Magic Square evaluate employees.

Liang Wenfeng said that the company has no assessment indicators. If a person hasn't contributed for a long time, the first thing they should reflect on is whether the company has placed him in the right position.

The reporter also asked what kind of company Magic Square would like to become in the end.

After thinking about it for a while, he said he hoped that one day, performance rewards and management fees would not be charged. After pausing for a moment, he added that what he really wanted to do was an open source strategy platform that ordinary investors could also use.

At that time, Magic Square was still a quantitative private equity firm, open source, and inclusive. Speaking from his mouth, it seemed somewhat far from business. However, Liang Wenfeng never seems to be satisfied with turning technology into an advantage for a few. He made money, but the next thing he thought about was how to unpack that set of abilities so that more people could get them.

Many of the things he later did at DeepSeek were outlined in this interview. Next to the thermos mug in 2018, the engineer, who is not good at gossiping, has already voiced these ideas.

In August 2021, Magic Square's management scale exceeded 100 billion, and the four kings were quantified together with Jiukun, Ming Gong, and Ling Jun.

After four months, the retracement came.

On December 28, 2021, Magic Square issued a public statement saying that the company experienced the biggest decline in performance since its establishment, and “feels deeply guilty” about this. Subsequently, Magic Square closed the fundraising channel and actively reduced the management scale. The numbers, which have been rapidly expanding in the past few years, are beginning to recede. The money machine that once operated at high speed showed signs of stalling for the first time.

Legends don't go all the way up.

In the few years that Magic Square contracted, funding declined, and applause from the outside world gradually waned. At the same time, those video cards that have already been bought, the computer room that has been built, and the young engineers gathered are beginning to be used for another unproven thing.

The years the Magic Square shrank happened to be the years DeepSeek nurtured. No one could see this at the time, including himself.

Epilogue

In January 2023, Magic Square announced the 2022 donation. The company donated 220 million yuan. Among them, an employee who signed “An Ordinary Little Pig” personally donated 138 million yuan. Outsiders speculated and guessed that the little pig was Leung Man-fung.

Three months later, Magic Square issued an announcement announcing its entry into the big model field. The poster quoted Truffer's words to the young director:

“You have to be insanely ambitious and insanely sincere.”

At the time, domestic big model startups were already crowded with big internet companies, star entrepreneurs, and capital. Magic Square suddenly swerved in from the side of the quantitative market, which seemed somewhat out of place. It has no cloud business, no search portal, and no huge user system. The only things it has in its hands are computing power, engineers, and a sum of money earned from the quantitative market.

In July 2023, Hangzhou Deepinqusuo Artificial Intelligence Basic Technology Research Co., Ltd. was established. The team consists of about 140 people. Many of the members are graduates of Tsinghua University and Peking University who have just left the campus. The average age is under 30.

Some people asked Leung Man-fung why not compete for those who are already famous and experienced.

He said that the criteria for selection have always been love and curiosity. The top 50 talents in the world may not be in China for a while, but China can train such people on its own.

In May 2024, DeepSeek-V2 was released, with an input of 1 yuan per million tokens and an output of 2 yuan. As a result, the price war for large domestic models was pushed to a new stage. Liang Wenfeng later described that they had no intention of making a pangasius that stirred up the market; they only accidentally became pangasius.

But low prices are just the result.

Dam Chung asked him why he had to go to the hardest place. Over the past few decades, the most familiar path for Chinese technology companies was to wait for the US to innovate at the bottom, then take back the technology, make products, sell markets, and run the scale. This path has gone fast, and it has actually made money.

Liang Wenfeng said that with the development of the economy, China must gradually change from a beneficiary of technology to a contributor, and must not always rely on the achievements of others.

In the IT revolution of the past 30 years, China has hardly ever actually participated in core technological innovation. Faster chips, better software, and updated architectures always arrive here after 18 months. Moore's Law is like a monsoon that arrives on time. People are used to standing on the shore and waiting, but they also gradually forget where the wind comes from.

Liang Wenfeng didn't want to wait any longer.

Now it is China's turn to contribute to the world.

Deepseek-v3 was released on December 26 of the same year. According to official disclosure, 2,048 Nvidia H800s were used in the final training phase, and the calculation cost was about US$5.576 million.

Low cost was his motive. The principal amount of 80,000 yuan, an inexpensive PTZ camera, a 560,000 yuan Fiat, and a training bill of 5576,000 US dollars. He always seems to find a path within limited circumstances. When there aren't enough resources, can the method be more accurate; when the equipment isn't expensive enough, can the system be smarter.

On January 20, 2025, DeepSeek-R1 was released, and inference capabilities entered the ranks of the world's top models and was opened under the MIT agreement. On the same day, Premier Li Qiang of the State Council hosted a symposium of experts, entrepreneurs, and representatives from UNESCO's health and sports fields. Liang Wenfeng was present and spoke about the development of large-scale domestic models.

Seven days later, DeepSeek caused Nvidia's market value to evaporate nearly $600 billion in a single day. He still went back to Wu Chuan during the Spring Festival. On the morning of the first day of the new year, he left the village.

On February 17, 2025, a private enterprise symposium was held. Liang Wenfeng sat in the first row. According to media reports, he did not speak at that meeting. He also rarely made public appearances for quite some time after that. The world is still discussing DeepSeek, and he himself has once again withdrawn from the crowd.

Liang Wenfeng's WeChat signature is rumored to “break up the world into mathematics.” I can't confirm this, but it's quite like him. Since disassembling the Feiyue radio, he has been doing the same thing all his life, opening the case, identifying the internal structure, and finding the rules that make it work.

Simmons founded Renaissance Technology at the age of 44. When Liang Wenfeng was 40 years old, he had spent 16 years in the quantitative market. Both of them once believed that behind the messy prices, there was an order that could be calculated. However, later, what Liang Wenfeng tried to model was no longer just the price.

On November 7, 2011, he wrote on Weibo that Fiat would stay in the endless alpine grasslands of Ali forever. It was the last diary he left for the internet. Since then, he has rarely recounted his life.

In the last Weibo that Liang Wenfeng left on the internet, the Fiat stopped forever on Ali's alpine prairie.

Over the years that followed, the wind blew through the wasteland, and the grass fell season by season. People don't know where that car went later. To those who read that Weibo, it seemed like it stayed there, slowly rusting off the tin, and eventually became part of a no-man's land.

That's where people came from. Those who walk out don't have to look back.

Twitter:https://twitter.com/BitpushNewsCN

Compare the TG exchange group:https://t.me/BitPushCommunity

Compare TG subscriptions:https://t.me/bitpush