Why did Liang Wenfeng create DeepSeek?

Source: AI Technology Review

This is the hottest technology star at the beginning of 2025. In just a few days, Liang Wenfeng's past from childhood to adulthood has been shown to the world, including his new house that he did not have time to decorate and the tent he used to sleep in the house, which have become symbols of his unique personality.

Unique personality is certainly talked about, but it is not the key to success. This unknown college student has only his thoughts and abilities to rely on in the past ten years.

Everyone is curious about this question, why did Liang Wenfeng create DeepSeek? This is certainly due to the factors of the times, as well as his personal experience that is very different from other large model researchers. But AI Technology Review believes that understanding what kind of person Liang Wenfeng is is the key to understanding this problem.

1 Finding talents does not require labels

Headhunters feel that it is too difficult to help Liang Wenfeng's company find people.

A headhunter who has been working closely with Huanfang since 21 years told us that recruiting people makes him "want to cry" because it is too difficult.

"Tsinghua undergraduate and doctoral degree, six top conference papers, you think there must be no problem, hey, how come the resume is directly rejected; a Tsinghua undergraduate and MIT doctoral degree, was eliminated in the second round of interviews."

If you want to find candidates within large companies, he believes that Huanfang and DeepSeek basically do not benchmark domestic companies, they only benchmark overseas large companies such as Google and Meta.

Another headhunter couldn't help but feel overwhelmed when talking about DeepSeek. "It's too picky. I recommended a young middle-level manager who had very good performance in ByteDance, but he didn't pass the interview. I was very surprised. I asked them, and they told me that this person has no passion for AI. They have all done some AI Agent-related projects, and generally don't make such comments."

Liang Wenfeng doesn't label talents. Regardless of academic background or past performance, he only looks at the person's personal ability and personal quality.

The extremely high talent threshold has created today's DeepSeek. Among the large model teams in China, DeepSeek's talent density may not be enough to be compared with top companies, but its talent density can definitely be said to be the first tier.

In addition to DeepSeek's high salary, it also has a management model that fully respects creativity and ideas to retain these talents. "No fixed team, no reporting relationship, no annual plan" is more about trust than management. The book "Netflix Culture Handbook" once said, "Excellent colleagues and difficult challenges are the biggest factors that attract people to work in the company." For AI practitioners, there is no greater challenge than AGI.

To do the most difficult thing, you must find the best people and give them sufficient resources and trust. Top talents who are trusted often bring huge explosive power, and this theory can be verified in the rise of Douyin.

During the Spring Festival in 2018, Douyin added more than 10 million new users per day. A product manager in charge of growth once mentioned that this growth project had no performance pressure at all, and after sending an email to the finance department, his account had an additional budget of more than 100 million yuan. He realized at the time, "Such a team, what can't it win?"

The same is true for DeepSeek. Those whose resumes were screened out must not have academic qualifications; those who failed the interview must not have ability problems; the demand for talent can be summarized in one sentence: Is this person someone who can be trusted to work together for AGI?

This is DeepSeek's talent view. Understanding this talent view is the first step to understanding Liang Wenfeng.

2 Minimalist values

Although he has been doing quantitative investment for many years, Liang Wenfeng does not think of himself as a financial person. He thinks of himself as, "I am doing AI, but I am doing quantitative scenarios."

Almost everyone who has communicated with Liang Wenfeng said that he is a person who will not be disturbed by the outside world, "his way of thinking is extremely pure, and he pays special attention to the first principles," "he speaks very slowly," and "he gets to the point as soon as he opens his mouth."

The characteristics of quantitative investment just fit his minimalist style - it does not need to deal with complex upstream and downstream industrial chains, but only needs to focus on pure market data.

To this day, Liang Wenfeng is still often immersed in his own technical world, focusing on solving problems. For example, when it comes to making a large model, he would tell others, "You can do it if you think it through, as long as you have a card", and other difficulties are not within the scope of consideration.

The same is true for money. Money is used for investment or charity. As long as it can be spent in the right place, the loss is not worth mentioning.

At the end of 2023, there was a large sign language model project aimed at helping the deaf and mute, and Liang Wenfeng was found to attract investment. Liang Wenfeng proposed that the advantage of this project is its outstanding public welfare, and the disadvantage is that the market size is limited. The hidden danger is that this is a project of a team of college students from a top university, and they may not persist for a long time.

Although it is very likely that there will be no return, he still proposed that as long as the team is willing to continue to advance the project, he is willing to invest.

In the past, Liang Wenfeng would spend 500 million yuan every year on investment or charity, and now he spends this money on DeepSeek. Stock trading is to make money, and investing in large models is for AGI, that's all.

DeepSeek has nearly 20,000 cards, and he is extremely generous with computing power. He promised the above-mentioned sign language large model team that the computing power cluster will be open to them at any time. But he is a bit "stingy" and has high requirements for the utilization rate of these nearly 20,000 cards, striving to use them fully and not to idle.

These two behaviors seem contradictory, but if explained from the perspective of minimalism, it will work: the existence of cards is to be used, and they can be used as much as possible, and never waste them.

3 Not limited to commercialization

Without spending a penny on advertising, DeepSeek's App only took 7 days to get 100 million users. What does Liang Wenfeng think of this miraculous growth?

An investor specifically asked Liang Wenfeng this question during the Spring Festival, but Liang Wenfeng seemed to not care about such a large amount of traffic at all. The answer the investor got was, "This is still a long way from AGI."

This is not Liang Wenfeng pretending. According to AI Technology Review, DeepSeek only arranged two or three people to be responsible for App maintenance, dialogue web page development, and recharge backend management. So it is normal that it is not easy to use.

DeepSeek's various deeds in the B-end market are more widely circulated. For example, their private deployment was previously priced at only 450,000 yuan, which not only included the right to use an H20 or 910b, but also came with a large model service with a one-year use period. For the same price, you can only rent the right to use 910b for one year on Huawei Cloud, which means that DeepSeek's large model is almost free.

Private deployment does not make money, and DeepSeek does not care whether it makes money from APIs. An employee of a large company that connects to DeepSeek complained that it has a "use it or not" temperament, it is always difficult to use, and it is never adjusted.

No matter how big the customer and the call volume are, it is not worth looking at it differently. All large companies have to queue up during peak hours, and the user experience is very poor. There are also many feedbacks from large customers, requiring DeepSeek to expand and expand again, and at least respond more smoothly, and not fail once in two requests, which is almost unbearable.

The outside world is noisy, but Liang Wenfeng does not seem to care about this matter.

How should this situation be solved? Many companies are troubled by this. According to some internal rumors, Liang Wenfeng believes that large companies are fully capable of finding ways to solve the problem of request failures on their own, and they should cover themselves instead of relying too much on DeepSeek to guarantee services.

This answer is almost enough to make people laugh.

It can be said that Liang Wenfeng now does not care about all the possibilities of commercialization.

Today, when many teams are investing in applications, Liang Wenfeng once said to a good friend, "Don't keep looking at the application and industry implementation. If you look at it now, you will only imprison yourself, because it's not time yet, and everything you think now is wrong. And you have invested more time, energy and money on the wrong path."

This is advice to friends and also his own practice. For Liang Wenfeng, investing energy in applications and commercialization is a wrong path no matter what he does.

And there is only one right path, and he is now on the right path.