Originally published on Weixin Official Accounts Platform
Original by “Waves (暗涌)”
July 17, 2024, 09:01
Written by: Yu Lili
Edited by: Liu Jing
Among China’s seven major large-model startups, DeepSeek (深度求索) is perhaps the most unassuming – yet it always manages to be remembered in unexpected ways.
One year ago, its element of surprise stemmed from its backer: the quantitative private-equity giant Huanfang—then the only company outside the tech giants to have tens of thousands of A100 chips in reserve. Now, DeepSeek itself has become the catalyst for China’s large-model price war.
In a May when the AI news cycle was relentless, DeepSeek suddenly shot to fame with the release of its open‑source model, DeepSeek V2. The model offered an unprecedented cost–performance ratio: inference costs were slashed to only 1 yuan per one million tokens—roughly one‑seventh the cost of Llama3 70B and one‑seventieth that of GPT‑4 Turbo.
Dubbed the “Pinduoduo of the AI world,” DeepSeek’s price drop set off a chain reaction. Giants like ByteDance, Tencent, Baidu, Alibaba, and others couldn’t help but follow suit—thus igniting a fierce price war in the Chinese large-model arena.
Note: Amid all the competitive “smoke,” one fact stands out: unlike many major companies burning cash on subsidies, DeepSeek is profitable.
Behind the scenes, DeepSeek achieved these results by innovating across the board on model architecture. The team introduced a brand‑new MLA (a novel multi‑head latent attention mechanism) that reduces memory usage to just 5–13% of that of the conventional MHA architecture. At the same time, its unique DeepSeekMoESparse structure minimizes computational cost—and all of these breakthroughs ultimately helped drive down the overall cost.
In Silicon Valley, DeepSeek is being hailed as “a mysterious force from the East.” The chief analyst at SemiAnalysis even commented that the DeepSeek V2 paper “might be the best one this year.” Former OpenAI employee Andrew Carr remarked that the paper “is full of astonishing insights” and has already influenced the training setups in his own models. Meanwhile, former OpenAI policy director and Anthropic co‑founder Jack Clark noted that DeepSeek “has hired a group of inscrutable geniuses” and predicted that Chinese‑made large models “will, like drones and electric vehicles, become an unignorable force.”
In an AI wave whose narrative is typically steered by Silicon Valley, this is a rare situation. Multiple industry insiders have told us that the strong reaction to DeepSeek V2 comes from its architectural innovations—a bold attempt rarely seen even among domestic large‑model companies or global open‑source base models. One AI researcher even said that since the advent of the attention mechanism, it has almost never been successfully altered on a large scale. “This is even an idea that gets nixed at the decision‑making stage because most people lack confidence.”
On the other hand, domestic large‑model companies have seldom ventured into architectural innovation—largely because few are willing to shatter the long‑held preconception that “the U.S. is better at zero‑to‑one technological innovation, while China excels in one‑to‑ten application innovation.” Moreover, pursuing innovation in model structure isn’t very “profitable” in the short term. New‑generation models will eventually be built by someone; Chinese companies can simply follow and build applications. But innovating in model architecture means there is no preset path, and it involves many failures along with enormous time and financial costs.
DeepSeek is clearly a contrarian. While most chatter that large‑model technology is destined to converge and that “following” is the smarter shortcut, DeepSeek values the knowledge gained along a more winding road. It believes that Chinese large‑model entrepreneurs should not only focus on application innovation—they can also join the global tide of technological innovation.
Indeed, many of DeepSeek’s choices stand out. So far, among the seven Chinese large‑model startups, DeepSeek is the only one that has abandoned the “have it all” approach—focusing solely on research and technology without developing consumer‑facing applications. It is also the only company that has not fully pursued commercialization, steadfastly choosing an open‑source path (and not even having raised funding). These factors often leave DeepSeek overlooked at the negotiating table—but on the other hand, it spreads organically through the community via word‑of‑mouth.
How was DeepSeek forged? To answer that, we sat down for an interview with DeepSeek’s famously reclusive founder, Liang Wenfeng.
A product of the Huanfang era, Liang (an ‘80s kid) has always been deeply immersed in technology behind the scenes. Even in the DeepSeek era, he remains low‑key, much like his research team—spending his days “reading papers, writing code, and participating in group discussions.”
Unlike many founders of quantitative funds who have stints at overseas hedge funds and hail from physics or mathematics backgrounds, Liang has always been local. He studied artificial intelligence in the Electronic Engineering Department at Zhejiang University early on.
Industry insiders and DeepSeek researchers alike describe Liang as a rarity in today’s Chinese AI scene—a person who combines strong infrastructure engineering skills with cutting‑edge model research, who can mobilize resources effectively, make high‑level strategic judgments while excelling in details, and who possesses a “terrifying learning ability” yet “doesn’t come off as a typical boss at all, but rather like a geek.”
This is an especially rare interview. In our conversation, this technological idealist offered a voice that is particularly scarce in China’s tech community today: he is one of the few who puts “a sense of right and wrong” before “a sense of profit and loss” and reminds us to recognize the inertia of our times and prioritize original innovation.
A year ago—when DeepSeek first emerged—we interviewed Liang in “Crazy Huanfang: The Path of a Hidden AI Giant’s Large Model”. Back then the rallying cry was “you must passionately harbor ambition, and be wildly sincere.” A year later, that slogan is now a call to action.
Below is our conversation with Liang Wenfeng.
Interviewer (Waves):
“After the release of the DeepSeek V2 model, a bloody price war erupted in the large-model arena. Some even called you the ‘catfish’ of the industry.”
Liang Wenfeng:
“We never intended to be a catfish; we simply ended up being one by accident.”
Interviewer (Waves):
“Were you surprised by this outcome?”
Liang Wenfeng:
“Very much so. We didn’t expect people to be so sensitive about the price. We simply worked at our own pace and priced based on our cost calculations. Our principle is neither to subsidize nor to charge exorbitant profits—the price reflects a modest profit over cost.”
Interviewer (Waves):
“Just five days later, Zhipu AI followed suit, and soon after, giants like ByteDance, Alibaba, Baidu, and Tencent dropped their prices. What happened?”
Liang Wenfeng:
“Zhipu AI reduced the price for an entry‑level product, while models in our category remained expensive. ByteDance was truly the first to follow—their flagship model dropped to our price, which then triggered other big companies to lower theirs. Since the cost structure for big companies is much higher than ours, we never expected anyone to lose money doing this. In the end, it turned into the classic internet‑era model of burning cash for user subsidies.”
Interviewer (Waves):
“From the outside, the price drop looks like an attempt to capture users—the kind of price war common in the internet era.”
Liang Wenfeng:
“Capturing users isn’t our primary goal. On one hand, our reduced cost came from our exploration into next‑generation model architectures; on the other, we believe that both APIs and AI should be accessible and affordable for everyone.”
Interviewer (Waves):
“Until now, most Chinese companies simply copied the current Llama architecture to build applications. Why did you choose to focus on innovating the model architecture?”
Liang Wenfeng:
“If your goal is to build applications, then sticking with Llama for a quick product launch is a reasonable choice. But our aim is AGI. This means we need to research new model architectures that can deliver greater capabilities with limited resources. It’s one of the fundamental research areas required to scale up to larger models. Besides the architecture, we’ve invested heavily in research on data construction, making models more human‑like, and more—and all of that is embodied in the models we release. Moreover, Llama’s architecture is, in terms of training efficiency and inference cost, likely two generations behind the state‑of‑the‑art abroad.”
Interviewer (Waves):
“Where do you think this gap primarily comes from?”
Liang Wenfeng:
“First, there’s a gap in training efficiency. We estimate that even the best domestic systems are about twice as inefficient as the best overseas systems in terms of model architecture and training dynamics—meaning we need twice as much compute to achieve the same effect. Then there’s data efficiency; again, we may need twice as much training data and compute to match the results. In total, that’s about four times as much compute. What we’re trying to do is continuously narrow these gaps.”
Interviewer (Waves):
“Most Chinese companies are determined to develop both the model and the application. Why has DeepSeek chosen to focus exclusively on research and exploration?”
Liang Wenfeng:
“Because we believe that what’s most important now is to participate in the global wave of innovation. For years, Chinese companies have been accustomed to others doing the technological innovation while we simply adopted and commercialized applications—but that isn’t something we should take for granted. In this wave, our starting point isn’t to make a quick profit but to be at the technological forefront and drive the ecosystem forward.”
Interviewer (Waves):
“In the internet and mobile internet era, many have come to think that the U.S. is great at technological innovation, while China is better at application development.”
Liang Wenfeng:
“We believe that as the economy develops, China must gradually become a contributor rather than always being a free rider. Over the past 30-plus years of the IT wave, we barely participated in genuine technological innovation. We’ve grown accustomed to Moore’s Law delivering better hardware and software every 18 months, and even Scaling Laws are taken for granted. In truth, these breakthroughs were created painstakingly over generations by the Western tech community. Because we didn’t participate in that process before, we overlooked their importance.”
Interviewer (Waves):
“Why did DeepSeek V2 manage to astonish so many people in Silicon Valley?”
Liang Wenfeng:
“In the midst of the daily barrage of innovations in the U.S., this is quite ordinary. They were surprised simply because here is a Chinese company entering the arena as an innovative contributor. Most Chinese companies are used to following rather than innovating.”
Interviewer (Waves):
“Choosing to focus solely on innovation seems extravagant in China—large‑model work is capital‑intensive, and not every company has the resources to focus only on research without first considering commercialization.”
Liang Wenfeng:
“Innovation certainly isn’t cheap, and the old habit of simply adopting others’ ideas was partly due to past national circumstances. But now, given the scale of China’s economy and the profits of giants like ByteDance and Tencent, capital is not what we lack—it’s the confidence and the know‑how to organize highly dense talent to innovate effectively.”
Q11: Why Do So Many Chinese Companies, Even Those Flush with Cash, Prioritize Quick Commercialization?
Interviewer (Waves):
“Why is it that Chinese companies—even large ones that aren’t short on cash—tend to prioritize rapid commercialization?”
Liang Wenfeng:
“For the past 30 years, we’ve focused solely on making money and neglected innovation. Innovation isn’t driven entirely by commerce; it requires curiosity and a creative spirit. We were simply bound by the old inertia—and that is only temporary.”
Interviewer (Waves):
“You’re a commercial organization, not a non‑profit research institute. By choosing to innovate and then open‑source your work, where do you form a competitive moat? For example, won’t your MLA innovation be quickly copied?”
Liang Wenfeng:
When it comes to disruptive technology, any moat built by keeping things closed is only temporary. Even if OpenAI stays closed, it won’t stop others from catching up.
“That’s why we concentrate our value in our team. Our colleagues grow and accumulate a vast amount of know‑how during this process, forming an organization and culture capable of continuous innovation—that is our real moat.
Open‑sourcing and publishing papers don’t cost us anything. For engineers, having others follow your work is a huge source of pride. In fact, open‑source is more of a cultural act than a commercial one. Giving away knowledge is an extra honor, and a company that does so gains cultural appeal.”
Interviewer (Waves):
“What do you make of market-believers such as Zhu Xiaohu, whose views seem to prioritize fast money?”
Liang Wenfeng:
“Zhu Xiaohu is self‑consistent, but his approach is more suited for companies looking to make quick profits. Just look at the most profitable companies in the U.S.—they are high‑tech firms that have built up strength over time.”
Interviewer (Waves):
“Technological superiority alone is rarely enough to maintain a permanent lead. What larger bet are you making?”
Liang Wenfeng:
We see that Chinese AI cannot remain in the follower position forever.
“We often say Chinese AI is one or two years behind the U.S., but the real gap is between originality and imitation. If that gap isn’t closed, China will always remain a follower. Certain explorations are simply unavoidable.
NVIDIA’s lead, for example, is not just the result of one company’s effort—it’s the outcome of a whole Western tech community and industry working together. They can see the next generation of trends and have roadmaps in hand. Chinese AI development needs such an ecosystem too. Many domestic chips fail because they lack a supporting tech community and only rely on second‑hand information. Someone in China must step up to the forefront of technology.”
Interviewer (Waves):
“DeepSeek currently exudes an early‑OpenAI idealism and is completely open‑source. Will you ever choose to go closed‑source—as OpenAI and Mistral once did?”
Liang Wenfeng:
“We will not go closed‑source. We believe that building a strong technological ecosystem comes first.”
Interviewer (Waves):
“Some media report that Huanfang is planning to spin off DeepSeek for an independent IPO, and that Silicon Valley’s AI startups inevitably end up tied to big companies. Do you have any funding plans?”
Liang Wenfeng:
“In the short term, we have no plans for financing. The problem we face isn’t money—it’s the embargo on high‑end chips.”
Interviewer (Waves):
“Many believe that building AGI is entirely different from doing quantitative work—quantitative projects can be done quietly, but AGI requires high‑profile moves and alliances to scale up the investment. What do you think?”
Liang Wenfeng:
“More investment does not necessarily generate more innovation. Otherwise, big companies could scoop up all the breakthroughs.”
Interviewer (Waves):
“Since you’re not developing applications right now, is that because you lack the operational know‑how?”
Liang Wenfeng:
“We believe this is a period of explosive technological innovation rather than an explosion of applications. In the long run, we hope to build an ecosystem where the industry directly uses our technology and outputs. We will focus on foundational models and cutting‑edge innovations, while other companies build B2B and B2C businesses on top of DeepSeek. If a full industry chain develops, there’s no need for us to build applications ourselves. Of course, if necessary, we can develop applications—but research and technological innovation will always come first.”
Interviewer (Waves):
“If you were to offer an API, why would someone choose DeepSeek instead of a tech giant?”
Liang Wenfeng:
“The future will likely be one of specialized division of labor. Foundational large models need continual innovation—and big companies have their own limits, which may not be ideal for this role.”
Interviewer (Waves):
“But can technological innovation really create a lasting competitive gap? You yourself have said that there are no absolute technological secrets.”
Liang Wenfeng:
“Technology has no secrets—but resetting a system takes time and cost. NVIDIA’s GPUs, for instance, have no true secrets and are relatively easy to copy. However, reorganizing a team and catching up with the next generation of technology takes considerable time. That’s why the actual moat remains wide.”
Interviewer (Waves):
“After your price drop, ByteDance was the first to follow—which suggests they felt threatened. What do you think about the new methods of competition between startups and big companies?”
Liang Wenfeng:
“To be honest, we don’t really care about that; we just happened to do what we did. Providing cloud services isn’t our main goal—our goal is still to achieve AGI.
At present, we haven’t seen any radically new solutions—and big companies don’t hold a clear advantage either. They have ready‑made user bases, but their cash flow businesses can also be burdensome, making them vulnerable to disruption.”
Interviewer (Waves):
“What do you think will be the eventual outcome for the other six large‑model startups apart from DeepSeek?”
Liang Wenfeng:
“Perhaps only two or three will survive. Right now, everyone is in the money‑burning stage. Those with clear self‑positioning and more refined operations are more likely to endure. Others may reinvent themselves. Valuable innovations won’t simply vanish—they’ll transform.”
Interviewer (Waves):
“During the Huanfang era, your competitive approach was described as ‘doing things your own way’ without worrying about horizontal comparisons. What is your fundamental perspective on competition?”
Liang Wenfeng:
“I always ask myself whether something can increase the efficiency of society and whether you can find a niche in the industrial value chain where you excel. As long as the end result is a more efficient society, it’s justified. Much of what happens in between is temporary, and over‑focusing on it only leads to distraction.”
Interviewer (Waves):
“Former OpenAI policy director and Anthropic co‑founder Jack Clark said that DeepSeek hired ‘a group of inscrutable geniuses.’ What kind of people are behind DeepSeek V2?”
Liang Wenfeng:
“There aren’t any inscrutable geniuses. Our team is made up of fresh graduates from top universities, PhD candidates (and even interns in their fourth or fifth year), as well as young professionals who have only been out of school for a few years.”
Interviewer (Waves):
“Many large‑model companies go overseas to recruit talent. Some say that the top 50 experts in this field might not be in Chinese companies. Where do your people come from?”
Liang Wenfeng:
“Our V2 model’s team is entirely local—none of our talent are returnees from overseas. It may be that the top 50 experts aren’t in China, but perhaps we can cultivate them ourselves.”
Interviewer (Waves):
“How did the MLA innovation come about? I heard that the idea originally stemmed from a young researcher’s personal interest?”
Liang Wenfeng:
“After summarizing some of the evolution patterns of the attention mechanism, one of our young researchers had a sudden inspiration to design an alternative. However, the process from idea to implementation was long—we formed a dedicated team, and it took several months to get it working.”
Interviewer (Waves):
“Your open and innovative organizational structure is known for its bottom‑up approach. During the Huanfang era, you rarely assigned tasks from the top down. With an unpredictable frontier like AGI, did you need to introduce more management measures?”
Liang Wenfeng:
“At DeepSeek everything is bottom‑up. We don’t pre‑assign divisions of labor—the roles form naturally. Every person comes with their own background and ideas, so there’s no need to push them. If someone encounters a problem, they naturally bring others in for discussion. However, when an idea shows potential, we do allocate resources from the top down.”
Interviewer (Waves):
“We’ve heard that DeepSeek is very flexible when it comes to reallocating compute resources (“cards”) and people.”
Liang Wenfeng:
“There is no limit on how anyone here can mobilize compute resources or personnel. If you have an idea, you can call upon the training cluster without approval. And because there are no strict hierarchical or departmental boundaries, you can flexibly engage anyone—as long as they’re interested.”
Interviewer (Waves):
“Your loose management style depends on having recruited highly passionate and driven people. I hear you’re very good at spotting talent using non‑traditional evaluation criteria.”
Liang Wenfeng:
“Our standard has always been passion and curiosity. Many candidates have unique experiences—and that’s very interesting. Their thirst for research far exceeds any concern for money.”
Interviewer (Waves):
“The Transformer was born at Google’s AI Lab and ChatGPT at OpenAI. What do you think is the difference in the value of innovation produced by big companies’ AI Labs versus a startup like yours?”
Liang Wenfeng:
“Whether it’s Google’s lab, OpenAI, or even the AI Labs of Chinese tech giants, they all have value. In the end, that OpenAI breakthrough also involved an element of historical chance.”
Interviewer (Waves):
“To a large extent, isn’t innovation also about serendipity? I noticed that in your office, the meeting rooms have doors that can be freely pushed open—your colleagues say that’s to leave room for serendipity. There’s even a story of someone casually overhearing a conversation during the birth of the Transformer and joining in, which eventually turned it into a general‑purpose framework.”
Liang Wenfeng:
“I believe that innovation is first and foremost a matter of belief. Why is Silicon Valley so rich in innovative spirit? Because they dare. When ChatGPT was released, there was a general lack of confidence here about pursuing frontier innovation—from investors to big companies, everyone thought the gap was too huge, so they stuck with applications. But innovation requires self‑confidence. And that confidence is usually most apparent among young people.”
Interviewer (Waves):
“You don’t raise funding and rarely make public announcements, so your social profile isn’t as high as those companies that are constantly in the news. How do you ensure that DeepSeek becomes the first choice for those building large models?”
Liang Wenfeng:
“Because we’re tackling the hardest problems.
The greatest magnet for top talent is the opportunity to solve the world’s toughest challenges.
In truth, top talent in China is under‑recognized. With so little hardcore innovation happening at a societal level, they haven’t had the opportunity to be noticed. What we’re doing—taking on the most difficult tasks—is inherently attractive to them.”
Interviewer (Waves):
“Recently, OpenAI’s release didn’t lead to GPT‑5, and many believe the technological curve is clearly slowing down. Some are even questioning the validity of Scaling Laws. What’s your take?”
Liang Wenfeng:
“We’re fairly optimistic—the industry overall appears to be on track. OpenAI isn’t a deity; it can’t always be at the forefront.”
Interviewer (Waves):
“How long do you think it will take to achieve AGI? Before releasing DeepSeek V2, you introduced models for code generation and mathematics and shifted from dense models to MOE. What are the milestones on your AGI roadmap?”
Liang Wenfeng:
“It might be 2, 5, or even 10 years—whatever the case, it will happen within our lifetimes. As for the roadmap, even within our company there’s no unified opinion. However, we are betting on three directions:
- Mathematics and Code – a natural testing ground for AGI, akin to Go; it’s a closed and verifiable system where high intelligence might be achieved through self‑learning.
- Multimodality – engaging with the real world is also essential for AGI.
- Natural Language – the core of human communication. We remain open to all possibilities.”
Interviewer (Waves):
“What do you think the ultimate landscape of large models will be?”
Liang Wenfeng:
“There will be specialized companies providing foundational models and services, with a long chain of specialized division of labor. On top of that, many others will build applications to meet the diverse needs of society.”
Interviewer (Waves):
“Over the past year, we’ve seen many changes among Chinese large‑model startups. For instance, someone like Wang Huiwen—who was very active at the start of last year—has since withdrawn, and newer companies are beginning to show differentiation.”
Liang Wenfeng:
“Wang Huiwen took on all the losses himself so that others could exit unscathed. He made a choice that was most disadvantageous to himself but beneficial to everyone—and I truly admire that.”
Interviewer (Waves):
“What are you focusing on most these days?”
Liang Wenfeng:
“My primary energy is devoted to researching the next‑generation large models. There are still many unresolved problems.”
Interviewer (Waves):
“Other startups are trying to have it both ways—developing the model and building products—because technology alone won’t secure a permanent lead. Is DeepSeek’s focus on model research a result of insufficient model capability?”
Liang Wenfeng:
“All conventional approaches are relics of the previous generation; they may not hold in the future. Using internet‑era business logic to discuss AI’s future profitability is like comparing General Electric to Coca‑Cola when Ma Huateng was starting out—it’s like trying to carve a new path in stone.”
Interviewer (Waves):
“Huanfang already had strong technological and innovative DNA and grew relatively smoothly. Is that why you’re optimistic?”
Liang Wenfeng:
“Huanfang did boost our confidence in technology‑driven innovation—but it wasn’t all smooth sailing. We went through a long period of accumulation. What outsiders saw was only the post‑2015 phase, but in fact, we have been working on this for 16 years.”
Interviewer (Waves):
“Now that the economy is slowing and capital is in a cooling cycle, do you think original innovation will be further suppressed?”
Liang Wenfeng:
“I don’t think so. As China’s industrial structure adjusts, there will be even more reliance on hardcore technological innovation. When people realize that the fast money of the past was largely due to being in the right place at the right time, they will be more willing to bend down and pursue genuine innovation.”
Interviewer (Waves):
“So you’re optimistic about the future?”
Liang Wenfeng:
“I grew up in a fifth‑tier city in Guangdong during the 1980s. My father was an elementary school teacher. In the 1990s, there were plenty of money‑making opportunities in Guangdong—many parents would come to our house, convinced that studying was pointless. But now, looking back, attitudes have changed. It’s hard to make money anymore—there might not even be an opportunity to drive a taxi. Times change for an entire generation.
In the future, hardcore innovation will only increase. It may not be easily understood now because society needs to be educated by facts. Once society sees that those who pursue hardcore innovation achieve success and recognition, collective attitudes will change. We just need a body of facts and time for the process.”
Image Source: IC Photo
Layout: Yao Nan
Scan the QR code below to follow DeepSeek’s Mini Program or read the original article.
End of Interview