Translated from a July 2024 interview with DeepSeek founder Liang Wenfeng—conducted shortly after the company’s open-source V2 model catapulted it to fame—this rare dialogue unveils how a Chinese startup dared to leapfrog giants and redefine innovation’s rules. (View Highlight)
Liang Wenfeng (DeepSeek Founder): We never intended to be a disruptor; it just happened by accident. (View Highlight)
Very surprised. We didn’t expect pricing to be such a sensitive issue. We were simply following our own pace, calculating costs, and setting prices accordingly. Our principle is neither to sell at a loss nor to seek excessive profits. The current pricing allows for a modest profit margin above our costs. (View Highlight)
An Yong: From an outsider’s perspective, price cuts seem like a tactic to grab users—typical of internet-era competition.
Liang Wenfeng: Grabing users wasn’t our primary goal. We reduced prices because, first, while exploring next-generation model structures, our costs decreased; second, we believe that both AI and API services should be affordable and accessible to everyone. (View Highlight)
If the goal is to develop applications, adopting Llama’s structure to quickly launch a product is a reasonable choice. However, our goal is AGI (Artificial General Intelligence), which requires us to explore new model structures to achieve superior capabilities within limited resources. This is foundational research for scaling up. Beyond architecture, we’ve studied data curation and human-like reasoning—all reflected in our models. Also, Llama’s training efficiency and inference costs lag behind cutting-edge global standards by about two generations. (View Highlight)
First, there’s a gap in training efficiency. We estimate that China’s best models likely require twice the compute power to match top global models due to structural and training dynamics gaps. Data efficiency is also half as effective, meaning we need twice the data and compute for equivalent results. Combined, that’s four times the resources. Our goal is to continuously narrow these gaps. (View Highlight)
An Yong: Most Chinese firms pursue both models and applications. Why is DeepSeek focusing solely on research?
Liang Wenfeng: Because we believe the most important thing right now is to participate global innovation. For years, Chinese companies have been accustomed to leveraging technological innovations developed elsewhere and monetizing them through applications. But this isn’t sustainable. This time, our goal isn’t quick profits but advancing the technological frontier to drive ecosystem growth. (View Highlight)
An Yong: The prevailing belief from the internet and mobile internet eras is that the U.S. leads in innovation, while China excels at applications.
Liang Wenfeng: We believe that with economic development, China must gradually transition from being a beneficiary to a contributor, rather than continuing to ride on the coattails of others. Over the past 30 years of the IT revolution, we barely participated in core tech innovation. (View Highlight)
Innovation is undoubtedly costly, and our past tendency to adopt existing technologies was tied to China’s earlier developmental stage. But today, China’s economic scale and the profits of giants like ByteDance and Tencent are globally significant. What we lack isn’t capital but confidence and the ability to organize high-caliber talent for effective innovation. (View Highlight)
For three decades, we’ve emphasized profit over innovation. Innovation isn’t purely business-driven; it requires curiosity and creative ambition. We’re shackled by old habits, but this is a phase. (View Highlight)
Liang Wenfeng: In disruptive tech, closed-source moats are fleeting. Even OpenAI’s closed-source model can’t prevent others from catching up.
Therefore, our real moat lies in our team’s growth—accumulating know-how, fostering an innovative culture. Open-sourcing and publishing papers don’t result in significant losses. For technologists, being followed is rewarding. Open-source is cultural, not just comm (View Highlight)
We believe that China’s AI cannot remain a follower forever. Often, we say there’s a one- or two-year gap between Chinese and American AI, but the real gap is between originality and imitation. If this doesn’t change, China will always be a follower. Some explorations are unavoidable. (View Highlight)
An Yong: DeepSeek currently exudes an idealistic vibe reminiscent of OpenAI’s early days, and you’re open-source. Do you plan to transition to a closed-source model in the future, as OpenAI and Mistral have done?
Liang Wenfeng:We won’t go closed-source. We believe that establishing a robust technology ecosystem matters more. (View Highlight)
An Yong: Are there fundraising plans? Media reports suggest Huanfang【1】 aims to spin off DeepSeek for an IPO. Silicon Valley AI startups inevitably align with big players—will you follow?.
Liang Wenfeng:No short-term plans. Our challenge has never been money; it’s the embargo on high-end chips. (View Highlight)
We believe that the current stage is a period of technological innovation, not application explosion. In the long term, we aim to establish an ecosystem where the industry directly uses our technologies and outputs. Others develop B2B/B2C services on our models while we focus on foundational research. If a complete industry chain forms, there’s no need for us to develop applications ourselves. That said, if necessary, we are fully capable of doing so. However, research and innovation will always remain our top priority. (View Highlight)
The future world will likely be one of specialized division of labor. Foundational AI models require continuous innovation, and big companies have their limits—they may not always be the best fit for this role. (View Highlight)
Secrets don’t exist, but replication takes time and cost. NVIDIA GPUs have no hidden magic—yet catching up requires rebuilding teams and chasing their next-gen tech. That’s the real moat. (View Highlight)
An Yong: What’s your core philosophy when it comes to competition?
Liang Wenfeng: I focus on whether something elevates societal efficiency and whether we can find our strength in the industry value chain. As long as the ultimate goal boosts efficiency, it’s valid. Many aspects are just temporary phases—over-focusing on them will only lead to confusion.
V2 Model: Built Entirely by Homegrown Talent (View Highlight)
An Yong: ack Clark, former policy lead at OpenAI and co-founder of Anthropic, remarked that DeepSeek has hired “some of those inscrutable wizards” who built DeepSeek V2. What defines these people?
Liang Wenfeng: No “inscrutable wizards” here—just fresh graduates from top universities, PhD candidates (even fourth- or fifth-year interns), and young talents with a few years of experience. (View Highlight)
An Yong: How did the MLA innovation emerge? We heard that the idea initially stemmed from a young researcher’s personal interest.
Liang Wenfeng: After summarizing the key evolutionary patterns of the mainstream Attention architecture, he had a sudden inspiration to design an alternative. However, turning an idea into reality is a long journey. We assembled a team and spent months validating it. (View Highlight)
An Yong: This kind of organic creativity seems tied to your flat organizational structure. In Huanfang, you avoided top-down mandates. But for AGI—a high-uncertainty frontier—do you impose more management?
Liang Wenfeng: DeepSeek remains entirely bottom-up. We also do not preassign roles; natural division of labor emerges. Everyone brings unique experiences and ideas, and they don’t need to be pushed. When they encounter challenges, they naturally pull others in for discussions. However, once an idea shows potential, we do allocate resources from the top down (View Highlight)
An Yong: We’ve heard that DeepSeek operates with remarkable flexibility in allocating computing resources and personnel.
Liang Wenfeng: There are no limits on accessing compute resources or team members. If someone has an idea, they can tap into our training clusters anytime without approval. Additionally, since we don’t have rigid hierarchical structures or departmental barriers, people can collaborate freely as long as there’s mutual interest. (View Highlight)
An Yong: Such loose management relies on hiring intensely driven individuals. It’s said that DeepSeek excels at identifying exceptional talent based on non-traditional criteria.
Liang Wenfeng: Our hiring standards have always been based on passion and curiosity. Many of our team members have unique and interesting backgrounds. Their hunger for research far outweighs monetary concerns. (View Highlight)
An Yong: Transformer was born in Google’s AI Lab, and ChatGPT emerged from OpenAI. In your opinion, how do corporate AI labs differ from startups in fostering innovation?
Liang Wenfeng: Whether it’s Google’s labs, OpenAI, or even AI labs at Chinese tech giants, they all provide significant value. The fact that OpenAI eventually delivered breakthroughs was partly historical chance. (View Highlight)
Liang Wenfeng: I believe innovation is, first and foremost, a matter of belief. Why is Silicon Valley so innovative? Because they dare to try. When ChatGPT debuted, China lacked confidence in frontier research. From investors to major tech firms, many felt the gap was too wide and focused instead on applications. But innovation requires confidence, and young people tend to have more of it. (View Highlight)
An Yong: Unlike other AI companies that actively seek funding and media attention, DeepSeek remains relatively quiet. How do you ensure that DeepSeek becomes the top choice for people looking to work in AI?
Liang Wenfeng: Because we are tackling the hardest problems. The most attractive thing for top-tier talent is the opportunity to solve the world’s toughest challenges. In fact, top talent in China is often underestimated because hardcore innovation is rare, which means they rarely get recognized. We offer what they crave. (View Highlight)
An Yong: The recent OpenAI event did not feature GPT-5, leading many to believe that the industry’s technological curve is slowing down, and some have begun questioning Scaling Law. What’s your perspective?
Liang Wenfeng: We remain optimistic. The industry’s progress is still in line with expectations. OpenAI isn’t divine; they can’t lead forever. (View Highlight)
An Yong: How long do you think it will take to achieve AGI? Before V2, you released code/math models and switched from dense to MoE【2】 . What’s your roadmap?
Liang Wenfeng: It could take two years, five years, or ten years—but it will happen within our lifetime. As for our roadmap, there’s no consensus even within our company. However, we are placing our bets on three directions:
1. Mathematics and code, which serve as a natural testbed for AGI—much like Go, they are enclosed, verifiable systems where self-learning could lead to high intelligence.
2. Multimodality, where the AI engages with the real world to learn.
3. Natural language itself, which is fundamental to human-like intelligence. (View Highlight)
An Yong: What do you envision as the endgame for large AI models?
Liang Wenfeng: There will be specialized companies providing foundational models and services, forming a long value chain of specialized divisions. More players will emerge to meet society’s diverse needs on top of these foundations. (View Highlight)
An Yong: Where do you currently focus most of your energy?
Liang Wenfeng: My main focus is on researching the next generation of large models. There are still many unresolved challenges. (View Highlight)
An Yong: Many other AI startups insist on balancing both model development and applications, since technical leads aren’t permanent. Why is DeepSeek confident in focusing solely on research? Is it because your models still lag?
Liang Wenfeng: All strategies are products of the past generation and may not hold true in the future. Discussing AI’s future profitability using the commercial logic of the internet era is like comparing Tencent’s early days to General Electric or Coca-Cola—it’s essentially carving a boat to mark a sword’s position, an outdated approach. (View Highlight)
An Yong: Returning to original innovation: With the economy slowing and capital cooling, will this stifle groundbreaking R&D?
Liang Wenfeng: Not necessarily. The restructuring of China’s industrial landscape will increasingly rely on deep-tech innovation. As quick-profit opportunities vanish, more will embrace real innovation. (View Highlight)
An Yong: So you’re optimistic about this?
Liang Wenfeng: I grew up in the 1980s in a fifth-tier city in Guangdong. My father was a primary school teacher. In the 1990s, there were plenty of opportunities to make money in Guangdong. Many parents would come to our home and argue that studying was useless. But looking back now, perspectives have changed. Making money isn’t as easy as it used to be—not even driving a taxi is a viable option anymore. Within just one generation, things have shifted. (View Highlight)
Hardcore innovation will only increase in the future. It’s not widely understood now because society as a whole needs to learn from reality. When this society starts celebrating the success of deep-tech innovators, collective perceptions will change. We just need more real-world examples and time to allow that process to unfold. (View Highlight)
