DeepSeek is set to face more pressure and challenges in the future. The race towards a universal model has just begun, and who will ultimately win depends on continuous investment in funding and technological iteration.

· A headhunter responsible for sourcing high-end tech talent in the large model field told The Paper Technology that DeepSeek's hiring logic is not much different from that of other companies in the large model sector. The core label for talent is "young and high-potential," meaning those born around 1998 with preferably no more than five years of work experience—"smart, STEM-educated, young, and with little experience."

· In the eyes of industry insiders, DeepSeek is fortunate compared to other large model startups in China. It faces no funding pressure, doesn't need to prove itself to investors, and doesn't have to balance model iteration with product application optimization. However, as a commercial company, after significant investment, it will inevitably face the same pressures and challenges as other model companies.

Which company was the hottest in China's large model circle in 2024? Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. (hereinafter referred to as DeepSeek) is certainly a strong contender. As the initiator of last year's mid-year large model price war, DeepSeek first entered the public eye. By the end of the year and the beginning of the new year, it had released the open-source model DeepSeek-V3 and the inference model DeepSeek-R1, completely igniting the large model community. People were amazed by its cost-effective training (reportedly, DeepSeek-V3 only cost $5.576 million to train) and applauded its open-source model and public technical reports. The release of DeepSeek-R1 excited many scientists, developers, and users, who even considered DeepSeek a strong competitor to OpenAI's o1 inference model.

How did this low-profile company manage to create a high-performing large model with such low training costs? What did it do right to achieve its current popularity? What challenges will it face in the future as it continues to navigate the "model circle"?

Algorithm Innovation Drastically Reduces Computing Costs

"DeepSeek invested early, accumulated a lot, and has its own unique features in algorithms," said an executive from a prominent large model startup in China when discussing DeepSeek. He believes that the core advantage behind DeepSeek's popularity is its algorithmic innovation. "Chinese companies, due to a lack of computing power, are more focused on cost-saving compared to OpenAI."

According to DeepSeek's disclosed information on DeepSeek-R1, it extensively used reinforcement learning technology in the post-training phase, significantly enhancing the model's inference capabilities with minimal annotated data. Its performance in tasks like mathematics, coding, and natural language reasoning rivals that of OpenAI's o1 official version.

Can DeepSeek Stay Hot?

DeepSeek-R1 API Pricing

DeepSeek founder Liang Wenfeng has repeatedly emphasized that DeepSeek is committed to forging a differentiated technical path rather than replicating OpenAI's model. DeepSeek must find more efficient ways to train its models.

"They used a series of engineering techniques to optimize the model architecture, such as innovatively using model blending methods, with the ultimate goal of reducing costs through engineering to achieve profitability," a veteran in the tech industry told The Paper Technology.

From DeepSeek's disclosed information, it is evident that the company has made significant progress in MLA (Multi-head Latent Attention) and its self-developed DeepSeekMOE (Mixture-of-Experts) structure. These technological designs reduce training computational resources, making DeepSeek models more cost-effective and improving training efficiency. According to data from research firm Epoch AI, DeepSeek's latest models are highly efficient.

In terms of data, unlike OpenAI's "massive data feeding" approach, DeepSeek uses algorithms to summarize and categorize data, selectively processing it before feeding it into the large model. This improves training efficiency and reduces costs. The emergence of DeepSeek-V3 achieves a balance between high performance and low cost, offering new possibilities for large model development.

"In the future, we might not need ultra-large-scale GPU clusters," said OpenAI founding member Andrej Karpathy after DeepSeek's cost-effective models were released.

Liu Zhiyuan, a tenured associate professor at Tsinghua University's Department of Computer Science, told The Paper Technology that DeepSeek's breakout success demonstrates our competitive advantage—achieving more with less through the highly efficient use of limited resources. The release of R1 shows that the gap between China and the U.S. in AI capabilities has significantly narrowed. The Economist also noted in its latest report: "DeepSeek is changing the tech industry with its low-cost training and innovative model design."

Demis Hassabis, CEO and co-founder of Google DeepMind, stated that while it's not entirely clear how much DeepSeek relies on Western systems for training data and open-source models, the team's achievements are indeed impressive. He acknowledged China's strong engineering and scaling capabilities but also pointed out that the West still leads and needs to consider how to maintain its edge in frontier models.

Years of Focused Accumulation

DeepSeek's innovations are not the result of overnight success but years of "incubation" and long-term planning. Liang Wenfeng is also the founder of the top quantitative hedge fund, Magic Square Quantitative. DeepSeek is believed to have fully utilized the funds, data, and computing power accumulated by Magic Square Quantitative.

Liang Wenfeng graduated with a bachelor's and master's degree from Zhejiang University's Department of Information and Electronic Engineering. Since 2008, he has led a team to explore fully automated quantitative trading using machine learning. In 2015, Magic Square Quantitative was established, and the following year, it launched its first AI model, with the first trading position generated by deep learning executed. In 2018, AI was established as the main development direction. In 2020, Magic Square invested over 100 million yuan in the AI supercomputer "Firefly One," which occupies the space of a basketball court and boasts computing power equivalent to 40,000 personal computers. In 2021, Magic Square invested 1 billion yuan in building "Firefly Two," equipped with 10,000 A100 GPU chips. At that time, there were no more than five companies in China with over 10,000 GPUs, and aside from Magic Square Quantitative, the other four were major internet companies.

In July 2023, DeepSeek was officially established, entering the field of general artificial intelligence, and has never sought external funding.

"With relatively ample computing power and no funding pressure, DeepSeek has focused solely on models without developing products in the past few years, making it more focused and capable of breakthroughs in engineering and algorithms compared to other domestic large model companies," said the aforementioned executive from a domestic large model company.

Additionally, as the large model industry increasingly moves towards closed systems, with OpenAI being jokingly called "CloseAI," DeepSeek's open-source models and public technical reports have won widespread praise from developers, allowing its technical brand to quickly stand out in the global large model market.

A researcher told The Paper Technology that DeepSeek's openness is remarkable, and the open-source models V3 and R1 have raised the benchmark for open-source models in the market.

Success Proves the Power of Youth

"DeepSeek's success also shows the power of young people. Essentially, this generation of AI development requires young minds," a person from a model company told The Paper Technology.

Previously, Jack Clark, former policy director at OpenAI and co-founder of Anthropic, believed that DeepSeek hired "a group of enigmatic geniuses." In response, Liang Wenfeng stated in an interview with self-media that there are no enigmatic geniuses, just graduates from top domestic universities, fourth and fifth-year PhD interns, and some young people who graduated only a few years ago.

From current media reports, it is clear that the DeepSeek team's most notable characteristics are their elite educational backgrounds and youth. Even at the team leader level, most are under 35 years old. With a team of fewer than 140 people, engineers and developers mostly come from top domestic universities like Tsinghua University, Peking University, Sun Yat-sen University, and Beijing University of Posts and Telecommunications, with relatively short work experience.

A headhunter responsible for sourcing high-end tech talent in the large model field told The Paper Technology that DeepSeek's hiring logic is not much different from that of other companies in the large model sector. The core label for talent is "young and high-potential," meaning those born around 1998 with preferably no more than five years of work experience—"smart, STEM-educated, young, and with little experience."

However, the headhunter also noted that large model startups are essentially still startups. It's not that they don't want to hire top AI talent from overseas, but the reality is that few top AI talents are willing to return.

An anonymous DeepSeek employee told The Paper Technology that the company has a flat management structure and a good atmosphere for free communication. Liang Wenfeng's whereabouts are often unpredictable, and most communication with him is done online.

The employee previously worked on large model technology development at a major domestic company but felt like a cog in a machine, unable to create value, and ultimately chose to join DeepSeek. In his view, DeepSeek is currently more focused on underlying model technology.

DeepSeek's work atmosphere is entirely bottom-up, with natural division of labor. There are no limits on the allocation of computing power and personnel. "Everyone brings their own ideas; there's no need to push. During exploration, when someone encounters a problem, they naturally gather people to discuss it," Liang Wenfeng said in a previous interview.

"It's Too Early to Say China's AI Has Surpassed the U.S."

U.S. business media Business Insider analyzed that the newly released R1 shows that China can compete with some of the top AI models in the industry and keep pace with the cutting-edge developments in Silicon Valley. Additionally, open-sourcing such advanced AI could pose a challenge to companies trying to profit from selling technology.

However, it might be too early to declare that "China's AI has surpassed the U.S." Liu Zhiyuan publicly stated that we need to be cautious about the shift from extreme pessimism to extreme optimism, thinking that we have completely surpassed and are far ahead—"we are far from it." Liu Zhiyuan believes that AGI technology is still rapidly evolving, and the future development path is still unclear. China is still in the catching-up phase, though it's no longer out of reach, it's still just keeping pace. "It's relatively easy to follow quickly on a path others have already explored. The bigger challenge is how to forge a new path in the fog."

"Things are moving too fast now; everyone is in a hurry and didn't realize DeepSeek would come out on top," someone close to DeepSeek lamented to The Paper Technology. The industry is changing too quickly, and it's impossible to predict what the next step will be. We can only wait and see what happens in the next Q3.

Demis Hassabis acknowledged China's strong engineering and scaling capabilities but also pointed out that the West still leads and needs to consider how to maintain its edge in frontier models.

Although Liang Wenfeng previously stated that DeepSeek only focuses on models and not products, as a commercial company, it's almost impossible to only focus on models without developing products. On January 15, the official DeepSeek app was released. Someone close to DeepSeek told The Paper Technology that commercialization is already on DeepSeek's agenda.

In the eyes of industry insiders, DeepSeek is fortunate compared to other large model startups in China, with no funding pressure, no need to prove itself to investors, and no need to balance model iteration with product application optimization. However, as a commercial company, after significant investment, it will inevitably face the same pressures and challenges as other model companies. "This breakout success has been a successful marketing move for DeepSeek on the eve of commercialization, but after true commercialization, it will need to face market scrutiny. Whether it can continue to ride the waves remains uncertain," said the aforementioned model company insider.

What is certain is that DeepSeek will face more pressure and challenges in the future. The race towards a universal model has just begun, and who will ultimately win depends on continuous investment in funding and technological iteration. However, industry insiders also believe that "for the large model industry, having a company like DeepSeek with genuine technical strength is a good thing."

DeepSeek is set to face more pressure and challenges in the future. The race towards a universal model has just begun, and who will ultimately win depends on continuous investment in funding and technological iteration.

· A headhunter responsible for sourcing high-end tech talent in the large model field told The Paper Technology that DeepSeek's hiring logic is not much different from that of other companies in the large model sector. The core label for talent is "young and high-potential," meaning those born around 1998 with preferably no more than five years of work experience—"smart, STEM-educated, young, and with little experience."

· In the eyes of industry insiders, DeepSeek is fortunate compared to other large model startups in China. It faces no funding pressure, doesn't need to prove itself to investors, and doesn't have to balance model iteration with product application optimization. However, as a commercial company, after significant investment, it will inevitably face the same pressures and challenges as other model companies.

Which company was the hottest in China's large model circle in 2024? Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd. (hereinafter referred to as DeepSeek) is certainly a strong contender. As the initiator of last year's mid-year large model price war, DeepSeek first entered the public eye. By the end of the year and the beginning of the new year, it had released the open-source model DeepSeek-V3 and the inference model DeepSeek-R1, completely igniting the large model community. People were amazed by its cost-effective training (reportedly, DeepSeek-V3 only cost $5.576 million to train) and applauded its open-source model and public technical reports. The release of DeepSeek-R1 excited many scientists, developers, and users, who even considered DeepSeek a strong competitor to OpenAI's o1 inference model.

How did this low-profile company manage to create a high-performing large model with such low training costs? What did it do right to achieve its current popularity? What challenges will it face in the future as it continues to navigate the "model circle"?

Algorithm Innovation Drastically Reduces Computing Costs

"DeepSeek invested early, accumulated a lot, and has its own unique features in algorithms," said an executive from a prominent large model startup in China when discussing DeepSeek. He believes that the core advantage behind DeepSeek's popularity is its algorithmic innovation. "Chinese companies, due to a lack of computing power, are more focused on cost-saving compared to OpenAI."

According to DeepSeek's disclosed information on DeepSeek-R1, it extensively used reinforcement learning technology in the post-training phase, significantly enhancing the model's inference capabilities with minimal annotated data. Its performance in tasks like mathematics, coding, and natural language reasoning rivals that of OpenAI's o1 official version.

Can DeepSeek Stay Hot?

DeepSeek-R1 API Pricing

DeepSeek founder Liang Wenfeng has repeatedly emphasized that DeepSeek is committed to forging a differentiated technical path rather than replicating OpenAI's model. DeepSeek must find more efficient ways to train its models.

"They used a series of engineering techniques to optimize the model architecture, such as innovatively using model blending methods, with the ultimate goal of reducing costs through engineering to achieve profitability," a veteran in the tech industry told The Paper Technology.

From DeepSeek's disclosed information, it is evident that the company has made significant progress in MLA (Multi-head Latent Attention) and its self-developed DeepSeekMOE (Mixture-of-Experts) structure. These technological designs reduce training computational resources, making DeepSeek models more cost-effective and improving training efficiency. According to data from research firm Epoch AI, DeepSeek's latest models are highly efficient.

In terms of data, unlike OpenAI's "massive data feeding" approach, DeepSeek uses algorithms to summarize and categorize data, selectively processing it before feeding it into the large model. This improves training efficiency and reduces costs. The emergence of DeepSeek-V3 achieves a balance between high performance and low cost, offering new possibilities for large model development.

"In the future, we might not need ultra-large-scale GPU clusters," said OpenAI founding member Andrej Karpathy after DeepSeek's cost-effective models were released.

Liu Zhiyuan, a tenured associate professor at Tsinghua University's Department of Computer Science, told The Paper Technology that DeepSeek's breakout success demonstrates our competitive advantage—achieving more with less through the highly efficient use of limited resources. The release of R1 shows that the gap between China and the U.S. in AI capabilities has significantly narrowed. The Economist also noted in its latest report: "DeepSeek is changing the tech industry with its low-cost training and innovative model design."

Demis Hassabis, CEO and co-founder of Google DeepMind, stated that while it's not entirely clear how much DeepSeek relies on Western systems for training data and open-source models, the team's achievements are indeed impressive. He acknowledged China's strong engineering and scaling capabilities but also pointed out that the West still leads and needs to consider how to maintain its edge in frontier models.

Years of Focused Accumulation

DeepSeek's innovations are not the result of overnight success but years of "incubation" and long-term planning. Liang Wenfeng is also the founder of the top quantitative hedge fund, Magic Square Quantitative. DeepSeek is believed to have fully utilized the funds, data, and computing power accumulated by Magic Square Quantitative.

Liang Wenfeng graduated with a bachelor's and master's degree from Zhejiang University's Department of Information and Electronic Engineering. Since 2008, he has led a team to explore fully automated quantitative trading using machine learning. In 2015, Magic Square Quantitative was established, and the following year, it launched its first AI model, with the first trading position generated by deep learning executed. In 2018, AI was established as the main development direction. In 2020, Magic Square invested over 100 million yuan in the AI supercomputer "Firefly One," which occupies the space of a basketball court and boasts computing power equivalent to 40,000 personal computers. In 2021, Magic Square invested 1 billion yuan in building "Firefly Two," equipped with 10,000 A100 GPU chips. At that time, there were no more than five companies in China with over 10,000 GPUs, and aside from Magic Square Quantitative, the other four were major internet companies.

In July 2023, DeepSeek was officially established, entering the field of general artificial intelligence, and has never sought external funding.

"With relatively ample computing power and no funding pressure, DeepSeek has focused solely on models without developing products in the past few years, making it more focused and capable of breakthroughs in engineering and algorithms compared to other domestic large model companies," said the aforementioned executive from a domestic large model company.

Additionally, as the large model industry increasingly moves towards closed systems, with OpenAI being jokingly called "CloseAI," DeepSeek's open-source models and public technical reports have won widespread praise from developers, allowing its technical brand to quickly stand out in the global large model market.

A researcher told The Paper Technology that DeepSeek's openness is remarkable, and the open-source models V3 and R1 have raised the benchmark for open-source models in the market.

Success Proves the Power of Youth

"DeepSeek's success also shows the power of young people. Essentially, this generation of AI development requires young minds," a person from a model company told The Paper Technology.

Previously, Jack Clark, former policy director at OpenAI and co-founder of Anthropic, believed that DeepSeek hired "a group of enigmatic geniuses." In response, Liang Wenfeng stated in an interview with self-media that there are no enigmatic geniuses, just graduates from top domestic universities, fourth and fifth-year PhD interns, and some young people who graduated only a few years ago.

From current media reports, it is clear that the DeepSeek team's most notable characteristics are their elite educational backgrounds and youth. Even at the team leader level, most are under 35 years old. With a team of fewer than 140 people, engineers and developers mostly come from top domestic universities like Tsinghua University, Peking University, Sun Yat-sen University, and Beijing University of Posts and Telecommunications, with relatively short work experience.

A headhunter responsible for sourcing high-end tech talent in the large model field told The Paper Technology that DeepSeek's hiring logic is not much different from that of other companies in the large model sector. The core label for talent is "young and high-potential," meaning those born around 1998 with preferably no more than five years of work experience—"smart, STEM-educated, young, and with little experience."

However, the headhunter also noted that large model startups are essentially still startups. It's not that they don't want to hire top AI talent from overseas, but the reality is that few top AI talents are willing to return.

An anonymous DeepSeek employee told The Paper Technology that the company has a flat management structure and a good atmosphere for free communication. Liang Wenfeng's whereabouts are often unpredictable, and most communication with him is done online.

The employee previously worked on large model technology development at a major domestic company but felt like a cog in a machine, unable to create value, and ultimately chose to join DeepSeek. In his view, DeepSeek is currently more focused on underlying model technology.

DeepSeek's work atmosphere is entirely bottom-up, with natural division of labor. There are no limits on the allocation of computing power and personnel. "Everyone brings their own ideas; there's no need to push. During exploration, when someone encounters a problem, they naturally gather people to discuss it," Liang Wenfeng said in a previous interview.

"It's Too Early to Say China's AI Has Surpassed the U.S."

U.S. business media Business Insider analyzed that the newly released R1 shows that China can compete with some of the top AI models in the industry and keep pace with the cutting-edge developments in Silicon Valley. Additionally, open-sourcing such advanced AI could pose a challenge to companies trying to profit from selling technology.

However, it might be too early to declare that "China's AI has surpassed the U.S." Liu Zhiyuan publicly stated that we need to be cautious about the shift from extreme pessimism to extreme optimism, thinking that we have completely surpassed and are far ahead—"we are far from it." Liu Zhiyuan believes that AGI technology is still rapidly evolving, and the future development path is still unclear. China is still in the catching-up phase, though it's no longer out of reach, it's still just keeping pace. "It's relatively easy to follow quickly on a path others have already explored. The bigger challenge is how to forge a new path in the fog."

"Things are moving too fast now; everyone is in a hurry and didn't realize DeepSeek would come out on top," someone close to DeepSeek lamented to The Paper Technology. The industry is changing too quickly, and it's impossible to predict what the next step will be. We can only wait and see what happens in the next Q3.

Demis Hassabis acknowledged China's strong engineering and scaling capabilities but also pointed out that the West still leads and needs to consider how to maintain its edge in frontier models.

Although Liang Wenfeng previously stated that DeepSeek only focuses on models and not products, as a commercial company, it's almost impossible to only focus on models without developing products. On January 15, the official DeepSeek app was released. Someone close to DeepSeek told The Paper Technology that commercialization is already on DeepSeek's agenda.

In the eyes of industry insiders, DeepSeek is fortunate compared to other large model startups in China, with no funding pressure, no need to prove itself to investors, and no need to balance model iteration with product application optimization. However, as a commercial company, after significant investment, it will inevitably face the same pressures and challenges as other model companies. "This breakout success has been a successful marketing move for DeepSeek on the eve of commercialization, but after true commercialization, it will need to face market scrutiny. Whether it can continue to ride the waves remains uncertain," said the aforementioned model company insider.

What is certain is that DeepSeek will face more pressure and challenges in the future. The race towards a universal model has just begun, and who will ultimately win depends on continuous investment in funding and technological iteration. However, industry insiders also believe that "for the large model industry, having a company like DeepSeek with genuine technical strength is a good thing."