Around noon on January 20, DeepSeek’s founder, Liang Wenfeng (梁文锋), had the company release the DeepSeek-R1 model, whose performance can match the official version of OpenAI’s o1. Immediately afterwards, he left for Zhongnanhai to attend a symposium on Premier Li Qiang’s government work report—he was the only representative from an LLM company present at the symposium. Other attendees are:
Zhang Hui (张辉) – Executive Director, Bank of China
Ren Shaobo (任少波) – Party Secretary, Zhejiang University (major hub for STEM education and a key incubator of tech talent/business innovation in China)
刘珺 (Liu Jun) – Deputy Party Secretary, Industrial and Commercial Bank of China (ICBC)
Zou Jingyuan (邹敬园) – World Champion in Men's Artistic Gymnastics
Wei Hongxing (魏洪兴) – Robotics Expert
Chen Xuedong (陈学东) – Robotics Expert
Chen Hongyan (陈红彦) – Deputy Director, National Library
Du Bin (杜斌) – Deputy Director, Peking Union Medical College Hospital
As the founder of DeepSeek, Liang Wenfeng is also well-known in China’s quantitative investment sector. Born in the 1980s in a fifth-tier city in Guangdong Province, he completed both his bachelor’s and master’s degrees at Zhejiang University, specializing in software engineering and AI in a 211 program. Liang is noted for his low-profile demeanor and technological idealism, rarely grants media interviews, and has few publicly available photos; however, his accomplishments in quantitative investment and AI are undeniable.
He began his work in quantitative hedging in 2008 and established High Flyer in 2015, which quickly rose to become one of China’s “Big Four” quantitative private funds, at one point managing assets exceeding 100 billion RMB. Under Liang’s leadership, High Flyer introduced its first AI-driven strategies in 2016 and transitioned its entire investment strategy to AI by 2017, marking it as an industry pioneer in quantitative innovation.
In July 2023, Liang founded DeepSeek to focus on R&D in AI large models. DeepSeek attracted industry attention for its innovative model architecture and cost-effectiveness, earning it the nickname “the Pinduoduo of AI.” Liang and his team proposed the MLA (Multi-head Latent Attention) architecture and the DeepSeekMoESparse structure, significantly reducing computational demands and memory usage, thus greatly lowering costs. Liang emphasizes that AI technology should serve the public, advocating for inclusive AI and open-source culture. He believes open sourcing is not only a technical decision but also a cultural one that fosters technological advancement. He remains optimistic about AGI (Artificial General Intelligence), expecting it to be realized within our lifetime.
Notably, as early as last December, OpenAI suggested that DeepSeek was only about six months behind them in research—a remark that, in retrospect, seems well-founded. Huanfang recently released its R1 inference model, which performs on par with OpenAI’s o1 across multiple benchmarks. It’s worth noting that o1’s preview version appeared just four months earlier, with its official release only one month ago.
In my view, whichever team in China can rapidly develop its own “o1” will gain a significant lead in the increasingly competitive field of top-tier domestic AI labs. Huanfang has already secured its seat at the table, while overseas, OpenAI, Google, and Anthropic continue to pioneer the cutting edge. I anticipate at least one or two more Chinese teams will soon introduce their own “o1.”
Another Chinese large model, Minimax-01, is also underestimated by many, despite making notable progress in critical areas such as communication bottlenecks, large-scale MoE (Mixture of Experts) training, and linear attention.
The gap between Chinese and U.S. models is visibly narrowing. In recent years, China has developed a tiered R&D ecosystem for AI: Huanfang focuses on frontier research and has led with the R1 inference model, while Jiaeyue Xingchen, Zhipu, and others have each made strides in multi-modal and B2B solutions. Crucially, even with limited resources, domestic teams can quickly replicate overseas achievements in MoE, multi-modal, and “o1”-level inference models at a fraction of the cost, thanks to having earlier international efforts as references.
That said, catching up in frontier research still requires huge computing resources. For example, Huanfang’s previously disclosed $5.5 million training cost is merely the tip of the iceberg. Upgrading to larger-scale models—like V4—demands far more computing power, and Huanfang currently faces shortages of GPUs. Moreover, from o1 to o3 to o4 or o5, the iterative growth of inference models could demand compute resources on par with, or even exceeding, those for pre-training. There seems to be no immediate ceiling.
It’s also worth mentioning that overseas pre-training has hit a bottleneck, and those teams have only recently started shifting toward post-training and testing paradigms. This has given domestic efforts a valuable window of opportunity, saving significant trial-and-error costs. And, simply restricting hardware exports may not fully halt the progress of Chinese models, as some teams have already secured resources and worked through engineering challenges like communication bottlenecks.
Given the relatively clear scaling roadmap, the increasingly evident commercial prospects, and ongoing resource infusion, I remain optimistic about the future of Chinese models. The primary uncertainty lies in potential export restrictions from the United States, which could make the next year or two particularly crucial.
CCTV News (Xinwen Lianbo) of Liang’s attendance to Premier Li’s Symposium:
On the afternoon of January 20, Li Qiang, member of the Standing Committee of the Political Bureau of the CPC Central Committee and Premier of the State Council, presided over a symposium with experts, entrepreneurs, and representatives from the fields of education, science, culture, health, and sports. The meeting aimed to gather opinions and suggestions on the draft of the Government Work Report (Consultation Draft).
During the symposium, Zhang Hui, Ren Shaobo, Liu Jun, Liang Wenfeng, Wei Hongxing, Chen Xuedong, Chen Hongyan, Du Bin, and Zou Jingyuan delivered speeches. Participants recognized that, despite the complex and severe challenges faced last year, China had strengthened macroeconomic regulation, introduced innovative policy measures, achieved rapid economic growth, and made progress in various fields—remarkable achievements that were hard-earned. They also shared their insights on addressing current development challenges and improving this year’s government work.
After listening to the speeches, Li Qiang noted that under the strong leadership of the CPC Central Committee with Comrade Xi Jinping at its core, the nation made concerted efforts to overcome difficulties last year, successfully achieving major development goals. The foundation for "stability" has become more solid, and the momentum for "progress" continues to grow. At the same time, challenges remain in China’s development, requiring a clear understanding of the situation and the ability to turn crises into opportunities. Li emphasized that China’s comprehensive advantages in systems, markets, industries, and talent remain intact, and the long-term positive trajectory of its economy has not and will not change. As long as the nation steadfastly focuses on its own priorities, stable and healthy economic growth is achievable.
Li Qiang stressed the need to adopt a more proactive fiscal policy and a moderately loose monetary policy to provide robust policy support for addressing challenges and promoting development. He called for deeper reform and greater openness to invigorate various market entities. Technology and innovation should drive the transformation of growth drivers, fostering new growth points for the economy. Greater efforts should also be made to improve people's livelihoods, boost income, reduce burdens, and enhance quality of life.
Wu Zhenglong attended the symposium.
The draft Government Work Report is currently open for feedback from all regions and departments.
Geotechpolitics and big tech role in warefare around the world beyond cyber security and intelligence misuse of Ai https://music.youtube.com/watch?v=mqFLXayD6e8&si=bfmVy_u1Eyx4zz5u
https://www.brighteon.com/1cab80fc-8c76-4e4f-be9c-a2a6e1adbcf6 can be correction big tech ungooglable recommendations unlike garbage google algorithms including spy companies connected executive and legislative branch of government for funding wars on false prophet google teachings to users of google and other us big tech companies using the private profit gain to fund war with out tech law beyond spying including public private misuse of authority administration even approval and assassinate from tech industry for their role in warfare of geotechpolitics including real life migration/refugee crisis in body and Mind manipulation by garbage algorithms recommendation especially online woke Ai recommendations and censorship of non woke ideology beyond user self consciousness and object/device consciousness to user migration of other platforms especially local platforms