Beijing – DeepSeek, a rising force in the global AI community, has been the subject of intense scrutiny following the unexpected fervor surrounding its DeepSeek-R1 model. However, high-quality information about the company remains scarce. To address this, Li Guangmi, founder and CEO of Shixiang, organized a closed-door discussion on January 26, 2025, bringing together dozens of leading AI researchers, investors, and industry practitioners to explore DeepSeek’s technical details, organizational culture, and its short, medium, and long-term impact.
This discussion sought to unveil a corner of the mysterious Eastern force amidst limited information. It is important to note that this discussion was a non-governmental technical exchange and does not represent the views of any specific individual or institution. As renowned Silicon Valley venture capitalist Marc Andreessen commented on DeepSeek-R1: As open source, a profound gift to the world. Therefore, the participants in this discussion also learned from DeepSeek, and in the spirit of open source, made public the collective thinking of the closed-door meeting. The following is a summary of the key points of the discussion. This summary was compiled by the Shixiang team, with minor edits by the author.
The Mystery of DeepSeek
The core takeaway from the meeting is that DeepSeek’s driving force lies in its unwavering commitment to pushing the boundaries of artificial intelligence.
-
Focus on Intelligence, Not Just Commercialization: Participants emphasized that founder and CEO Liang Wenfeng is the core of DeepSeek, a technically proficient leader driven by the pursuit of intelligence rather than immediate commercial gains. He is not in the same mold as figures like Sam Altman, known for their business acumen.
-
Efficient Resource Utilization: DeepSeek has earned a strong reputation for being the first to release reproductions of MoE and o1, capitalizing on early timing. However, the potential for further advancement remains significant. The challenge lies in maximizing limited resources, focusing them on the most promising areas. The team’s research capabilities and culture are commendable, and with access to more resources, such as 100,000 to 200,000 GPUs, they could achieve even greater breakthroughs.
-
Rapid Advancements in Long Context Capabilities: DeepSeek has demonstrated impressive progress in long context capabilities between the preview and official release of its models. They have achieved a 10K long context using conventional methods.
-
Resource Speculation: While Scale.ai’s CEO suggested DeepSeek possesses 50,000 GPUs, the actual number is likely lower. Public information indicates that DeepSeek has approximately 10,000 older A100 GPUs and potentially 3,000 H800 GPUs acquired before export restrictions.
Conclusion
The closed-door discussion on DeepSeek highlights the company’s unique position in the AI landscape. Its focus on fundamental intelligence, efficient resource utilization, and rapid technical advancements have propelled it to the forefront of the field. While challenges remain, particularly in securing sufficient resources, DeepSeek’s commitment to open-source principles and its technically driven leadership suggest a promising future for the company and the broader AI community. The insights from this meeting underscore the importance of vision and dedication in driving innovation, even in the face of limited resources.
References
- Zhang, X. (2024, January 29). 一场关于DeepSeek的高质量闭门会：比技术更重要的是愿景 [A High-Quality Closed-Door Meeting on DeepSeek: Vision is More Important Than Technology]. Tencent Technology. [Original article link]
Views: 0