How many more “Guoda Ya” does DeepSeek have? After analyzing 27 papers, we found a group of “polygon warriors”
2026-09-04 08:00Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
5mentions
SummaryAI generated
After approximately 60 to 70 members of the ByteDance Seed team left, DeepSeek’s R&D team managed to maintain its core R&D network. The retention rate of initial members reached 87%, and currently, 21 highly skilled researchers continue to support technical iteration. Although key authors such as Guo Daya have left, the team still has 328 R&D authors, among whom 79 specialize in three or more technical areas, forming an efficient collaboration mechanism of “battalions + teams”. The team’s papers focus on underlying issues such as system efficiency, long context caching costs, and the stability of large model training, rather than simply competing for rankings. Organically, the team adopts a thin hierarchical structure with natural division of labor, emphasizing verification before integration, and without explicit overtime assessments.
The initial retention rate of DeepSeek’s R&D team reached 87%. Although core authors such as Guo Daya have left, 21 highly skilled developers continue to support technological iteration. Analysis shows that the team consists of 328 R&D authors, 79 of whom specialize in three or more technical areas, forming an efficient collaboration mechanism of “battalion + teams”. Their papers focus on underlying issues such as system efficiency, long context caching costs, and stability of large model training, rather than simply competing for rankings. Organically, the team adopts a thin hierarchical structure, emphasizing verification before integration, with no clear overtime assessment. Although about 60 to 70 members of ByteDance’s Seed team have left, DeepSeek’s core R&D network remains stable, with a talent density significantly higher than expected.