陈德里
Senior Researcher at DeepSeek AI. Building next-generation large language models.
Core contributor to DeepSeek-V1, V2, V3, V4, DeepSeek-R1 (Nature Cover), DeepSeek-Coder, DeepSeek-MoE architecture, etc.
B.S. & M.S. from Peking University · Previously Tencent WeChat AI.
All opinions are my own. | INTP-T | 人心惟危，道心惟微 ｜ #AGIforEveryone
Four survey papers produced by the AutoResearch pipeline — fully autonomous literature review, experiment design, and paper writing powered by LLM agents. The latest (Paper #4 on self-play) is validated by an original 285B-parameter GRPO reinforcement-learning experiment and reached an 8.6/10 peer-review score.
The story of AutoResearch Paper #4 — a self-play survey the framework pushed from a rough draft to an 8.6/10 strong-accept across 16 review rounds, backed by an original 285B-parameter GRPO experiment, five-persona review, and a final round of theory hardening. 一篇自我博弈综述如何被自主框架推到 8.6。EN/CN bilingual.
A major update to the AutoResearch project — new Paper #3 on long-horizon decision-making, Papers #1 & #2 republished with 141 new references, and the full paper-writing skill group open-sourced. EN/CN bilingual.
A candid reflection on how Code Agents broke the classic INTP overthink-never-execute cycle, the new bottlenecks they reveal, and 9 survival rules for using them without losing your mind. EN/CN bilingual.
