Shengyu Liu (刘胜与)
Shengyu Liu is a machine learning systems engineer at DeepSeek. According to his homepage, he has worked at DeepSeek-AI in Hangzhou since April 2025, focusing on machine learning systems (MLSys) and kernel design and optimization.
Background and notable work
The following is based on the account on his homepage.
- Education: he entered the Turing Class at the School of EECS, Peking University in September 2021 (September 2021 to July 2025), advised by Associate Professor Xin Jin; his research areas are machine learning systems (MLSys), machine learning compilers, and distributed systems.
- Supercomputing: as team leader of the Peking University Supercomputing Team, he led the team to first place at the 10th ASC student supercomputer competition and second place at SC23.
- Open source and research: he built swiftLLM, a research-oriented inference system of roughly 2,000 lines of code that claims vLLM-comparable performance; papers listed on his homepage include DistServe (OSDI 2024), LoongServe (SOSP) and RLHFuse (NSDI 2025) in LLM serving systems.
- Work at DeepSeek: his homepage records, in the team’s voice, the releases of DeepSeek-V3.2, the DeepSeek-V4 preview and DeepSeek-V4.1-Flash, as well as DeepSelect, an efficient top-k implementation for DeepSeek Sparse Attention; it also describes FlashMLA, an efficient MLA decoding kernel, as delivering 5–15% speedups on compute-intensive tasks and up to 660 TFlops on an NVIDIA H800 SXM5 GPU.
Mentioned in the show
Shengyu Liu does not appear in the episodes quoted on this page; the following are the hosts’ relays and comments. In Weekly #003, the hosts discussed his article twice:
- In the chapter on using GPT-6 Astra to generate data and train small models, while discussing that AI can already write kernels, Yang Pan says “the article by the DeepSeek guy says GPT wrote it” (see the line in the Chinese transcript); Orange adds that “people inside DeepSeek are already worried about losing their jobs… like a soldier in a battle who has nothing left to do” (see that line).
- In the chapter on enterprise data and log boundaries, Guizang says “as for Shengyu Liu’s article we mentioned earlier, it spread widely, including on Twitter” (see that line); Yang Pan attributes its virality to the combined topics of a “DeepSeek prodigy” and “criticizing Anthropic” (see that line).
The hosts’ relays and comments are not statements by Shengyu Liu, and the article’s claims should be read in the original.
Frequently asked questions
Who is Shengyu Liu?
A machine learning systems engineer at DeepSeek-AI who has worked on MLSys and kernel design and optimization since April 2025; earlier he studied in Peking University’s Turing Class and led the PKU supercomputing team. See his homepage. (Editorial addition, based on his own homepage.)
What is the “article by Shengyu Liu” mentioned on the show?
Per the Weekly #003 hosts, it is an article that spread widely that week on Zhihu, Twitter and other platforms, with topics including that AI can already write production-grade kernels; the hosts attribute its virality to his “DeepSeek employee” status combined with “criticizing Anthropic”. This page relays the hosts’ account from the transcript only, and the article’s claims should be read in the original; see the relevant chapter in the Chinese transcript. (Editorial addition, based on the transcript.)