Vincent Zhong — Reading

Reading

Books

EducatedTara Westover2018Robinson CrusoeDaniel Defoe1719The Kite RunnerKhaled Hosseini2003

Essays

AI research is a max-performance domainJason WeiYou are judged by your best five works, not your average, so one exceptional skill beats being well-rounded.2025On the Slow Death of ScalingSara HookerThe link between compute and performance is weakening, and scaling alone misses where progress comes from next.2025What Is Taste?Yoonho LeeTaste is a fast, wordless filter you pick up through practice and by working near good people.2026The “it” in AI models is the datasetJames BetkerTrain long enough and every architecture lands in the same place. The model is the data.2023An Opinionated Guide to ML ResearchJohn SchulmanPicking the right problem matters more than raw skill. How to pick, keep moving, and keep growing.202095%-ile isn't that goodDan LuuMost people who do a thing never try to fix their obvious mistakes, so modest deliberate practice puts you near the top.2020

Source code

DeepGEMMDeepSeekFor kernels. Read the source itself, not the README.DeepEPDeepSeekFor communication, and what large-scale MoE parallelism really means. Read the kernels.DeepInferDeepSeekA state-of-the-art model runner, built for DeepSeek's own models.

Blogs

Yifan YangGPU inference at Anthropic, previously NVIDIAShort, exact posts on CuTe layouts, TMA, tensor core swizzles, PDL and Blackwell MMA descriptors.Changyi YangRL infrastructureWhy MLA and MTP fight each other, delta weight sync for RL, and how SGLang loads weights and serves LoRA.