Essays & notes
I do research at OpenAI. About me ↗
The questions behind my PPO notes: why rewards cannot simply backpropagate through sampled tokens, what advantage and GAE estimate, why on-policy rollouts still need probability ratios, and what clipping actually changes. Includes numerical examples, runnable gradient checks, and a focused reading path.
My recommended systems companion for becoming an LLM engineer: a guide to AI Agents in Depth, the model–harness boundary, context and tool design, outcome-based evaluation, and a practical reading path.
Why I strongly recommend Shiyu Zhao’s Mathematical Foundations of Reinforcement Learning, the questions that clarified RL for me, and a practical learning path from Bellman equations to LLM post-training.
How BM25, dense retrievers, SPLADE, and ColBERT represent and match evidence. Worked examples of MaxSim, compression, hybrid fusion, and the difference between scoring quality and retrieval quality.
How I think about noisy labels, data curation, and evaluation that deserves to be trusted, with lessons from MIT and three reproducible teaching notebooks.
The knowledge map behind my preparation for an OpenAI research engineering role: seven connected areas, concrete failure cases, diagnostic exercises, and a method for turning study into usable skills.
从 Facebook 早期校园网络、增长团队、S-1 和 Alex Schultz/Andy Johns 等材料出发，拆解可迁移的增长原则：留存优先、magic moment、边际用户、增长系统，以及它们对 ChatGPT 类产品的启发。
真正改变人生的，往往不是一次宏大顿悟，而是把自己从惯性里叫醒的几个小习惯：先行动、追问为什么、走出舒适区、慢下来，并把“但是”换成“如何”。
更快的学习通常不是靠更用力，而是靠更少的认知噪音、更好的理解压缩、更主动的提问与输出，以及把休息当成学习回路的一部分。
The useful form of visualization is not fantasy about the trophy but rehearsal of the process. Outcome fantasies can drain motivation; process rehearsal can reduce surprise, strengthen neural patterns, and make hard moments feel familiar enough to execute.