Archive
All writing
| Date | Title |
|---|---|
| Dec 14, 2025 | How we inference (2025) After a year working at an inference engine start up, I have witnessed the great evolution of inference optimization in 2025. This blog concludes several famous breakthroughs. |
| Nov 28, 2025 | Daily Vibe Coding Share Recently, I have been invited to share some of my experiences with vibe coding. AI has profoundly changed the way I think and how I work. It is a great pleasure to witness such a transformation whe… |
| Mar 9, 2025 | RAG & Agent Share Last week I shared some my exp of RAG and Agent for all members in the company. Here's a desensitized copy of my notes. Though something I cannot share here, but these are enough for starters to le… |
| Feb 14, 2025 | 5 Years invention Patent My first-author invention patent just got approved after 5 years! Though technology has evolved since then, I'm still deeply thankful for this milestone. Looking back at those nights spent writing … |
| Feb 6, 2025 | DS-R1 & GRPO Code DSR1 paper https://arxiv.org/pdf/2501.12948 As we have read the GRPO algo in previous blog https://lihaorui.com/2025/02/05/deepseek-math-reading/ here we continue read deepseek r1's training pipeline. |
| Feb 5, 2025 | Deepseek Math Reading [](https://yaih.dawn.ee/image/SPdl) [](https://yaih.dawn.ee/image/SJKt) [](https://yaih.dawn.ee/image/Sqw7) **Proximal Policy Optimization (PPO)** is an actor-critic RL algorithm widely used in the… |
| Jan 25, 2025 | Prompt Attack Defense This blog is a note of Google's prompt attack and defense presentation at 2025 Google Cloud Export Summit Shenzhen. Video Source: 【提示词注入防御最佳实践】 https://www.bilibili.com/video/BV1DLwEeaEDa |
| Jan 24, 2025 | Swarm Code Reading Source code: https://github.com/openai/swarm My opensourced demo of how to modify and use swarm: https://github.com/haoruilee/huggingface-swarm |
| Jan 24, 2025 | Multi Agent RL & Web3 *Update at 2025.02.08: I write an implement of this idea https://github.com/haoruilee/Principal-Agent-Contract* Recently I read this paper *Principal-Agent Reinforcement Learning: Orchestrating AI … |
| Jan 22, 2025 | Tensor Parallelism and NCCL Recommend Reading: https://uvadlc-notebooks.readthedocs.io/en/latest/tutorial_notebooks/scaling/JAX/tensor_parallel_simple.html **Tensor Parallelism** optimizes the computation of large matrix oper… |
| Jan 22, 2025 | From RL to RLHF Note: This blog is a small part of these sources, recommend read them all. $1 categorizes techniques for aligning large language models (LLMs) into **four main themes** with subtopics: |
| Jan 21, 2025 | Invest Harvey Index **"Rich is having money (or assets) you haven't spent."** Over the past year, I've been exploring steady ways to grow wealth SLOWLY BUT SURELY. After countless trials, I've finally developed what I… |
| Jan 6, 2025 | 一天很漫長 但十年卻很短暫 Sam Altman 是Y Combinator這家創投公司的CEO，他最近剛滿三十歲。朋友問他這十年來他所經歷的，有什麼可以給別人建議的…他就洋洋灑灑寫了這個清單，我看完深深有感觸，就翻譯了一下… I turned 30 last week and a friend asked me if I’d figured out any life advice in the past decade… |
| Dec 23, 2024 | Qwen QwQ Guid In the rapidly evolving landscape of artificial intelligence, language models have transcended traditional boundaries, enabling applications that were once deemed impossible. Among these, QwQ https… |
| Dec 8, 2024 | Paper reading Online Learning Interesting online learning paper that solve the whole problem using pure math. Nice intro to the whole field. Data stream $D=\left\{\left(x_t, y_t\right) \mid t \in\{1,2, \ldots, T\}\right\}$, $x_… |
| Dec 6, 2024 | Bitcoin Whitepaper Summary *2024.12.06, write at the day bitcoin price hit $100,000* FTAV posts between June 2011 and today may have communicated the idea that bitcoin is a negative-sum game being played on a protocol that’s… |
| Oct 25, 2024 | Mathematical Models and Life Choices There are many interesting mathematical models in the world that can, to some extent, offer us a lot of insights. Chinese Version: https://zhuanlan.zhihu.com/p/871793375 |
| Jul 16, 2024 | Why GPT cannt compare Cuz this is a virus meme on Chinese social media, this post is first wirtten in Chinese. 最近有大量的群聊在讨论为什么GPT无法正确地比较9.11和9.9 可以看到无论是微软的Copilot，还是GPT4o，或者调用GPT API的外部网站，都无法正确地比较9.11和9.9的大小，最好玩的是即使程序已经明… |
| Jul 14, 2024 | Real Meaning of Rank Here, we first discuss a mathematical question that has long puzzled students of engineering and even physics: What is the real meaning of area, and how is it generalized to higher dimensions? |
| Jun 26, 2024 | Threads and Processes ***Robert Love on Quora:*** Here is the analogy I use in *Linux Kernel Development.* **Processes are the abstraction of running programs** : A binary image, virtualized memory, various kernel resou… |
| Jun 16, 2024 | Read AlexNet in 2024 After 12 years, reading Alexnet again can still give us insights. This work comes from UoT, Alex Krizhevsky is first author, his others famous paper: |
| Feb 17, 2021 | About this site This is my resume site, read `` to jump. If the webpages load slowly, consider switching the server: In order to improve the speed of access, this site uses a large number of .webp format pictures,… |
| Feb 16, 2021 | acknowledgement 其中为加速网页打开速度，对MIc-Theme改动为$1 |