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.

PPO, One Question at a Time: From Rewards to Policy Updates

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.

Agent Harness Engineering: Why I Recommend AI Agents in Depth

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.

Learning RL from the Math: 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.

Retrieval, Explained: BM25, Dense Embeddings, SPLADE, and ColBERT

How I think about noisy labels, data curation, and evaluation that deserves to be trusted, with lessons from MIT and three reproducible teaching notebooks.

Before I Trust the Model, I Audit the Data

A practical mental model for PyTorch: tensor shapes, shared storage, autograd, and model state. Worked examples connect indexing and parameter registration to nanoGPT, gradient checks, KV caches, and training-loop correctness.

PyTorch, Beyond the API: Shapes, Storage, Gradients, and State

how therapy changed my life, at least so far

Jim Rohn — Life Philosophy Notes

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.

How to Become an LLM Engineer: A Knowledge Map and Learning Method

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。

Facebook 增长不是增长黑客：哪些策略可以迁移到 ChatGPT 这样的产品？

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”。

把自己从自动驾驶里叫醒 (4min read)

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。

学得更快，不是更紧：从 Reddit 和知乎提炼的高效学习线索

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.

Visualization Is Not About Winning. It Is About Rehearsal.

Most big decisions are not problems to solve but bets to place. A better framework: expand your option set, identify the uncertainties that actually matter, run cheap tests, prefer reversibility, and build adaptation into the decision itself.

You Don't Need Certainty. You Need a Decision System

对 80,000 Hours《What makes for a dream job?》的中文提炼：职业满足感更接近“擅长 + 助人 + 好环境 + 与生活匹配”，而不是单纯追逐高薪、低压或所谓天命热爱。

理想工作不是高薪低压，而是这六个要素 (8min read)

Top 5 Most Surprisings From Claude Code Source Map

🧠 Why We Feel Empty in a World That Gives Us Everything

Zen Is Not Calm. Zen Is a Weapon.

Claude Code: A Highly Agentic Coding Assistant — A Deep, Practical Review 💻

Elon Musk’s 15 Daily Prompts That Rewired How I Think About Hard Problems

极端优秀 vs 一般优秀

Hanlon’s Razor: The Mental Model That Reduces Stress and Drama

Mental Model IV - Habit Management

Mental Model III - Emotion Management

Mental Model II - Cognitive Management

Mental Model I - Learning Management

A Peek into Elon Musk's Success: Insights from a Visionary

Life with your most potential

Why You Need a Product Roadmap (And When You Don’t)🚀 

Mastering Work-Life Balance in Tech: Lessons from a Staff Developer

🧠 Mastering Mentorship: Insights from a FAANG company

🌟 What I Wish Someone Had Told Me 🌟 by Sam Altman

Hidden Technical Debt in Machine Learning Systems

How to Implement MLOps: A Guide to Elevating Your Machine Learning Practices

7 Best Practices for MLOps: Optimizing Team Collaboration and Model Performance

3 Paradigms of LLM4Rec 🌟☄️

Mastering the Art of the 1:1 Meeting: Transformative Tips for Managers 🌟✨

what would happen if AI does all the jobs for us as human?

Post-Employment Economics

what does the people working in Silicon Valley say about LLM?

What does LLM mean to your career? A poll from Silicon Valley Tech Managers/Staffs (10mins read)

how Google secretly dominate the ads market? Chrome!

3P Cookies and Chrome/Google (5min read)

🚀 Enhancing Retail Success with Advanced AI and Machine Learning Techniques 🚀 (5min read)

what are the real differences between sr manager and director?

🌟 Demystifying the Roles: Senior Managers vs. Directors in Big Tech 🚀 (3min read)

You should at least think this through when considering the use of LLM in RecSys

🚀 The Rise of RecLLM: A New Era for Recommendations (10min read)

2 kings in the era of large language model; one is close source, the other is open source

OpenAI vs Meta (3min read)

OpenAI vs Google (3min read)

Will OpenAI create the next-gen cloud computing platform

OpenAI vs AWS Cloud (3min read)

my guesstimate/bestimate in an ordered list

Who is targeted by OpenAI? (3min read)

OpenAI DevDay: The biggest day after human walk out of Africa (7min read)

how a third player can challenge the leading big 2?

Amazon Ads at First Glance (3min read)

When Dr Qi Lu speaks, people listen - HYSTA Keynote (3min read)

Something you don’t know about ChatGPT - part 1 (3min read)

Excellent life advices from Kevin Kelly with my own editorials (7min read)

a new tool for fast development of web app api

Accelerating API Development: A Dive into FastAPI (7min read)

we should develop multidisciplinary perspectives to solve real world problems

Bridging Disciplines: A Glimpse into "Latticework: The New Investing" by Robert G. Hagstrom (5min read)

search and learning are the two most scalable and future-ready algorithms

Unveiling The Sweet Side of AI’s Bitter Lesson: A Dive into ChatGPT and Generative AI (3min read)

Llama 2 Everywhere: Bridging the Gap between Minimal Hardware and AI (5min read)

Mindmap 🍁

Elon Musk by Walter Isaacson - Part II (5min read)

Elon Musk by Walter Isaacson - Part I (5min read)

 France's Crucial Influence in the AI and Generative Models Landscape 🤖 (5min read)

Why writing blogs? or why should you write? (3min read)

The art of doing nothing, but the most important things

Productivity Guide from Marc Andreessen - The Art of Doing Nothing... But The Most Important Things 🎨 (3min read)

be adaptive, be like water, flow around and make yourself a great product first!

🌐🎓 Welcome to the Tech Wonderland: A Mega-Guide to Career & Education Choices 🌟💼 (5min read)

The groundbreaking study on ChatGPT reveals a systematic left-leaning political bias, raising critical questions about AI's influence on public opinion and democracy, and calling for collaboration between technology leaders, lawmakers, and society to uphold human dignity in the algorithmic age. 🤖💬🌍

🤖 Measuring Political Bias in ChatGPT - What Does This Mean for Society? 🤖 (7min read)

Explore the evolution of Large Language Models (LLMs) and the shift from GPUs to alternative hardware. Dive into the world of FPGAs, ASICs, and optical computing, and discover what the future holds for AI technology.

LLM Open Challenges 3: Do we always need GPUs? (3 min)

RAVEN, a groundbreaking AI model, utilizes retrieval-augmented mechanisms to excel in in-context learning, outperforming the ATLAS model and opening new possibilities in AI. 🚀

RAVEN: Unleashing the Power of In-Context Learning 🚀 (3min read)

Platypus: Revolutionizing Large Language Models 🚀 (5min read)

 Huggingface.co: A Journey from Start-up to AI Superstar 🌟 (5min read)

a fine-tuned llama2 can help answer some of your medical questions

Introducing DoctorGPT: Your Private AI Doctor 🩺💻 (3min read)

🧠💡 Self-Alignment with Instruction Backtranslation": A Groundbreaking Approach to Language Model Training 🚀📊 (5min read)

 🚀💰Together AI's Game-Changing Price Reduction: A Deep Dive into Faster Inference and Optimized Pricing

Unleashing the Power of GPT-4 with Wolfram Alpha and Code Interpreter: A Comprehensive Test (5min read)

google found that synthetic-data intervention to reduce models' tendency to agree with users' incorrect opinions

🚀 Tackling Sycophancy in Language Models: A Comprehensive Insight by Google AI 🚀 (5min read)

ethical and reliable evaluations of llm is critical

🤖💡Trusting the Machines: A Comprehensive Guide to Evaluating Large Language Models' Alignment (5min read)

🧠💼 D-Bot: Revolutionizing Database Administration with AI 🚀 (5min read)

if you need an open agi projects, take a look at here

Exploring Open-Source AGI Projects: Use Cases and Comparisons (5min read)


ChatGPT vs. Stack Overflow study reveals 52% of ChatGPT's answers are incorrect, yet preferred 39% of the time for their comprehensive language.

Who Answers It Better? ChatGPT vs. Stack Overflow in Software Engineering Questions (3min read)

A universal language for state-of-the-art audio synthesis

🎶 AudioLDM 2: A Symphony of AI-Generated Sound 🎧 (5min read)

Terence Tao talks about how mathematical think is related to every day life.

Unraveling Life's Equation: Mathematical Thinking in Everyday Life 🌐🔢 (3min read)

a multi-agent assembly line approach, mirroring human workflows for complex problem-solving

🤖 MetaGPT: Building the Future of AI Collaboration with an Assembly Line Approach 🏭 (5min read)

you need personalization, simplicity, integrity and extensibility

LLM Open Challenges 1: How to improve efficiencies of chat interface? (3min read)

llm agents are used to simulate societal-economical events

Large Language Models as Simulated Economic Agents: A New Frontier in Economics (3min read)

how to generate ML/AI ideas in your job and manager afterwards

ML Team Leader Q&A Part 3 - Ideas and Executions (5min read)

what careers are you looking for in the ML/AI community

ML Team Leader Q&A Part 2 - ML Career Path (5min read)

ML Team Leader Q&A Part 1 - ML Team First Glance (5min read)

where you can find distributed ML and software system designs

Distributed System Design 🔬 (10min read)

how is it different to build LLM for non-english languages

🌐 LLM Open Challenges 2: Large Language Models for Non-English Languages: Challenges and Perspectives 🚀 (3min read)

balancing and transparency of the three levers ensure projects fit into bigger picture

🎛️Your Levers: Scope, Resources, and Time - The Triad of Project Management (3min read)

engineering seem slower due to decreased productivity per head

🐢 Why is Engineering Getting So Slow? Unraveling the Paradox (3min read)

Navigating The Core Trio: Search, Recommendation, and Advertising Algorithms in the Digital Realm

Search, Recommendation System, and Generative AI 🧲 part 3 (5min read)

leadership fosters purpose and passion in teams, rather than fear or intimidation

📏The Stick: Navigating the Delicate Balance of Leadership (3min read)

what does the role of machine learning manager at FAANG look like?

Navigating the Role of a Machine Learning Manager at Amazon 🧭 (3min read)

each company has their own choice, how about yours?

🚀 Monorepo vs. Polyrepo: A Technical Exploration 🚀 (3min read)

If a system returns what you need, does it matter how it was produced?

Search, Recommendation System, and Generative AI 🧲 part 2 (3min read)

How can we find the intersections between information generation, distribution and consumption?

Search, Recommendation System, and Generative AI 🧲 part 1 (3min read)

"Rockstars" and "Superstars" each requiring tailored environments to excel.

🎸Rockstars vs. Superstars: The Power Players in Tech Teams 🚀 (3min read)

managers broaden their influence within a company by building cross-departmental relationships.

🚀 Expanding Your Influence: The Power of Networking Inside Your Company 🚀 (3min read)

human are usually wrong, especially on the prediction of future.

The Pioneering Spirit of AI: A Journey Toward Human-Level Intelligence 🚀 (5min read)

to be a good manager, you need to understand the gear shift for cycling on uphill and downhill.

🚴‍♂️ Shifting Gears: The Uphills and Downhills of Software Development Cadence in ML Projects 🏔️ (3min read)

This blog post provides a comprehensive exploration of Altman's proposal, drawing from various disciplines and historical lessons. It's a complex and multifaceted issue that warrants continued discussion and exploration. The future is bright, but the path we choose will determine how we get there.

🌐 Envisioning the Future: An answer to Sam Altman’s proposal 🌐 (3min read)

Praxis Synthesis

Tech Dive

Mood Diary

Roaming Rhymester

Career Growth

Management Science

Level Up

Fresh Dispatch

Book Rumination

LLM+AIGC

Personal Philosophy

Interdivisual Psychology

Home

往期整理

Search

words

Contact

About

engineering

books

recommendations

learning

password

icon

Created time

date

type

slug

status

title

summary

Site Build

Table

Announcement 🎏

Life Dashboard | Notion

pretraining = unsupervised learning posttraining will increase targeted capability: a new area of capability or strenghtened capability in an existing area.

Traditional software development requires meticulous planning. With machine learning, it's better to jump right in.