ML / LLM notes for learning, building, and interview preparation.

When I started preparing for industry roles, my materials were scattered across lectures, papers, documentation, and bookmarks. Finding a good explanation took time. Working out what to read next took more. These notes grew out of trying to bring that material together.

Here you'll find ML and LLM notes in English and Chinese, with examples, diagrams, code, and project walkthroughs. They're free to read and will stay that way. You can start with the basics, review for an interview, or explore a field outside your own. The aim is to make useful material easier to find and follow, even if you're new to it.

English or 中文: use the switch at the top of any note. It opens the same article in the other language; Chinese notes keep familiar English technical terms. The site starts in English and remembers your choice in this browser. A direct link to a language opens that version.

Start here

Choose for the topic, not your job title. You might know systems well and still want an introduction to RL.

How to use the notes

1. Pick something you want to understand. Use a route above, or the topic guide. You don't need to read the site in order.
2. Follow the example, then try it. Work through a calculation or run the code. Continue to the derivation and tradeoffs when you want more detail. We're still improving this progression in some chapters.
3. Keep your place. Choose “Save this page” under a note's title, and find it later under “My reading” at the top right; a note reopens where you stopped. No account or fork is needed. Records stay in this browser; export a JSON backup to keep a copy or move devices.

Time estimates mean focused learning time, not a daily target. Split a session if you need to; there is no streak to maintain.

What's in each section

• Foundations — How models work, learn, and are evaluated. The study guide is a starting point for language models.
• Interview prep — ML / LLM review, ML coding, Python, and reusable algorithm patterns.
• Engineering practice — Recommendation, RAG, and post-training systems: how data, training, serving, and evaluation fit together.
• Tech history — Developments in power, chips, networks, software, and AI, with links to original sources.
• Career notes — What I tried during internship and new-grad searches, and what I'd do differently.

Diagrams show the main idea without requiring interaction. The controls are there when you want to change an input or compare outcomes.

Practice with your AI agent (interview coach skill)

Install the notes into an agent that reads skills, such as Claude Code or Codex, and it becomes an interviewer: one question at a time, graded against the notes, with follow-up questions, and a record of what you got wrong so the next session starts there. The notes and questions come from the site's read-only data API and update with the site; your answers and scores stay on your machine.

npx skills add https://github.com/tianyi-zhang-02/cooking-agi --skill cooking-agi-interview

Then ask, for example, "Mock interview for an LLM inference engineer role, 5 questions" or "What am I weak at? Quiz me on that". See skills/README.md for details. The skill was contributed by @Juggernaut0825.

What's inside

A few articles to start reading now:

• How models work: Transformer walkthrough · BERT · CLIP and multimodal learning.
• How models learn: Training basics · Deep RL · Post-training · LLM-as-a-Judge.
• How models get used: Inference · Search and retrieval · Agents and tools.
• What's moving: Frontier watch (world models, self-improvement, test-time compute) · AI infrastructure and perception, robots and driving, both drafts pending review.

For practice, head to ML coding, Python, algorithm patterns, or system design.

Still being worked on: some explanations and source comparisons need more review. The coverage and backlog separates material that exists from material that has been checked. Small tested examples are not claims of full-model training or benchmark reproduction.

Roadmap

Now

Better notes, not more of them. Every note has been re-read once in both languages, and in October the facts were checked again against primary sources: errors fixed, stale numbers updated, repeated explanations folded into one place. The newest series (AI infrastructure, perception and driving, three frontier series, scaling, diffusion LMs and quantization) are drafts until the site owner has read them. If something reads awkwardly, is hard to follow, or seems wrong, please tell us; every article also ends with a feedback button.

Next: a faster way to the answer: better search and quicker pages, then an MCP server, so Claude, ChatGPT and Cursor can search the notes and quiz you from the same data as the interview coach skill. For Jetson, physical simulation, and reviewing the vision, driving and robotics drafts, we'd like people who have worked in those areas to join us. See the roadmap for the full plan; the scope of the upcoming v1.0 interview-prep release is in release scope.

Topic coverage

How many notes each topic has, and how far this review pass has got. The numbers update on every site deploy.

How we got here

Details of every merge are in the changelog at the bottom of this page.

Contributors and reviewers

Everyone with a commit or co-author credit, the reviewers who approve changes in each area, and where readers went next if they chose to share it. The card is rebuilt from git history, GitHub and site/collaboration.toml on every deploy and every hour. Commit counts are not a ranking: questions, reviews and corrections count too. To be credited for a review or correction, or to share your next stop (only what you choose to show), open an issue: credit me · share my next stop.

Join us!

If a passage is confusing, something has changed, or a topic is missing, we'd like to hear about it. You don't need to have the answer before asking a question.

• Give feedback: open an issue and link the part you're referring to. You can flag a correction, outdated information, an unclear explanation, a missing topic, or a site problem.
• Make a change: small fixes can go straight into a PR. For a new article or navigation change, start with a content proposal so we can coordinate.
• Pass it on: if the notes helped you learn or prepare, share the link with a friend or classmate. Sharing an employer or job-search outcome is never required. The people behind the notes are listed under Contributors and reviewers.

These are our working notes, not an official curriculum. Our experience is mainly in ML / LLM and MLE / research preparation, and it shapes what we cover. Some material will be incomplete or out of date. We share public knowledge and preparation methods, not company-specific interview information or question banks. We can't predict questions or promise an offer.

Free to read, open to reuse. There are no paid chapters or premium preparation routes. Original code uses MIT; original notes use CC BY 4.0. Keep attribution, source links, and the relevant notices, and mark your changes. Third-party materials retain their own terms. We don't endorse others' edited or redistributed versions; warranty and liability terms are in the licenses.

Run locally

This repository contains the notes, examples, and static-site source. No database is needed. From the repository root, with Python 3.12 recommended:

python3 -m pip install markdown pygments numpy
python3 site/build.py --serve

Open http://localhost:8000. To build without starting a server, run python3 site/build.py; output goes to _site/.

python3 site/collaboration.py
python3 -m unittest discover -s site/tests
python3 site/leakcheck.py
python3 site/paritycheck.py --strict
python3 site/build.py

Why this exists

Moving from academic research toward industry, I often found that something I recognized in a paper was much harder to explain or implement myself. Writing it down helped me notice those gaps. Sharing the notes gives other people a chance to point them out too.

I hope this makes it easier to get started without already having a background in the field or someone to hand you a reading list. It won't replace study and practice. Where others explain something better, we'll link to them; where we lack experience, we'd like to learn from people who have it. Comparing notes across fields may reveal connections we wouldn't find on our own.

If it saves you some searching and leaves a little more time for the rest of your life, that's a good reason to keep working on it.

Changelog

Every change merged into the main branch, by date, split into major and minor updates. Links go to the pull requests.