LLM Wiki & Second Brain

Ghi chú chỉ sinh lời khi có người bảo trì. Câu hỏi là ai. Notes only pay off when someone maintains them. The question is who.

Zettelkasten, Second Brain, RAG và LLM Wiki là bốn câu trả lời cho cùng một vấn đề: công bảo trì kiến thức tăng nhanh hơn giá trị nó mang lại. Trang này giúp bạn chọn đúng — kể cả chọn không dùng gì. Zettelkasten, Second Brain, RAG and LLM Wiki are four answers to one problem: the cost of maintaining knowledge grows faster than the value it returns. This page helps you choose — including choosing nothing.

Three sources ingested one by one; wiki pages and links accumulate. CLAUDE.md: luật bảo trì wiki CLAUDE.md: the wiki's maintenance rules raw/ wiki/
LLM Wiki khi chạy: mỗi nguồn mới là một lượt LLM đọc và ghi (nét đứt) — xong là hết. Thứ ở lại là trang và liên kết (nét liền). Ba nguồn đã thành tám trang, mười hai liên kết, và còn tăng tiếp. Đây là điều RAG không làm: RAG tìm lại từ đầu mỗi lần hỏi. An LLM Wiki at work: each new source is one LLM read-and-write pass (dashed) — then it's gone. What stays are pages and links (solid). Three sources became eight pages and twelve links, and it keeps growing. RAG doesn't do this: it re-finds everything from scratch on every question.

Second Brain là mục tiêu, LLM Wiki là một cách xây nó. Chúng không đối đầu nhau — điều hay bị hiểu nhầm nhất khi hai thuật ngữ xuất hiện cùng nhau. Second Brain is the goal; LLM Wiki is one way to build it. They don't compete — the most common confusion when the two terms appear together.

Chọn cái nàoWhich one to use

Bắt đầu bằng câu hỏi có cần gì không. Phần lớn người bỏ cuộc với ghi chú vì dựng hệ thống cho một nhu cầu không tồn tại.Start by asking whether you need anything. Most people who give up on notes built a system for a need that didn't exist.

Baseline: không dùng gì cảBaseline: use nothing

  • Bạn không dùng lại kiến thức đó để viết, quyết định hay dạy lại.
  • Chủ đề ngắn hạn hoặc dưới khoảng 10 nguồn — dán hết vào một chat context dài là đủ.
  • Thông tin tìm lại trên web trong 30 giây.
  • You won't reuse it to write, decide or teach.
  • The topic is short-lived or under roughly 10 sources — paste it all into one long-context chat.
  • You can find it on the web again in 30 seconds.
du-an-x/notes.md   một file, ghi theo ngày, mới nhất ở trên
du-an-x/nguon/     PDF, link đã lưu
+ tìm toàn văn:    Ctrl+F, Spotlight, rg "từ khoá"
project-x/notes.md   one file, dated, newest on top
project-x/sources/   PDFs, saved links
+ full-text search:  Ctrl+F, Spotlight, rg "keyword"

Chỉ nâng cấp khi cùng một kiểu "đau" lặp lại ba lần trong một tháng.Upgrade only when the same pain repeats three times in a month.

  1. Bạn có dùng lại kiến thức này để làm ra thứ gì không?Will you reuse this knowledge to make something?
    Không → BaselineNo → Baselinebookmark + tìm kiếmbookmarks + search
  2. Mọi nguồn dán vừa một lần vào chat, và bạn chỉ hỏi vài lần?Do all sources fit in one chat, and you'll only ask a few times?
    Có → BaselineYes → Baselinemột chat context dàione long-context chat
  3. Nguồn nhiều, hỏi lặp lại lâu dài. Mục tiêu chính của bạn?Many sources, repeated questions over time. Your main goal?

Tổ hợp đáng thử nhất: PARA cho dự án, mỗi chủ đề nghiên cứu lớn là một LLM Wiki nằm trong Resources/, kèm một thư mục bạn tự viết nhận định để phần hiểu vẫn là của bạn.The combination most worth trying: PARA for projects, each big research topic as an LLM Wiki inside Resources/, plus a folder where you write your own takes so the understanding stays yours.

Bốn khái niệmThe four concepts

Xếp theo năm ra đời. Đọc liền mạch, đây là câu chuyện con người tự bảo trì, rồi máy tìm hộ, rồi máy bảo trì hộ.Ordered by year of origin. Read in sequence, it's a story: humans maintain, then machines retrieve, then machines maintain.

1952

Zettelkasten

Người bảo trìHuman-maintained Niklas Luhmann

Mỗi thẻ một ý, viết bằng lời của bạn, nối với nhau thành mạng. Giá trị nằm ở liên kết, không ở từng thẻ. Luhmann tích luỹ khoảng 90.000 thẻ.One idea per card, in your own words, linked into a web. The value lives in the links, not the cards. Luhmann built up about 90,000 of them.

Dùng choUse for
luận án nhiều năm, viết dài hạn, xây quan điểm riêngmulti-year theses, long-form writing, building your own view
ĐượcGain
hiểu sâu vì phải viết lại; ý tưởng mới từ liên kết chéodeep understanding from restating; new ideas from cross-links
MấtLose
chậm; lợi ích đến sau nhiều tháng, đa số bỏ cuộc trước đóslow; payoff comes after months, most quit first

Đổi tốc độ ghi chép lấy độ hiểu sâu và ý tưởng của riêng mình.Trades note-taking speed for deep understanding and ideas of your own.

Dấu hiệu: bạn cần ý tưởng mới từ việc nối những thứ đọc cách nhau nhiều năm.Sign: you need new ideas from connecting things read years apart.

2017

Second Brain

Người bảo trìHuman-maintained Tiago Forte — CODE + PARA

Lưu thứ đáng giữ, xếp theo việc đang làm (PARA: thanh dài là thứ mở hằng ngày), chắt lọc dần và biến thành sản phẩm (CODE).Keep what's worth keeping, organize by the work in progress (PARA: longer bar, opened more often), refine gradually and turn it into output (CODE).

Dùng choUse for
nhiều dự án song song, làm content, giấy tờ cá nhânparallel projects, content work, personal paperwork
ĐượcGain
bắt đầu trong 30 phút; thứ cần dùng luôn gần tay; tái dùng đầu rastart in 30 minutes; what you need stays close; outputs get reused
MấtLose
dễ thành kho chứa; liên kết giữa các ý yếu; sống nhờ review hàng tuầneasily becomes a hoard; weak links between ideas; lives or dies by the weekly review

Đổi chiều sâu liên kết giữa các ý tưởng lấy tốc độ biến thông tin thành sản phẩm.Trades depth of connection between ideas for speed in turning information into output.

Dấu hiệu: bạn biết đã lưu nó ở đâu đó nhưng không tìm lại được lúc cần.Sign: you know you saved it somewhere but can't find it when needed.

2020

RAG

Máy truy xuấtMachine-retrieved Lewis et al. — NotebookLM, ChatGPT file upload

Cắt tài liệu thành chunk. Mỗi câu hỏi tìm vài chunk liên quan rồi để LLM trả lời từ đó. Không ghi lại gì — lần sau làm lại từ đầu.Chunk the documents. Each question fetches a few relevant chunks and the LLM answers from them. Nothing is written back — next time starts over.

Dùng choUse for
chatbot trên tài liệu sản phẩm, tra cứu quy định, kho tài liệu khổng lồproduct-docs chatbots, policy lookup, huge document stores
ĐượcGain
dựng ngay; quy mô gần như vô hạn; trích dẫn tới đoạn gốcinstant setup; near-unlimited scale; citations to the source passage
MấtLose
không tích luỹ; câu hỏi cần nhìn toàn kho chỉ thấy vài mảnhnothing accumulates; whole-corpus questions only see fragments

Đổi khả năng tổng hợp tích luỹ lấy chi phí bảo trì gần bằng không và quy mô lớn.Trades cumulative synthesis for near-zero maintenance and large scale.

Dấu hiệu: đáp án nằm gọn trong vài đoạn văn cụ thể.Sign: the answer sits within a few specific passages.

2026

LLM Wiki

Máy bảo trìMachine-maintained Andrej Karpathy — raw / wiki / schema

LLM đọc nguồn, biên dịch thành wiki markdown có liên kết, rồi giữ nó luôn mới. Bạn chọn nguồn và đặt câu hỏi; LLM lo sổ sách.The LLM reads sources, compiles them into an interlinked markdown wiki, and keeps it current. You pick sources and ask; the LLM does the bookkeeping.

ingest → query → lint

Dùng choUse for
nghiên cứu nhiều tuần, đọc sách dày, wiki nội bộ teamweeks-long research, long books, internal team wikis
ĐượcGain
kiến thức tích luỹ; công bảo trì gần bằng không; mâu thuẫn được ghi nhận ngayknowledge compounds; near-zero upkeep; contradictions flagged on arrival
MấtLose
lỗi cũng tích luỹ; bạn có thể hiểu ít hơn; tốn tokenerrors compound too; you may understand less; token-hungry

Đổi quyền kiểm soát câu chữ và một phần sự thấu hiểu lấy kho kiến thức tự bảo trì, tích luỹ theo thời gian.Trades control over the wording, and part of your own understanding, for a self-maintaining base that compounds over time.

Dấu hiệu: bạn liên tục thêm nguồn về một chủ đề và hay hỏi câu cần nhiều nguồn.Sign: you keep adding sources on one topic and ask questions spanning several.

Bảng đánh đổiTrade-offs

Các trục thật của chủ đề: ai bỏ công, kiến thức có tích luỹ không, tin được đến đâu, và bạn có hiểu thêm không.The real axes of this topic: who does the work, whether knowledge accumulates, how far you can trust it, and whether you understand more.

Zettelkasten Second Brain RAG LLM Wiki
Ai bảo trìWho maintains BạnYouBạnYou Không ai — máy tìm lạiNobody — machine re-findsLLM, bạn giám sátLLM, you supervise
Chi phí khởi độngSetup cost ThấpLowRất thấpVery low Rất thấp (app có sẵn)Very low (ready-made apps)Trung bình (agent + schema)Medium (agent + schema)
Công mỗi nguồn mớiEffort per source CaoHighTrung bìnhMedium Gần bằng 0Near zeroThấp (đọc diff)Low (review the diff)
Tích luỹCompounding Mạnh, chậmStrong, slowVừaModerate KhôngNoneMạnh, nhanhStrong, fast
Truy vết, độ tinTraceability, trust Cao — chính bạn viếtHigh — you wrote itCaoHigh Vừa — trích tới chunkMedium — cites chunksVừa; thấp nếu không lintMedium; low without lint
Bạn hiểu sâu thêmYour understanding Cao nhấtHighestVừaMedium ThấpLowThấp nếu không tự đọcLow unless you read it
Quy mô hợp lýSensible scale Hàng nghìn thẻ, nhiều nămThousands of cards, yearsTheo số dự ánBy project count Gần như không giới hạnPractically unlimited~100 nguồn với index; lớn hơn cần công cụ tìm kiếm~100 sources with an index; beyond needs search
Rủi ro lớn nhấtBiggest risk Bỏ cuộc sớmQuitting earlyThành kho chứaBecoming a hoard Câu hỏi tổng hợpSynthesis questionsLỗi lan truyềnError propagation

Sai lầm hay gặpAnti-patterns

Sai lầmMistake
Vì sao saiWhy it's wrong
Sửa thế nàoHow to fix
Lưu mọi thứ "phòng khi cần"Saving everything "just in case"
Vì sao sai: Why: Lưu tạo cảm giác đã học nhưng chưa xử lý gì — collector's fallacy.Saving feels like learning, but nothing was processed — the collector's fallacy.
Sửa: Fix: Chỉ lưu khi viết được "cái này dùng cho ___".Save only if you can finish "this is for ___".
Đổi app liên tụcConstantly switching apps
Vì sao sai: Why: Mỗi lần chuyển mất vài tuần, kiến thức không tăng.Each move costs weeks; knowledge doesn't grow.
Sửa: Fix: Dùng file markdown thuần, ít nhất 3 tháng mới đánh giá.Use plain markdown files; judge after 3 months.
Dựng cây thư mục trước khi có nội dungBuilding folders before content
Vì sao sai: Why: Phân loại trước khi hiểu luôn sai và phải làm lại.Categorizing before understanding is always redone.
Sửa: Fix: Bắt đầu phẳng; tạo thư mục khi nhóm vượt 10 file.Start flat; add a folder once a group passes 10 files.
Để LLM viết wiki mà không đọcLetting the LLM write the wiki unread
Vì sao sai: Why: Kho thông minh lên, bạn thì không; lỗi tích luỹ không ai thấy.The base gets smarter, you don't; compounding errors go unseen.
Sửa: Fix: Đọc diff mỗi lần ingest; tự viết nhận định của mình.Read each ingest's diff; write your own takes.
Batch-ingest hàng trăm nguồn ngày đầuBatch-ingesting hundreds of sources on day one
Vì sao sai: Why: Schema chưa ổn, lỗi nhân lên hàng trăm trang.The schema isn't settled; errors multiply across pages.
Sửa: Fix: Giám sát 10–20 nguồn đầu, rồi mới batch.Supervise the first 10–20 sources, then batch.
Cho LLM sửa raw/Letting the LLM edit raw/
Vì sao sai: Why: Mất nguồn sự thật, không còn gì để kiểm chứng wiki.You lose the source of truth to verify against.
Sửa: Fix: Cấm trong schema; kiểm bằng git.Forbid it in the schema; check with git.
Câu trả lời hay nằm lại trong chatGood answers left in chat
Vì sao sai: Why: Mất phần tổng hợp bạn đã trả token để có.You lose synthesis you paid tokens for.
Sửa: Fix: Lưu vào wiki/analyses/.File it into wiki/analyses/.
Dùng RAG cho câu hỏi toàn khoUsing RAG for whole-corpus questions
Vì sao sai: Why: Top-k chỉ thấy vài mảnh; câu trả lời thiếu mà trông vẫn đủ.Top-k sees fragments; the answer is incomplete yet looks complete.
Sửa: Fix: Dùng wiki đã biên dịch, hoặc đưa cả kho vào context nếu vừa.Use a compiled wiki, or load the whole corpus if it fits.

Ba câu cần nhớThree things to remember

  1. Kiến thức chỉ sinh lời khi được bảo trì; chọn hệ thống thực chất là chọn ai bảo trì.Knowledge only pays off when maintained; choosing a system is choosing who maintains it.
  2. RAG tìm lại từ đầu mỗi lần; LLM Wiki biên dịch một lần rồi giữ cho mới — nhưng raw/ mới là sự thật.RAG re-finds every time; an LLM Wiki compiles once and keeps it current — but raw/ is the truth.
  3. Giao sổ sách cho LLM, giữ phần suy nghĩ cho mình.Hand the bookkeeping to the LLM; keep the thinking for yourself.