5 cặp khái niệm AI hay bị lẫn 5 AI pairs people mix up

Năm cặp khái niệm nghe giống nhau, và cái giá của việc lẫn chúng Five pairs that sound alike, and what mixing them up costs

Cột bên trái luôn nghe như đã đủ. Cột bên phải mới là thứ khiến hệ thống tin được. Lẫn một cặp mất vài tuần; lẫn hai cặp mất một quý. The left column always sounds like enough. The right column is what makes a system trustworthy. Mixing up one pair costs weeks; mixing up two costs a quarter.

“Một vòng lặp không có đích sẽ chạy cho tới khi bạn nhận ra.” “A loop without a goal runs until you notice.”
PROMPT CONTEXT MODEL HARNESS RAG MEMORY GOAL ? nạp lại / reload loop retry — — — runtime ────── fixed

nét đứt = quyết định lúc chạy · nét liền = cố định lúc viết code dashed = decided at runtime · solid = fixed when you write the code

Chọn cái nào trong 30 giây Pick one in 30 seconds

Phần này dùng hằng ngày. Phần tra cứu chi tiết nằm bên dưới. This is the part you use daily. The reference cards are below.

Trước hết: khi nào không cần gì cả First: when you need none of this

  • Chạy dưới ~20 lần một ngày và có người đọc đầu ra trước khi dùng. Fewer than ~20 runs a day and a person reads the output before it is used.
  • Việc có luật rõ ràng: regex, tra bảng, một câu SQL. Luật rẻ hơn và đúng 100%. The task has clear rules: a regex, a lookup, one SQL query. Rules are cheaper and right 100% of the time.
  • Bạn chưa định nghĩa được thế nào là đầu ra đúng — chưa có gì để tự động hoá. You cannot define a correct output yet — there is nothing to automate.

Phương án rẻ nhất: một prompt viết tay, một người bấm chạy, một spreadsheet ghi kết quả. Chạy hai tuần, rồi chỉ dựng đúng thứ chữa chỗ đau. The cheapest alternative: one hand-written prompt, one person pressing run, one spreadsheet. Run it two weeks, then build only what treats the pain.

goal  →  harness  →  loop  →  memory  →  RAG  →  fine-tune

Thứ tự thêm khi thật sự đau. Fine-tune gần như luôn là bước cuối. The order to add things once it really hurts. Fine-tuning is almost always last.

Đầu ra có người đọc trước khi dùng không?
│
├─ CÓ ──── Chạy bao nhiêu lần một ngày?
│           ├─ < 20  ──► Prompt tay. Dừng. Không dựng gì thêm.
│           └─ ≥ 20  ──► Prompt + LOOP, kèm trần số vòng và trần tiền.
│
└─ KHÔNG ── Bắt buộc có HARNESS: schema + retry + log. Không ngoại lệ.
            │
            └─ Model đang thiếu cái gì?
               │
               ├─ Thiếu THÔNG TIN
               │   ├─ đổi hằng tuần / lượng lớn ──────► RAG
               │   └─ cố định và nhỏ ────────────────► nhét thẳng context
               │
               ├─ Thiếu TRÍ NHỚ giữa các phiên ──────► MEMORY (store → recall)
               │
               ├─ Sai PHONG CÁCH / ĐỊNH DẠNG
               │   ├─ prompt sửa được ───────────────► sửa prompt
               │   └─ prompt 1000+ token vẫn trượt ──► FINE-TUNE
               │
               └─ Chạy mãi không dừng ───────────────► thiếu GOAL,
                                                       không thiếu model
Does a person read the output before it is used?
│
├─ YES ─── How many runs per day?
│           ├─ < 20  ──► Run the prompt by hand. Stop. Build nothing.
│           └─ ≥ 20  ──► Prompt + LOOP, with an iteration cap and a spend cap.
│
└─ NO ──── A HARNESS is mandatory: schema + retry + logs. No exceptions.
           │
           └─ What is the model missing?
              │
              ├─ Missing INFORMATION
              │   ├─ changes weekly / large corpus ──► RAG
              │   └─ fixed and small ───────────────► put it in context
              │
              ├─ Missing MEMORY across sessions ────► MEMORY (store → recall)
              │
              ├─ Wrong STYLE / FORMAT
              │   ├─ a prompt change fixes it ──────► fix the prompt
              │   └─ 1000+ token prompt still misses ► FINE-TUNE
              │
              └─ It never stops ────────────────────► you are missing a GOAL,
                                                      not a better model

Năm cặp The five pairs

Câu đánh đổi nằm ở dải màu cuối mỗi thẻ. Đọc nó trước, phần còn lại là chi tiết. The trade sits in the coloured band at the foot of each card. Read that first; the rest is detail.

1

Loop vs Goal

điều khiển vòng chạy run control

LOOP

nó chạy lại bao nhiêu lầnhow often it runs again

↻run→wait 5m→run↻

Cơ chế gọi lại cùng một tác vụ theo nhịp. Bản thân nó không biết gì về kết quả. A mechanism that calls the same task again on a rhythm. It knows nothing about the result.

dùng chouse for
  • Quét hộp thư mỗi 10 phút để gắn nhãnScanning an inbox every 10 minutes to label it
  • Cron 6h sáng gộp 40 nguồn tinA 6am cron merging 40 news sources
  • Thử lại khi API lỗi hoặc đầu ra trượt kiểm traRetrying when an API drops or a check fails
đượcyou gain
  • Chạy khi bạn không có mặtIt runs while you are not there
  • Bắt sự kiện mới trong vài phútNew events picked up within minutes
  • Dựng mất mười phútTen minutes to build
mấtyou lose
  • Không có cửa ra thì đốt tiền âm thầmWith no exit it burns money quietly
  • Một lỗi lặp 288 lần một ngày nếu chạy mỗi 5 phútOne bug repeats 288 times a day at 5-minute intervals
  • Loop lồng loop nhân chi phí, không cộngNested loops multiply cost, they do not add

Dấu hiệu:Recognise it: mô tả được bằng câu “cứ mỗi… thì…” là loop. if you can say “every time X, do Y”, it is a loop.

vs

GOAL

khi nào nó được phép dừngwhen it is allowed to stop

start→work→done?

Một vị từ boolean trên trạng thái: xong(state) → true|false. Không phải một câu mô tả nguyện vọng. A boolean predicate over state: done(state) → true|false. Not a sentence describing an intention.

dùng chouse for
  • “Test suite pass” — kiểm bằng exit code“The test suite passes” — an exit code
  • “Không còn field bắt buộc nào trống” — kiểm bằng schema“No required field is empty” — a schema
  • “Đã gửi mail và có message id” — kiểm bằng API“Email sent, message id returned” — an API
đượcyou gain
  • Chi phí có trần và giải thích đượcCost capped and explainable
  • Lý do thoát là dữ liệu, không phải phỏng đoánThe exit reason is data, not a guess
  • Có tỷ lệ thành công để so giữa các phiên bảnA success rate you can compare across versions
mấtyou lose
  • Goal sai nguy hơn không goal: hệ thống tối ưu đúng cái saiA wrong goal is worse than none: it optimises the wrong thing
  • Goal đo được thường không phải goal bạn quan tâmThe measurable goal is often not the one you care about
  • Yêu cầu đổi thì goal cũ âm thầm saiWhen requirements change, old goals go stale silently

Dấu hiệu:Recognise it: viết không nổi hàm true/false cho “xong chưa” thì bạn mới có mong muốn. if you cannot write a true/false function for “is it done”, you only have a wish.

đánh đổithe trade Chọn loop là đổi quyền kiểm soát thời điểm dừng để lấy khả năng chạy khi bạn không có mặt. Choosing a loop trades control over when it stops for the ability to run while you are away.

đánh đổithe trade Chọn goal là đổi sự linh hoạt của tác vụ để lấy một điểm dừng chứng minh được. Choosing a goal trades task flexibility for a stopping point you can prove.

2

Prompt vs Harness

giao tiếp và độ tin cậyinterface and trust

PROMPT

thứ bạn hỏiwhat you ask

ask→answer

Toàn bộ chữ đưa vào một lần gọi. Không có bước nào kiểm tra kết quả — người kiểm tra là bạn. All the text in one call. Nothing checks the result — you are the check.

dùng chouse for
  • Nháp một email khó, bạn đọc lại rồi gửiDrafting a hard email you will read before sending
  • Tóm tắt tài liệu để tự đọcSummarising a document for yourself
  • Khám phá: thử năm cách trong mười phútExploring: five approaches in ten minutes
đượcyou gain
  • Sửa trong năm giây, không deployChanged in five seconds, no deploy
  • Không code, không hạ tầngNo code, no infrastructure
  • Vòng thử–sai–sửa nhanh nhấtThe fastest try–fail–fix cycle available
mấtyou lose
  • Không lặp lại được ổn địnhNot reliably repeatable
  • Không ai phát hiện khi đầu ra sai — nó vẫn trôi chảyNothing catches a wrong answer — it still reads fluently
  • Đổi model là prompt có thể hỏng lặng lẽA model change can break it silently

Dấu hiệu:Recognise it: bạn vừa đọc vừa sửa đầu ra thì prompt là đủ. if you both read and fix the output, a prompt is enough.

vs

HARNESS

cách bạn kiểm soát và tin nóhow you control and trust it

inputs→checks→output

Lớp code bao quanh lần gọi. Đầu ra bị coi là không đáng tin cho tới khi qua kiểm tra. The code around the call. Output is untrusted until it passes checks.

dùng chouse for
  • Trích hoá đơn thành JSON ghi thẳng vào databaseInvoices into JSON written straight to a database
  • Phân loại ticket với danh sách nhãn cho phépTicket classification with an allowlist of labels
  • Agent được gọi tool, mỗi tool có sandboxAn agent with tools, each one sandboxed
đượcyou gain
  • Sai thành lỗi bắt được, không phải dữ liệu bẩn ba tuần sauWrong becomes a caught error, not dirty data three weeks later
  • Đổi model là quyết định về số liệu, không phải canh bạcModel swaps become a decision about numbers
  • Có tỷ lệ pass theo dõi theo ngàyA pass rate you can watch daily
mấtyou lose
  • Tốn code gấp nhiều lần promptSeveral times more code than a prompt
  • Checks quá chặt loại cả câu trả lời đúngChecks that are too strict reject correct answers
  • Harness cũng có bug, và khó nhìn thấy hơnThe harness has bugs too, and harder to see

Dấu hiệu:Recognise it: đầu ra đi thẳng sang hệ thống khác mà không ai đọc → cần harness, không phải prompt tốt hơn. output flowing into another system unread means a harness, not a better prompt.

đánh đổithe trade Chọn prompt là đổi độ ổn định để lấy tốc độ thử. Choosing a prompt trades stability for speed of iteration.

đánh đổithe trade Chọn harness là đổi tốc độ thay đổi để lấy đầu ra đủ tin để không cần ai đọc. Choosing a harness trades speed of change for output nobody has to read.

3

Context vs Memory

trạng tháistate

CONTEXT

cái nó thấy ngay lúc nàywhat it sees right now

this turn→erased

Toàn bộ token của lần gọi hiện tại. Có trần cứng, và biến mất sau khi trả lời. Every token in the current call. A hard ceiling, gone after the reply.

dùng chouse for
  • File đang mở trong editorThe file open in the editor
  • Mười lượt hội thoại gần nhấtThe last ten turns of a conversation
  • Hợp đồng 20 trang cần trả lời ngay bây giờA 20-page contract you need answered now
đượcyou gain
  • Chính xác nhất: không có bước truy hồi chọn nhầmMost accurate: no retrieval step to pick the wrong thing
  • Thay đổi có hiệu lực tức thìChanges take effect instantly
  • Không database, không job nềnNo database, no background job
mấtyou lose
  • Trần token là trần cứng, chạm sớm hơn bạn nghĩThe token ceiling is hard, and closer than you think
  • Càng dài càng đắt và càng chậmLonger means more expensive and slower
  • Thông tin ở giữa context dài dễ bị bỏ quaMaterial in the middle of a long context gets skipped

Dấu hiệu:Recognise it: bỏ ra mà lần gọi hiện tại sai ngay thì đó là context. if removing it breaks the current call, it is context.

vs

MEMORY

cái nó lấy lại được về sauwhat it can pull back later

store→recall

Kho ngoài model. Nó chỉ có tác dụng khi nội dung được nạp ngược vào context. A store outside the model. It only matters once its contents are loaded back into context.

dùng chouse for
  • Hồ sơ người dùng: ngôn ngữ, cách xưng hô, định dạngA user profile: language, address, formats
  • Tóm tắt phiên trước để khỏi kể lạiSummaries of past sessions so nobody repeats themselves
  • Quyết định dự án: “đã chốt, không bàn lại”Project decisions: “settled, do not reopen”
đượcyou gain
  • Liên tục giữa các phiên, thứ context không làm đượcContinuity across sessions, which context cannot give
  • Giảm token: chỉ nạp phần liên quanFewer tokens: load only what is relevant
  • Cá nhân hoá thật, không hỏi lại mỗi lầnReal personalisation instead of asking again
mấtyou lose
  • Ghi sai một lần thì sai mãi, lan sang mọi phiên sauWritten wrong once, wrong forever, in every later session
  • Recall nhầm tệ hơn không nhớA wrong recall is worse than no memory
  • Bạn vừa tạo ra một kho dữ liệu cá nhân phải chịu trách nhiệmYou now own a personal data store you are liable for

Dấu hiệu:Recognise it: người dùng bực vì “tôi nói rồi mà” thì đó là memory. when a user says “I already told you”, it is memory.

đánh đổithe trade Chọn context là đổi tính bền lâu để lấy độ chính xác tức thời. Choosing context trades durability for immediate accuracy.

đánh đổithe trade Chọn memory là đổi việc phải quản lý vòng đời dữ liệu để lấy sự liên tục giữa các phiên. Choosing memory trades managing a data lifecycle for continuity between sessions.

4

RAG vs Fine-tuning

nguồn kiến thứcsource of knowledge

RAG

truy hồi, không train lạiretrieve, do not retrain

data→embed→fetch

Tìm đoạn liên quan ngay trước khi gọi. Chất lượng câu trả lời bị chặn trên bởi chất lượng bước tìm. Find the relevant passages just before the call. Answer quality is capped by retrieval quality.

dùng chouse for
  • Hỏi đáp tài liệu nội bộ đổi hằng tuầnQ&A over internal docs that change weekly
  • Bot hỗ trợ đọc changelog mới nhấtA support bot reading the latest changelog
  • Chính sách, hợp đồng — nơi bắt buộc dẫn nguồnPolicies and contracts, where citation is mandatory
đượcyou gain
  • Cập nhật bằng cách sửa tài liệu: phút, không phải ngàyUpdates by editing a document: minutes, not days
  • Trích dẫn được nguồnSources can be cited
  • Xoá tài liệu là xoá luôn khả năng trả lời từ nóDeleting a document deletes the ability to answer from it
mấtyou lose
  • Tìm sai thì model bịa rất trôi chảy từ đoạn saiBad retrieval makes the model invent fluently from the wrong passage
  • Tốn thêm token và độ trễ mỗi lần gọiExtra tokens and latency on every call
  • Không dạy được phong cách hay định dạngIt teaches nothing about style or format

Dấu hiệu:Recognise it: câu trả lời đúng thay đổi khi tài liệu thay đổi. the correct answer changes when the documents change.

vs

FINE-TUNING

train lại chính modelretrain the model itself

data→train→weights

Bạn không nhồi sự thật vào model, bạn dịch chuyển xu hướng của nó. You are not injecting facts, you are shifting the model's tendencies.

dùng chouse for
  • Ép giọng văn hoặc định dạng mà prompt dài vẫn trượtEnforcing a voice or format long prompts still miss
  • Rút system prompt 2000 token xuống 100Shrinking a 2000-token system prompt to 100
  • Cho model nhỏ đạt chất lượng model lớn trên một việc hẹpA small model matching a large one on one narrow task
đượcyou gain
  • Prompt ngắn hơn: rẻ và nhanh hơn mỗi lần gọiShorter prompts: cheaper and faster per call
  • Hành vi ổn định, ít phụ thuộc cách hỏiStable behaviour, less sensitive to phrasing
  • Chạy được model nhỏ hơn ở lưu lượng lớnA smaller model at volume
mấtyou lose
  • Cần hàng trăm–hàng nghìn ví dụ sạch; làm sạch là phần tốn nhấtHundreds to thousands of clean examples; cleaning is the slowest part
  • Mỗi lần đổi yêu cầu là train và đánh giá lạiEvery change means retraining and re-evaluating
  • Không cập nhật được sự thật; khoá bạn vào một phiên bảnFacts cannot be updated; it locks you to one version

Dấu hiệu:Recognise it: lặp cùng một hướng dẫn phong cách mọi lần gọi mà vẫn trượt. the same style instruction repeated every call, still missed.

đánh đổithe trade Chọn RAG là đổi độ trễ và token mỗi lần gọi để lấy khả năng cập nhật kiến thức mà không train lại. Choosing RAG trades latency and per-call tokens for updating knowledge without retraining.

đánh đổithe trade Chọn fine-tune là đổi khả năng sửa nhanh để lấy hành vi ổn định và chi phí mỗi lần gọi thấp hơn. Choosing fine-tuning trades fast fixes for stable behaviour and a lower cost per call.

5

Prototype vs Production

mức độ trưởng thànhmaturity

PROTOTYPE

chứng minh ý tưởngprove the idea

build→demo

Bản nhỏ nhất đủ trả lời một câu hỏi: có khả thi không, có đáng làm không. The smallest thing that answers one question: is it feasible, is it worth doing.

dùng chouse for
  • Notebook chạy tay trên 50 dòng dữ liệuA notebook run by hand over 50 rows
  • Thử ba model trước khi cam kếtTrying three models before committing
  • Kiểm tra giả định đắt nhất trước tiênTesting the most expensive assumption first
đượcyou gain
  • Học nhanh nhất trên mỗi giờ bỏ raThe fastest learning per hour available
  • Bỏ đi không tiếc vì chưa ai phụ thuộcFree to throw away, nobody depends on it
  • Không cần auth, giám sát hay quy trìnhNo auth, monitoring or process
mấtyou lose
  • Số liệu demo không phải số liệu thậtDemo numbers are not real numbers
  • Chưa gặp ca xấu vì không có xử lý lỗiNo error handling, so no bad cases seen yet
  • Demo tốt tạo áp lực deploy thẳngA good demo creates pressure to ship it as-is

Dấu hiệu:Recognise it: chỉ bạn chạy nó và bạn chấp nhận nó hỏng. you are the only one running it and you accept that it breaks.

vs

PRODUCTION

chạy thậtrun it for real

auth→evals→alerts

Câu hỏi chuyển từ “nó có chạy không” sang “nó còn đúng không, và tôi biết trước hay sau người dùng”. The question shifts from “does it work” to “is it still right, and do I find out before the users”.

dùng chouse for
  • 50 người trong công ty dùng hằng ngày50 people in the company using it daily
  • API mở cho bên ngoàiAn API exposed to the outside
  • Bất kỳ thứ gì tự động chạm dữ liệu khách hàngAnything that touches customer data automatically
đượcyou gain
  • Biết nó hỏng trước khi người dùng báoYou know it broke before users report it
  • Chi phí dự đoán được và có trầnCost is predictable and capped
  • Giới hạn được phạm vi thiệt hạiThe blast radius is bounded
mấtyou lose
  • Mỗi thay đổi tốn nhiều lần so với lúc prototypeEvery change costs several times the prototype price
  • Eval không bảo trì thì mục và cho cảm giác an toàn giảAn unmaintained eval set rots into false confidence
  • Cảnh báo quá nhiều thì không ai đọcToo many alerts means nobody reads them

Dấu hiệu:Recognise it: nó hỏng lúc hai giờ sáng và có người quan tâm. it breaks at 2am and somebody cares.

đánh đổithe trade Chọn prototype là đổi độ tin cậy để lấy tốc độ học. Choosing a prototype trades reliability for speed of learning.

đánh đổithe trade Chọn production là đổi tốc độ thay đổi để lấy quyền được người khác phụ thuộc vào mình. Choosing production trades speed of change for the right to have others depend on you.

Bảng đánh đổiTrade-off table

Đọc theo cột “sửa khi sai”: nó dự đoán nỗi đau sáu tháng tới chính xác hơn cột chi phí. Read down the “time to fix” column: it predicts your pain six months out better than cost does.

khái niệmconcept chi phí dựngcost to build chi phí mỗi lần chạycost per run sửa khi saitime to fix có biết khi hỏng?notice failures? hạ tầng thêmextra infra
Promptrất thấpvery lowthấplowgiâysecondskhôngnokhôngnone
Harnesscaohighthấp + retrylow + retriesgiờ → ngàyhours → dayscóyeslog, storagelogs, storage
Looprất thấpvery lownhân theo số vòngmultiplied per turnphútminuteschỉ khi có logonly with logsscheduler
Goalthấplowkhông đổiunchangedphút → giờminutes → hourscóyeskhôngnone
Contextkhôngnonecao, tăng theo độ dàihigh, grows with lengthtức thìimmediatekhôngnokhôngnone
Memorytrung bìnhmediumthấplowngàydayshiếm khirarelydatabase
RAGcaohightrung bìnhmediumphútminuteschỉ khi đo recallonly if you measure recallindex, job cập nhậtindex, refresh job
Fine-tunerất caovery highthấp nhất ở quy mô lớnlowest at volumengày → tuầndays → weekschỉ khi có evalonly with evalspipeline traintraining pipeline
Prototyperất thấpvery lowkhông quan trọngdoes not matterphútminuteskhông cầnnot requiredkhôngnone
Productionrất caovery highdự đoán đượcpredictablengàydayscó, theo thiết kếyes, by designauth, eval, alertauth, evals, alerts

Anti-patternsAnti-patterns

Mười cách quen thuộc để mất vài tháng. Ten familiar ways to lose a few months.

sai lầmmistake vì sao saiwhy it is wrong sửa thế nàothe fix
Dựng agent có loop trước khi có goal kiểm chứng đượcBuilding a looping agent before a checkable goal existsNó không biết lúc nào xong, chạy tới khi bạn nhìn hoá đơnIt cannot tell when it is done, so it runs until you see the billViết hàm xong() trước, rồi mới viết loopWrite the done() function first, then the loop
Fine-tune để model “thuộc” tài liệu công tyFine-tuning so the model “knows” the company docsTrọng số không chạy theo tài liệu; model chỉ bịa mượt hơnWeights do not track documents; the model just invents more smoothlyRAG cho sự thật, fine-tune chỉ cho hành viRAG for facts, fine-tuning only for behaviour
Nhét cả 200 trang vào context “cho chắc”Pasting all 200 pages into context to be safeĐắt, chậm, và phần giữa bị bỏ quaExpensive, slow, and the middle gets skippedTruy hồi và đưa vào 3–5 đoạn liên quanRetrieve and inject 3–5 relevant passages
Coi prompt là harness: “tôi đã dặn nó trả JSON”Treating a prompt as a harness: “I told it to return JSON”Đúng 97% nghĩa là 3% phá vỡ pipeline phía sau97% correct means 3% breaks the pipeline downstreamValidate schema, retry hai lần, rồi trả lỗiValidate the schema, retry twice, then error
Gọi là memory nhưng thực chất nối toàn bộ lịch sửCalling it memory when it is the whole history concatenatedĐó chỉ là context phình to; chi phí tăng tuyến tính rồi trànThat is a swelling context; cost grows linearly then overflowsTóm tắt có cấu trúc, chỉ recall phần liên quanStructured summaries, recall only what is relevant
Demo tốt nên deploy luônDeploying because the demo went wellDemo là n nhỏ đã chọn lọc, chưa gặp ca xấuA demo is a small cherry-picked n that never met a bad caseBộ eval 30–50 ca, chạy song song trước khi mở30–50 eval cases, shadow-run before opening it
Retry cho tới khi ra kết quả đẹpRetrying until the answer looks niceNhân chi phí và giấu mất lỗi thậtMultiplies cost and hides the real failureTrần hai lần, ghi lý do trượt, để nó failCap at two, log the reason, let it fail
Chia đoạn RAG cứng theo 500 ký tựChunking RAG at a fixed 500 charactersCắt giữa câu, giữa bảng, giữa đoạn mãIt cuts through sentences, tables and codeChia theo heading và ranh giới ngữ nghĩaChunk on headings and semantic boundaries
Cảnh báo mọi exceptionAlerting on every exceptionỒn quá nên không ai đọc, gồm cả cảnh báo thậtToo noisy to read, including the real onesCảnh báo theo tỷ lệ trượt và ngân sách mỗi giờAlert on failure rate and hourly budget
Đánh giá bằng “cảm giác câu trả lời hay”Evaluating by “the answers feel good”Không so sánh được giữa hai phiên bảnTwo versions cannot be comparedBộ eval cố định, chấm điểm, ghi lại theo thời gianA frozen eval set, scored and tracked

Ba câu cần nhớThree sentences to keep

Nếu chỉ giữ lại ba dòng từ trang này, giữ ba dòng sau. If you keep only three lines from this page, keep these.

  1. Một vòng lặp không có đích sẽ chạy cho tới khi bạn nhận ra — viết điều kiện dừng trước khi viết vòng lặp. A loop without a goal runs until you notice — write the stop condition before the loop.
  2. Context là cái model đọc lần này, memory là cái bạn ghi lại để đọc lần sau; memory chỉ có tác dụng khi được nạp ngược vào context. Context is what the model reads this time, memory is what you write down for next time; memory only matters once loaded back into context.
  3. Thiếu thông tin thì RAG, thiếu hành vi thì fine-tune; prompt cho một câu trả lời tốt, harness cho câu trả lời tốt mọi lần. Missing information means RAG, missing behaviour means fine-tuning; a prompt gets one good answer, a harness gets them every time.