Tuesday Morning: AI & Machine Learning Vocabulary 🌅
Good morning! Tuesday’s theme is AI & Machine Learning — the vocabulary you need when you’re explaining an LLM feature to a non-technical stakeholder, debugging a RAG pipeline out loud with a teammate, or just talking about your work at a meetup. Today’s word is one every engineer says a dozen times a day but rarely says clearly: inference.
🌟 Word of the Day: Inference
IPA Pronunciation: /ˈɪnfərəns/
Vietnamese: quá trình model đưa ra dự đoán/kết quả từ input (khác với training — lúc này model không học thêm, chỉ áp dụng những gì đã học)
3 Example Sentences:
- “Training this model took three days, but inference only takes 200 milliseconds per request.”
- “We moved inference to a smaller quantized model to cut our GPU costs in half.”
- “The latency spike wasn’t in the network layer — it was inference time on the larger model.”
Learn more:
- Cambridge Dictionary — inference
- YouGlish — hear “inference” in real videos
- What is Inference? — IBM Technology (YouTube)
📚 Vocabulary Table: AI & Machine Learning
| Phrase | Vietnamese | Example |
|---|---|---|
| hallucinate | model tạo ra thông tin sai nhưng nghe có vẻ đúng | ”The model hallucinated a function that doesn’t exist in our codebase — always verify generated code before merging.” |
| fine-tune | huấn luyện thêm một model có sẵn trên dữ liệu chuyên biệt | ”We fine-tuned the base model on our support tickets so it understands our product terminology.” |
| latency | độ trễ, thời gian chờ giữa request và response | ”Users start noticing latency above 300 milliseconds — that’s our budget for the whole pipeline.” |
| retrieval | truy xuất thông tin liên quan trước khi model tạo câu trả lời (phần “R” trong RAG) | “Retrieval quality matters more than prompt engineering — if you fetch the wrong documents, no prompt will save you.” |
| context window | giới hạn lượng văn bản model có thể “nhìn thấy” cùng lúc | ”We hit the context window limit, so we had to summarize the earlier conversation before continuing.” |
🗣️ Pronunciation Guide
Let’s break down inference /ˈɪnfərəns/ syllable by syllable:
- in- → /ɪn/ — short “i,” like in “bin”
- -fer- → /fər/ — unstressed, almost swallowed — sounds like “fer” in “offer”
- -ence → /əns/ — soft ending, don’t over-pronounce the final “e”
⚠️ Common mistake: Vietnamese speakers often stress the second syllable (“in-FER-ence”) — but native speakers put the stress on the first syllable: “IN-fer-ence.” The rest trails off quickly.
Practice sentence — read this out loud 3 times:
“Our inference latency dropped after we fine-tuned a smaller model for retrieval.”
Say it slowly the first time, focus on stress the second time, then read it at natural speed the third time.
✍️ Exercises
Exercise 1: Fill in the blank
- “The chatbot ______ a customer’s account number — it just made it up.” (hallucinated)
- “We need to ______ this model on our internal documentation before launch.” (fine-tune)
- “Our ______ pipeline fetches the top 5 documents before the model generates an answer.” (retrieval)
Show Answers
- hallucinated
- fine-tune
- retrieval
Exercise 2: Translate to English
- “Việc training mất ba ngày, nhưng inference chỉ mất vài trăm mili giây.”
- “Model bị hallucinate khi chúng tôi hỏi về một tính năng chưa tồn tại.”
- “Chúng tôi đã fine-tune model để giảm độ trễ (latency).”
Show Answers
- “Training took three days, but inference only takes a few hundred milliseconds.”
- “The model hallucinated when we asked about a feature that doesn’t exist.”
- “We fine-tuned the model to reduce latency.”
💬 Idiom of the Day: “Under the hood”
Vietnamese meaning: bên trong, cơ chế vận hành thực sự (thường dùng khi nói về phần kỹ thuật ẩn sau một giao diện đơn giản)
Examples:
- “The chatbot looks simple to users, but under the hood it’s running retrieval, reranking, and three separate model calls.”
- “Don’t worry about what’s under the hood for this demo — just show the product manager the final output.”
📺 Recommended Watching
- IBM Technology — Inference vs. Training explained (YouTube)
- YouGlish: “inference” in real tech talks
- 3Blue1Brown — How LLMs actually work (YouTube)
🎯 Today’s Challenge
Next time you talk about an AI feature at work — in standup, in a PR description, anywhere — say “inference” and “latency” out loud instead of writing them silently. Stress the first syllable of inference: “IN-fer-ence.” Speaking technical words before you type them is one of the fastest ways to make pronunciation stick.
See you this evening for practice! 🚀