Good morning! Today we’re talking about how to describe the AI systems most of us build with every day — not the model itself, but everything around it: how you prompt it, how you ground it in real data, and what happens when it goes wrong.

📖 Word of the Day: Hallucinate

  • Pronunciation (IPA): /həˈluː.sɪ.neɪt/
  • Vietnamese meaning: (AI) tự bịa ra thông tin sai mà vẫn nói như thật; “ảo giác” (thông tin do AI tạo ra không có thật)
  • Example sentences:
    1. “The model hallucinated a function name that doesn’t exist in our codebase.”
    2. “We added a retrieval step specifically to stop the assistant from hallucinating dates.”
    3. “If the answer isn’t grounded in a source document, always double-check — it might be hallucinating.”

🔗 Cambridge Dictionary: hallucinate 🔗 YouGlish: hear “hallucinate” in real videos

🧩 Vocabulary Table

PhraseVietnameseExample
Ground (v.) — to base an answer on real, verifiable datacăn cứ vào dữ liệu thật”We ground every response in the retrieved documents, not the model’s memory.”
Retrieve (v.) — to fetch relevant information from a knowledge sourcetruy xuất thông tin”The system retrieves the top 5 matching chunks before generating an answer.”
Chunk (n./v.) — a small piece of text split from a larger document for retrievalđoạn văn bản nhỏ (được chia ra)“We chunk the PDF into 500-token pieces before indexing it.”
Prompt template — a reusable structure for instructing a modelmẫu câu lệnh (prompt)“Our prompt template always includes the user’s role and the task context.”
Context window — the amount of text a model can “see” at oncecửa sổ ngữ cảnh”That document is too long to fit in the model’s context window.”

🗣️ Pronunciation Guide

Word of the Day breakdown: hallucinate — /həˈluː.sɪ.neɪt/

  • Syllables: ha-LU-ci-nate (4 syllables, stress on the second syllable)
  • hə — soft “uh” sound, very light, almost swallowed
  • LUː — the stressed syllable, long “oo” sound like in “blue” — this is where your voice goes up
  • sɪ — short, quick “sih”
  • neɪt — rhymes with “gate” or “late”

Common mistake: Vietnamese speakers often stress the first syllable (“HA-lu-ci-nate”). Push the stress to the second syllable instead — say “luuu” a little louder and longer than the rest.

Practice sentence (read aloud 3 times):

“If the model hallucinates, we ground the next prompt in retrieved data instead of trusting its memory.”

✍️ Exercises

Exercise 1 — Fill in the blank:

  1. Before answering, the system will _______ the top matching documents from our knowledge base. (retrieve / hallucinate)
  2. We split the manual into smaller _______ so the retrieval step can find the exact paragraph we need. (chunks / windows)
  3. The model _______ a customer name that was never in the input — we caught it in review. (grounded / hallucinated)
Show answers
  1. retrieve
  2. chunks
  3. hallucinated

Exercise 2 — Translate to English:

  1. “Chúng tôi chia tài liệu thành các đoạn nhỏ trước khi lập chỉ mục.”
  2. “Câu trả lời này không có căn cứ — model có thể đang bịa thông tin.”
Show answers
  1. “We chunk the document into smaller pieces before indexing it.”
  2. “This answer isn’t grounded — the model might be hallucinating.”

💬 Idiom of the Day: “Take it with a grain of salt”

  • Vietnamese meaning: đừng vội tin hoàn toàn, hãy nghi ngờ một chút — dùng khi thông tin có thể không chính xác
  • Example 1: “The demo output looked impressive, but take it with a grain of salt — we hadn’t grounded it in real data yet.”
  • Example 2: “Take any ungrounded AI answer with a grain of salt until you’ve verified the source.”

📺 Recommended Watching

  1. YouGlish — search any word from today’s lesson to hear it in real, unscripted videos: youglish.com
  2. Rachel’s English — clear, detailed breakdowns of American pronunciation patterns, useful for tricky multi-syllable words like “hallucinate”: YouTube: Rachel’s English
  3. AssemblyAI YouTube channel — short, plain-English explainers on RAG, retrieval, and grounding concepts, good for picking up the vocabulary in context: YouTube: AssemblyAI

🎯 Today’s Challenge

Next time you review an AI-generated answer at work, describe it out loud in one sentence using at least two of today’s words — for example: “This answer isn’t grounded in our docs, so it might be hallucinating.” Say it before you type it.

See you tomorrow morning! 🌅

Export for reading

Comments