Tuesday Evening: AI Vocabulary Review + Listening Comprehension Strategies

Tonight you review everything from today — morning AI/ML vocabulary and noon’s RAG pipeline terms. You’ll also learn how to understand fast-spoken technical English in podcasts and conference talks. The goal: don’t just recognize these words in text. Hear them.

🌟 Word of the Day: Context Window

Pronunciation/ˈkɒntɛkst ˈwɪndəʊ/
Vietnamesecửa sổ ngữ cảnh — lượng thông tin mà model AI có thể “nhìn thấy” và xử lý cùng một lúc
StressCON-text WIN-dow

🔊 Pronunciation links: Cambridge - context · YouGlish - context window

3 real-world examples:

  1. “Our model has a 200k token context window — it can process an entire codebase in one pass.”
  2. “The problem with long calls to the LLM isn’t speed, it’s that we’re filling the context window with irrelevant history.”
  3. “RAG works around the context window limitation by only loading the relevant chunks instead of the full document.”

Why it matters: Context window is the #1 constraint in practical LLM applications. If you can explain it clearly, you understand why RAG, chunking, and reranking exist.

📋 Today’s Full AI Vocabulary Review

Morning Session Words (AI/ML Fundamentals)

WordVietnameseUsage Example
inference /ˈɪnfərəns/suy luận — chạy model để ra kết quả”Inference latency is 120ms — acceptable for our use case.”
fine-tune /ˈfaɪn tjuːn/tinh chỉnh model trên dữ liệu của mình”We fine-tuned the base model on our support tickets.”
embedding /ɪmˈbɛdɪŋ/vector số biểu diễn ý nghĩa của văn bản”Each document is converted to an embedding before storage.”
hallucinate /həˈluːsɪneɪt/model bịa thông tin không có trong training data”The model hallucinated a function that doesn’t exist in our SDK.”

Noon Session Words (RAG Pipeline)

WordVietnameseUsage Example
grounding /ˈɡraʊndɪŋ/neo đậu vào dữ liệu thực — ngăn hallucination”Grounding improved our answer accuracy from 60% to 92%.“
chunk /tʃʌŋk/đoạn văn bản nhỏ để embed và tìm kiếm”We chunk documents into 512-token segments.”
rerank /ˌriːˈræŋk/sắp xếp lại kết quả theo relevance thực”After retrieval, we rerank with a cross-encoder model.”
semantic search /sɪˈmæntɪk sɜːtʃ/tìm kiếm theo ý nghĩa, không chỉ từ khóa”Semantic search finds related content even with different wording.”

🎧 Listening Comprehension: How to Understand Technical English Podcasts

This is the skill nobody teaches you. You can read technical English well, but when a native speaker talks at 160 words per minute in a podcast, you catch maybe 60%. Here’s why and how to fix it.

The 3 Reasons You Miss Words

1. Connected speech. Native speakers don’t say words separately. They run them together:

  • “contextwindow” → sounds like one word: “CONtextWINdow”
  • “workingonit” → sounds like: “WORKinonit”
  • “didyou” → sounds like: “didja”

2. Reduced vowels. Unstressed syllables collapse to /ə/ (schwa):

  • “inference” → “INF-rəns” (not “IN-fer-ence”)
  • “production” → “prə-DUC-shən”
  • “a model” → “ə MOD-əl”

3. Speed + technical density. A speaker says “retrieval-augmented generation with cross-encoder reranking” at 180 WPM. You’ve never heard it spoken — you only read it.

The Fix: Active Listening Protocol

Step 1 — Pre-load the vocabulary. Before listening to an AI podcast, spend 3 minutes reviewing the key terms out loud. Tonight you reviewed them above — you’re already primed.

Step 2 — Listen twice. First pass: get the main idea, don’t stress about every word. Second pass: focus on the sentences you missed.

Step 3 — Shadow the host. Pause after each sentence. Repeat it out loud, matching their rhythm — not just the words.

Step 4 — Slow it down. Most podcast apps let you set 0.75x speed. Use it when a section is dense. This isn’t cheating — it’s training.

Podcast / TalkWhy It’s GoodSpeed
Lex Fridman (AI episodes)Clear diction, deep technical1.0x
Latent Space podcastReal AI engineering, fast0.85x first time
Anthropic / OpenAI blog posts read aloudMatches your reading vocab1.0x
Google DeepMind YouTubeBritish + American accents0.9x

🗣️ Pronunciation Practice

Today’s challenge sentence (say this 5 times, each time a little faster):

“We use retrieval-augmented generation with semantic search and cross-encoder reranking to ground the model in our verified knowledge base.”

Stress map:

  • re-TRIE-val-aug-MEN-ted gen-er-A-tion → stress on TRIE and MEN and A
  • se-MAN-tic SEARCH → stress on MAN and SEARCH
  • CROSS-en-coder RE-rank-ing → stress on CROSS and RE
  • GROUND the MOD-el → stress on GROUND and MOD

Connected speech tip:

  • “retrieval-augmented” → say as one compound: “retrieval-AUGMENTED”
  • “to ground the model in” → “to-GROUND-the-model-in” runs together

✏️ Exercise 1: Fill in the Blank (Review)

Choose from: context window, grounding, fine-tune, embeddings, hallucinate, chunk, rerank

  1. “We had to _______ the model on our proprietary data — the base model didn’t know our product vocabulary.”
  2. “Without _______, the assistant would confidently answer questions about features that don’t exist yet.”
  3. “We _______ each PDF into 400-token pieces before creating _______.”
  4. “After retrieval, we _______ the results to pick the most relevant 5 for the context.”
  5. “The model started to _______ because we filled its _______ with too much irrelevant history.”
✅ Answers
  1. fine-tune
  2. grounding
  3. chunk / embeddings
  4. rerank
  5. hallucinate / context window

✏️ Exercise 2: Translate to English

  1. “Chúng tôi dùng semantic search để tìm các chunks liên quan, sau đó rerank trước khi đưa vào context window của model.”
  2. “Vấn đề là model đang hallucinate — nó không được grounding vào tài liệu thực tế của chúng tôi.”
  3. “Sau khi fine-tune, embeddings của model phản ánh chính xác hơn thuật ngữ nội bộ của công ty.”
✅ Suggested Answers
  1. “We use semantic search to find relevant chunks, then rerank them before loading into the model’s context window.”
  2. “The problem is the model is hallucinating — it’s not grounded in our actual documentation.”
  3. “After fine-tuning, the model’s embeddings more accurately reflect the company’s internal terminology.”

💡 Idiom of the Day: “In the weeds”

Vietnamese: Bị sa vào chi tiết quá mức — mất cái nhìn tổng quan vì tập trung quá vào từng bước nhỏ

Tech context examples:

  • “I got in the weeds trying to optimize the reranking threshold — and missed the bigger issue that our chunking strategy was wrong.”
  • “Don’t get in the weeds with token counting in the demo — just show the stakeholders what it can do.”

In a meeting: If a technical discussion gets too deep for non-technical attendees: “Let’s not get in the weeds here — I can take the details offline.”

💬 Speaking Challenge (60 seconds)

Set a timer for 60 seconds and answer this question out loud:

“Explain to a new junior engineer — in plain English — what RAG is, why we need grounding, and what happens if we don’t have it.”

Use at least 5 words from today’s vocabulary. Don’t read from notes — just talk.

Suggested structure:

  1. What RAG stands for and does (10 seconds)
  2. What grounding means and why it matters (20 seconds)
  3. What happens without it — hallucination example (20 seconds)
  4. How your team solved it (10 seconds)

No perfect answer. The goal is to say it fluently and confidently, as you would in a real onboarding conversation.

🌙 Evening Challenge

Before you go to sleep tonight, do ONE of these:

Option A — Voice memo: Record yourself explaining what RAG is in 30 seconds. Play it back. Notice your rhythm and confidence.

Option B — Dictionary lookup: Go to Cambridge Dictionary and listen to the pronunciation of 3 words from today: inference, embedding, retrieval. Then say each one out loud.

Option C — Shadow drill: Find any sentence in an AI article and read it out loud 3 times, each time a little faster.

Any one of these takes under 2 minutes. Done consistently, they add up faster than any app or course.

Tomorrow: Wednesday Morning — Meetings & Communication (how to interrupt politely, check understanding, and ask for clarification in technical discussions) 🎤

Export for reading

Comments