Tuesday Evening: Review + Listening Strategies
Evening is the consolidation phase. The vocabulary you’ve seen today — inference, hallucination, RAG, token, LLM, grounding, fine-tuning, context window, temperature, system prompt — will stick much better if you revisit them once more before sleep. This is how memory works: spaced repetition within a single day is one of the most effective learning techniques.
Tonight’s session has two parts: a vocabulary review, and practical strategies for understanding fast, native-speed technical English in real contexts.
Word of the Day
embeddings /ɪmˈbed.ɪŋz/ — vector nhúng (biểu diễn số của từ hoặc khái niệm)
Pronunciation breakdown:
- ɪm — short “i” + “m”
- BED — stressed syllable, like “bed”
- ɪŋz — ends with “-ingz”
Full word: im-BED-ingz (stress on BED)
What it means: Embeddings are numerical representations of text. When a model “understands” a sentence, it converts it into a list of numbers (a vector) that captures its meaning. Similar concepts end up with similar numbers, which is why semantic search works.
Examples:
“We store customer support tickets as embeddings so the search system can find semantically similar issues.”
“The RAG pipeline converts your query into an embedding, then finds the most similar embeddings in our knowledge base.”
“Embedding models are separate from generation models — they’re optimized for similarity, not output.”
Explaining it simply: “Think of embeddings like coordinates on a map. Words with similar meanings are placed close together. When you search, the system finds coordinates near your query.”
Vietnamese tip: “Vector nhúng” là bản dịch chính xác, nhưng trong giao tiếp kỹ thuật hàng ngày, hãy dùng “embeddings” trực tiếp — người nghe sẽ hiểu ngay. Số nhiều (embeddings) thường dùng hơn số ít (embedding).
Vocabulary Review Table
Here are all 10 AI terms from today — morning and noon sessions combined. Test yourself: cover the “Plain English” column and try to explain each term before revealing it.
| Term | IPA | Plain English (cover and guess first!) |
|---|---|---|
| inference | /ˈɪn.fər.əns/ | Running a trained model on new input to get output |
| LLM | /ˌel.el.ˈem/ | Large Language Model — a neural network trained on massive text data |
| prompt | /prɒmpt/ | The text instruction you give the AI model |
| token | /ˈtoʊ.kən/ | Smallest unit of text the model processes (~¾ of a word) |
| RAG | /ræɡ/ | Retrieving relevant documents and adding them to the model’s context |
| agent | /ˈeɪ.dʒənt/ | An AI system that takes autonomous actions to complete multi-step tasks |
| hallucination | /həˌluː.sɪˈneɪ.ʃən/ | When a model generates false information stated with false confidence |
| fine-tuning | /ˈfaɪn.tjuːnɪŋ/ | Training a general model further on specific data for a specific task |
| grounding | /ˈɡraʊndɪŋ/ | Connecting AI output to verifiable, factual sources |
| context window | /ˈkɒntekst ˈwɪndəʊ/ | The total text the model can process at once in one conversation |
| embeddings | /ɪmˈbed.ɪŋz/ | Numerical vector representations of text that capture semantic meaning |
Pronunciation Practice
Say this sentence aloud. It contains 7 of today’s words. Aim for smooth, connected speech — not word-by-word:
“Our RAG pipeline converts the user’s prompt into embeddings, retrieves relevant documents within the context window, and passes everything to the LLM for inference — which dramatically reduces hallucination compared to a raw system prompt alone.”
First pass: Say it slowly, clearly. Second pass: Say it at normal conversation speed. Third pass: Record it on your phone. Listen back. Where did you hesitate?
The goal is not speed — it’s fluency. Fluency means not pausing to search for the word.
Exercise 1: The Explanation Challenge
Explain RAG to someone who knows nothing about AI. Use plain English, no jargon. Maximum 3 sentences.
Model answer
“RAG stands for Retrieval-Augmented Generation. Before the AI writes a response, it searches through a collection of documents — like your company’s help center — and pulls the most relevant pieces. Then it uses those real documents to write its answer, so the response is based on actual facts rather than what the model guessed.”
Key technique: Start with the acronym expansion, explain the process step by step, end with the “why it matters” (reduces hallucination).
Exercise 2: Pronunciation Rapid Fire
Say each word or phrase 3 times fast. Focus on clarity, not speed.
- inference — IN-fər-əns, IN-fər-əns, IN-fər-əns
- hallucination — ha-LOO-ci-NA-tion, ha-LOO-ci-NA-tion, ha-LOO-ci-NA-tion
- embeddings — im-BED-ingz, im-BED-ingz, im-BED-ingz
- fine-tuning — FINE-TOO-ning, FINE-TOO-ning, FINE-TOO-ning
- context window — CON-text WIN-doh, CON-text WIN-doh, CON-text WIN-doh
After the rapid fire, say this sentence naturally:
“Fine-tuning reduces hallucination better than prompt engineering, but only if your embeddings and context window are well-designed.”
Listening Strategies for Technical English
Learning AI vocabulary in English is one thing. Hearing it spoken fast by native speakers in meetings, podcasts, or YouTube videos is another challenge. Here are the strategies that actually work:
1. Listen for stressed syllables, not every sound
Native speakers compress unstressed syllables. “Inference” sounds like “INf-rence” — the middle syllable nearly disappears. Train your ear to catch the stressed syllable first and fill in the rest.
2. Use podcasts with transcripts
Recommended: The TWIML AI Podcast (twimlai.com) — every episode has a transcript. Listen for 5 minutes, pause, read the transcript, listen again. Repeat.
3. Watch AI demo videos on mute, then with sound
Watch a 3-minute AI demo (Anthropic’s Claude demos, OpenAI releases) on mute first. Predict the vocabulary. Then watch with sound. Notice which words you recognized vs. missed.
4. Repeat after experts
Find a 30-second clip of a technical explanation — Karpathy, Sutskever, or any Anthropic/OpenAI announcement. Repeat each sentence exactly as they say it: same rhythm, same intonation, same stress.
5. The “shadowing” technique for AI terms
When you hear an AI term you recognize, whisper it silently at the same time as the speaker. This builds the muscle memory for automatic recognition.
Idiom of the Day
”state of the art”
Meaning: The most advanced, current best-in-class version of something.
In AI context:
“GPT-5.6 Sol is currently state-of-the-art on the Agents’ Last Exam benchmark.”
“Our current model was state-of-the-art six months ago — it’s already been surpassed by three newer releases.”
Pronunciation: STATE-uv-thuh-ART — often said quickly as “state-of-the-art” in one flow.
Vietnamese: “Hiện đại nhất,” “đỉnh cao hiện tại.” But use “state of the art” directly in tech English — it’s a standard compound adjective.
60-Second Speaking Challenge
Set a timer for 60 seconds. Record yourself explaining: “What is a RAG system and why does it reduce hallucination?”
Use these words: RAG, embeddings, context window, grounding, hallucination.
When the timer stops, listen back. Count how many of the 5 target words you used. Aim for all 5.
If you get all 5 naturally and the explanation sounds clear, you’ve mastered today’s vocabulary at the speaking level — not just recognition.
Evening Challenge
Before bed tonight, try to use one AI vocabulary word from today in a real context — a Slack message, a code comment, a PR description, or a conversation. Even something small:
“Fixed the prompt to reduce hallucination on edge cases.” “Added retrieval step — implementing basic RAG before we fine-tune.”
Using words in real output is the final step that converts passive recognition into active vocabulary you’ll use automatically.
Tomorrow: we continue with more technical English topics. Good work today.