📅 Tuesday, July 21 · Morning Session · Topic: AI & Machine Learning Vocabulary ⏱ ~20 minutes · 🎯 Intermediate
Welcome to Tuesday Morning — AI & Machine Learning day! 🤖 Today we focus on the vocabulary you need when building, debugging, or discussing LLM-powered systems. These words show up in design docs, code reviews, incident reports, and interviews. Let’s make them yours.
🌟 Word of the Day — hallucinate
/həˈluː.sɪ.neɪt/ 🇻🇳 ảo giác / bịa đặt — In AI context: when an LLM generates false or fabricated information with complete confidence, as if it were fact.
Why it matters: This is one of the most important concepts in production AI systems. If you work with LLMs, you will encounter hallucinations — and you need to explain them clearly to teammates, stakeholders, and clients.
3 example sentences:
- “The model hallucinated a citation that doesn’t exist — we need a grounding layer to prevent this in production.”
- “Our RAG pipeline significantly reduced hallucinations by anchoring answers to retrieved documents.”
- “When the LLM hallucinates, it doesn’t know it’s wrong — that’s what makes it so dangerous for critical applications.”
🔗 Cambridge Dictionary — hallucinate | YouGlish — hear it in real speech
📚 Vocabulary Table — 5 Key AI/ML Phrases
| Phrase | Vietnamese | Example Sentence |
|---|---|---|
| grounding | neo đậu thực tế | ”Grounding the model with external data reduces hallucinations.” |
| retrieval-augmented | tăng cường truy xuất | ”We use a retrieval-augmented approach to keep answers factual.” |
| inference pipeline | luồng suy luận | ”The inference pipeline processes 1,000 requests per second.” |
| context window | cửa sổ ngữ cảnh | ”Our model has a 200K-token context window.” |
| fine-tuned | tinh chỉnh | ”This model was fine-tuned on medical records.” |
🎤 Pronunciation Guide — “hallucinate”
Break it down syllable by syllable:
| Syllable | Sound | Tip |
|---|---|---|
| hal- | /həl/ | Short, relaxed — like the “hul” in “hull” |
| -lu- | /ˈluː/ | Stressed syllable — long “loo” sound, say it louder |
| -ci- | /sɪ/ | Short “si” — like “sit” without the -t |
| -nate | /neɪt/ | “nayt” — rhymes with “late” |
✅ Correct: huh-LOO-si-nayt
❌ Common mistake: HAL-loo-si-nayt (wrong stress on first syllable)
🗣️ Practice Sentence — Read Aloud 3×
“The hallucinating model confidently stated false facts about our inference pipeline.”
Read it slowly once, then at normal speed twice. Focus on stressing the second syllable of huh-LOO-ci-nate.
✏️ Exercise 1 — Fill in the Blank
Choose the correct word or phrase: hallucinate / grounding / retrieval-augmented / inference pipeline / context window / fine-tuned
- “We deployed a __________ generation system so the chatbot only answers based on our internal knowledge base.”
- “The model started to __________ when asked about events after its training cutoff — it invented news stories.”
- “Adding __________ to the prompt reduced false answers by 40% in our A/B test.”
- “The __________ is optimized for low latency — we cache embeddings to avoid recomputing them every request.”
- “Our 200K-token __________ lets the model analyze an entire codebase in a single call.”
- “We __________ the base model on our company’s support tickets to make it more domain-specific.”
✅ Click to reveal answers
- retrieval-augmented
- hallucinate
- grounding
- inference pipeline
- context window
- fine-tuned
✏️ Exercise 2 — Translate to English
Translate these sentences into natural English using today’s vocabulary:
- “Mô hình đã bịa đặt một câu trả lời rất tự tin — chúng ta cần cơ chế neo đậu thực tế.”
- “Luồng suy luận của chúng ta xử lý 5,000 yêu cầu mỗi phút mà không bị trễ.”
- “Mô hình được tinh chỉnh trên dữ liệu pháp lý hoạt động tốt hơn nhiều so với mô hình gốc.”
✅ Click to reveal suggested answers
- “The model hallucinated a very confident answer — we need a grounding mechanism.”
- “Our inference pipeline handles 5,000 requests per minute without latency issues.”
- “The model fine-tuned on legal data performs much better than the base model.”
💡 Idiom of the Day — “garbage in, garbage out"
"Garbage in, garbage out” (GIGO)
/ˈɡɑːrbɪdʒ ɪn ˈɡɑːrbɪdʒ aʊt/ 🇻🇳 Dữ liệu xấu vào, kết quả xấu ra — If the input data is poor quality, the output will be equally poor, no matter how sophisticated the system is. Originally a computing maxim, now widely used in AI/ML.
2 examples:
- “We spent three months on model tuning before realizing the root issue was data quality — classic garbage in, garbage out.”
- “The prompt was vague and unstructured, so the LLM hallucinated a useless answer. Garbage in, garbage out — always applies.”
📺 Recommended Watching
- Andrej Karpathy — Intro to Large Language Models (1 hour, worth every minute) — The clearest explanation of how LLMs work, including hallucinations. Karpathy’s pacing is perfect for English learners.
- 3Blue1Brown — But what is a GPT? Visual intro to transformers (27 min) — Visual, intuitive explanation of the architecture behind every major LLM. Great vocabulary in context.
- Fireship — RAG in 100 Seconds (2 min) — Fast, dense, and covers retrieval-augmented generation perfectly. Watch it three times.
🎯 Daily Challenge
Ask an AI chatbot 3 questions today, and try to spot one hallucination — then fact-check it.
Suggested questions to try:
- “What papers did [a specific researcher] publish in 2024?”
- “What is the exact API endpoint for [a service you know well]?”
- “Summarize the changelog of [a library] version X.Y.Z.”
When you find a hallucination, try describing it in English using today’s vocabulary:
- “The model hallucinated a paper title — it doesn’t exist.”
- “There was no grounding — the answer was completely fabricated.”
📊 Today’s Vocabulary at a Glance
| Term | IPA / Pronunciation | Vietnamese |
|---|---|---|
| hallucinate | /həˈluː.sɪ.neɪt/ | ảo giác / bịa đặt |
| grounding | /ˈɡraʊndɪŋ/ | neo đậu thực tế |
| retrieval-augmented | /rɪˈtriːvəl ˈɔːɡmentɪd/ | tăng cường truy xuất |
| inference pipeline | /ˈɪnfərəns ˈpaɪplaɪn/ | luồng suy luận |
| context window | /ˈkɒntekst ˈwɪndoʊ/ | cửa sổ ngữ cảnh |
| fine-tuned | /faɪn tjuːnd/ | tinh chỉnh |
🌅 Great start to Tuesday! The terms you learned today — especially hallucinate and grounding — will make you sound fluent and credible in any AI engineering conversation. Come back at noon for Communication & Phrases! 💪