Good morning! It’s Tuesday — AI day. Today we cover the vocabulary you need to talk about language models, agents, and AI systems like a professional. These are real words your teammates use in standups, Slack threads, and design docs. By the end of this session, you’ll use them naturally.
🌟 Word of the Day: Hallucinate
IPA Pronunciation: /həˈluːsɪneɪt/
Vietnamese meaning: Ảo giác / Bịa đặt thông tin (nói về AI tạo ra thông tin sai)
In AI, to hallucinate means an LLM generates information that sounds confident and fluent but is factually incorrect, invented, or unsupported by its training data. The model doesn’t “know” it’s wrong — it just produces the most statistically likely next token, even when that leads to false claims.
“The model hallucinated a citation — the paper doesn’t exist, but it gave us a real-looking author name and journal title.”
🗣️ 3 Example Sentences
- “We had to add grounding to our RAG pipeline because the model kept hallucinating details about our product that weren’t in the knowledge base.”
- “Never trust an LLM’s output for legal or medical content without verification — they hallucinate with complete confidence.”
- “The new Opus 5 model hallucinates significantly less on factual queries than earlier versions, but it still happens on niche topics.”
🔗 Practice Links
- Cambridge Dictionary: https://dictionary.cambridge.org/dictionary/english/hallucinate
- YouGlish (hear native speakers): https://youglish.com/pronounce/hallucinate/english
📋 AI & Machine Learning Vocabulary Table
| Phrase | Vietnamese | Example Sentence |
|---|---|---|
| inference | suy luận / chạy mô hình | ”Inference costs 15x more than training — most of our AI budget is actually inference.” |
| token budget | giới hạn token | ”Keep the system prompt short — every token counts against our token budget per request.” |
| grounding | neo vào dữ liệu thực | ”We ground the model by injecting the user’s actual account data into every prompt.” |
| RAG | Retrieval-Augmented Generation | ”Our chatbot uses RAG — it retrieves relevant docs before generating the answer.” |
| fine-tune | tinh chỉnh mô hình | ”We fine-tuned GPT-4o on our support tickets to make it understand our product’s terminology.” |
🎯 Pronunciation Guide
Word of the Day Breakdown: hallucinate
| Syllable | Sound | Stress |
|---|---|---|
| hal | /həl/ | unstressed — soft “hul” |
| LU | /ˈluː/ | primary stress — long “loo” like “blue” |
| ci | /sɪ/ | unstressed — short “si” like “sit” |
| nate | /neɪt/ | ends with “nayt” — like “late” |
Full word: hul-LOO-si-nayt
Common mistakes for Vietnamese speakers:
- ❌ “hal-lu-si-NATE” (stress on the last syllable)
- ❌ “ha-LU-si-net” (wrong final vowel)
- ✅ “hul-LOO-si-nayt” — stress the second syllable strongly, end with “-nayt”
Related forms:
- hallucination /həˌluːsɪˈneɪʃən/ — “Hallucination is one of the biggest challenges in deploying LLMs for production use.”
- hallucinating /həˈluːsɪneɪtɪŋ/ — “The model is hallucinating product names that don’t exist in our catalog.”
🔊 Practice Sentence (read aloud 3x)
“To reduce hallucination in production, our team combined retrieval-augmented generation with explicit grounding instructions in the system prompt.”
Break it down:
- to re-DUCE hal-LU-ci-NA-tion
- in PRO-duc-tion
- our TEAM com-BINED re-TRIEV-al-AUG-ment-ed gen-er-A-tion
- with ex-PLIC-it GROUND-ing in-STRUC-tions
✏️ Exercise 1: Fill in the Blank
Use: (hallucinate / inference / token budget / grounding / RAG / fine-tune)
- “The API call is slow because ______ on large models takes hundreds of milliseconds per request.” (chạy mô hình)
- “We use ______ to connect the LLM to our internal knowledge base so it answers questions about our specific product.” (RAG)
- “Our system prompt is 4,000 tokens — we need to reduce it to stay within the ______.” (giới hạn token)
- “The model started ______ customer names that don’t appear anywhere in the data it was given.” (bịa đặt thông tin)
- “We ______ the model on 10,000 internal support tickets to teach it our company’s specific vocabulary.” (tinh chỉnh)
- “We added ______ by injecting the user’s actual order history into each prompt to prevent the model from inventing purchase details.” (neo vào dữ liệu)
✅ Answers
- inference
- RAG
- token budget
- hallucinating
- fine-tuned
- grounding
✏️ Exercise 2: Translate to English
Translate these Vietnamese AI/tech sentences into natural English:
- “Mô hình đang ảo giác tên sản phẩm không tồn tại trong cơ sở dữ liệu của chúng ta.”
- “Chi phí suy luận tốn kém hơn nhiều so với training — hầu hết ngân sách AI của chúng ta là chi phí chạy mô hình.”
- “Chúng tôi dùng RAG để model có thể trả lời câu hỏi về tài liệu nội bộ của công ty.”
- “Giữ system prompt ngắn gọn — mỗi token đều tính vào ngân sách token của chúng ta.”
✅ Sample Answers
- “The model is hallucinating product names that don’t exist in our database.”
- “Inference is far more expensive than training — most of our AI budget is actually inference cost.”
- “We use RAG so the model can answer questions about the company’s internal documents.”
- “Keep the system prompt short — every token counts against our token budget.”
💡 Idiom of the Day: “Garbage in, garbage out”
Meaning: The quality of a system’s output depends entirely on the quality of its input. If you give an AI bad data, a poorly written prompt, or missing context, you’ll get bad results — no matter how powerful the model.
Vietnamese: Đầu vào rác thì đầu ra cũng rác / Chất lượng đầu ra phụ thuộc vào chất lượng đầu vào
Origin: A classic computer science principle (often abbreviated GIGO) dating to the 1960s, now more relevant than ever in the age of LLMs. The saying captures a universal truth: computational tools amplify what you put in, they don’t fix it.
2 Real Examples in AI:
- “The model’s recommendations were terrible in our first demo. Classic garbage in, garbage out — we were giving it unfiltered, noisy user feedback with no structure.”
- “Good prompt engineering is how you avoid garbage in, garbage out. Spend more time on what you ask the model than on which model you pick.”
When to use it:
- Explaining why an AI feature performed poorly in testing
- Arguing for investment in data quality or prompt design
- Warning a stakeholder that switching models won’t fix a data quality problem
📺 Recommended Watching
For AI & LLM English Vocabulary
-
Andrej Karpathy (YouTube) — Former Tesla AI Director and OpenAI founding member. His videos explain neural networks, LLMs, and AI concepts in deep but accessible English. Watch “Let’s build GPT from scratch” for the most authentic technical AI English you’ll find anywhere. → Search: “Andrej Karpathy youtube”
-
3Blue1Brown — Neural Networks series — Beautiful visual explanations of how neural networks work, narrated in very clear, precise English with a measured pace that’s ideal for non-native speakers absorbing technical content. → Search: “3blue1brown neural network”
-
Latent Space Podcast — A podcast where two AI engineers discuss real production AI systems, new research, and the business of LLMs. Advanced English, but very authentic — this is how AI engineers actually talk about their work. Great for listening practice once you know the core vocab. → Search: “Latent Space podcast youtube”
🎯 Tuesday Challenge: Explain Hallucination in Your Own Words
The specific practice challenge:
Before your first meeting today, write 2-3 sentences in English explaining what LLM hallucination is, as if you were explaining it to a non-technical colleague or product manager.
Example:
“Language models sometimes generate information that sounds completely confident but is factually wrong — we call this hallucination. The model doesn’t know it’s incorrect; it’s just predicting the most likely next word based on patterns it learned. That’s why we always need to verify AI output for anything factual, legal, or financial.”
Write your own version, then say it out loud once. This is one of the most common AI explanations you’ll give in your career — getting it smooth and natural in English now means you’ll never stumble when a stakeholder asks “wait, what does it mean when the AI hallucinates?”
Bonus: Use the idiom — “This is a classic garbage in, garbage out problem — if we don’t ground the model with real data, we’ll keep seeing hallucinations.”
🗓️ Coming Up This Week
- Wednesday (Architecture): Bulkhead, blast radius, circuit breaker, idempotent — reliability pattern vocabulary
- Thursday (Professional): “Circle back,” “action item,” “let’s take this offline” — meeting English for tech professionals
- Friday (Career): Interview English — how to talk about your technical decisions and leadership experience
- Saturday (Social): Weekend plans and casual tech-person small talk
- Sunday (Review): Consolidate this week’s best words with spaced repetition
Start the day with one real sentence. If you used “hallucinate” or “inference” at work today in English, that’s a win. See you tomorrow morning! 🚀