English Lesson — Tuesday Noon: AI Vocabulary Deep Dive
☀️ Noon focus: The AI buzzwords your team uses every day — and how to explain them clearly to clients, product managers, and anyone who isn’t a developer.
📖 Word of the Day: Hallucinate /həˈluːsɪneɪt/
Vietnamese: ảo giác / bịa đặt — Khi AI “bịa ra” thông tin trông có vẻ thật nhưng hoàn toàn sai
In AI and LLM context, a model hallucinates when it generates text that sounds confident and plausible but is factually incorrect or completely made up. The model is not lying intentionally — it is producing the statistically likely next token, which happens to be wrong.
3 Example Sentences:
- “The chatbot hallucinated a legal citation — it gave the case name and year, but the case never existed.”
- “We had to add a verification step because the model kept hallucinating API endpoint names that are not in our documentation.”
- “Always fact-check AI-generated summaries; even the best LLMs hallucinate when asked about recent or niche topics.”
🔗 Resources:
📚 Vocabulary Table
| Phrase | Vietnamese | Example Sentence |
|---|---|---|
| inference | suy luận / chạy mô hình | ”Running inference on a large image model requires a GPU with at least 16 GB of VRAM.” |
| embedding | nhúng vector / biểu diễn số | ”We store each document as an embedding so the search engine can find semantically similar results.” |
| fine-tune | tinh chỉnh mô hình | ”We fine-tuned the base model on our support tickets so it understands our product’s terminology.” |
| token | đơn vị văn bản nhỏ nhất (từ / mảnh từ) | “This prompt is 1,200 tokens — close to the context limit, so trim the system prompt.” |
| prompt engineering | kỹ thuật viết câu lệnh cho AI | ”Good prompt engineering can improve output quality without changing the model at all.” |
🗣️ Pronunciation Guide
Word: hallucinate /həˈluːsɪneɪt/
Break it down syllable by syllable:
| Syllable | Sound | Tip |
|---|---|---|
| hal | /həl/ | Soft “h”, short schwa vowel — like the start of “hello” said quickly |
| lu | /luː/ | Long “oo” sound — like “loo” (British for bathroom) |
| ci | /sɪ/ | Short “i” — like “sit”, not “see” |
| nate | /neɪt/ | Like the name “Nate” — long “a” |
Full stress pattern: hal-LU-ci-nate (stress on the second syllable)
Practice sentence — say it 3 times out loud:
“The AI model hallucinated facts that seemed convincing but were completely wrong.”
Tips:
- Do not rush “hallucinated” — give each syllable space: hal-LU-ci-na-ted
- Stress “convinced” and “completely” to make the sentence sound natural
- Record yourself, then compare with a native speaker on YouGlish: hallucinate
✏️ Exercises
Exercise 1: Fill in the Blank
Use one of these words to complete each sentence: inference, embedding, fine-tune, token, prompt engineering
- “We store each FAQ answer as an ___________ in a vector database so users get semantically relevant search results.”
- “The API returned a 429 error — we are sending too many ___________s per minute and hitting the rate limit.”
- “Running ___________ on this 70-billion-parameter model takes 8 seconds per request — we need to optimize.”
- “Instead of training a new model from scratch, we decided to ___________ GPT-4o on our internal dataset.”
- “The junior dev improved the output just by changing the system prompt — that’s the power of ___________.”
✅ Show Answers
- embedding
- token
- inference
- fine-tune
- prompt engineering
Exercise 2: Translate to English
Translate these Vietnamese sentences about AI concepts into natural English. Focus on using the vocabulary from today’s lesson.
- “Mô hình AI đã bịa đặt ra tên của một công ty không tồn tại trong báo cáo tài chính.”
- “Chúng tôi cần tinh chỉnh mô hình vì nó không hiểu thuật ngữ trong ngành y tế của chúng tôi.”
- “Mỗi câu được chuyển đổi thành một vector nhúng trước khi được lưu vào cơ sở dữ liệu.”
- “Câu lệnh của bạn quá dài — hãy cắt bớt để tiết kiệm token.”
- “Kỹ thuật viết câu lệnh tốt có thể tạo ra sự khác biệt lớn mà không cần thay đổi mô hình.”
✅ Show Answers
- “The AI model hallucinated the name of a company that does not exist in the financial report.”
- “We need to fine-tune the model because it does not understand our medical industry terminology.”
- “Each sentence is converted into an embedding vector before being stored in the database.”
- “Your prompt is too long — cut it down to save tokens.”
- “Good prompt engineering can make a big difference without changing the model at all.”
💡 Idiom of the Day: Garbage in, garbage out
Vietnamese: Đầu vào rác thì đầu ra cũng rác — Nếu bạn cho AI (hoặc bất kỳ hệ thống nào) dữ liệu tệ, bạn sẽ nhận lại kết quả tệ.
Origin: A classic computer science saying from the 1960s. It is more relevant than ever in the AI age — model quality is only as good as the training data and inputs it receives.
2 Examples:
-
“We spent three months fine-tuning the model on mislabeled data and got terrible results. Garbage in, garbage out — we had to clean the dataset first.”
-
“The client kept asking why the AI assistant gave vague answers. We looked at their prompts — they were one-sentence instructions with no context. Classic garbage in, garbage out.”
How to use it at work:
“Before we blame the model, let’s audit the training data. If the input is low quality, garbage in, garbage out.”
🎭 Mini Dialogue
Scene: A developer (Dev) and a product manager (PM) are in a Slack huddle reviewing a demo of their new AI-powered assistant.
PM: “The demo looked great, but in one case the assistant said our product supports a feature that we definitely don’t have. Did it just make that up?”
Dev: “Yes — that’s called hallucination. The model generated a confident-sounding answer, but it was completely fabricated. It’s a known limitation of large language models.”
PM: “So it just… lies? How do we prevent that in production?”
Dev: “It’s not lying intentionally — it’s predicting the most statistically likely next word, and sometimes that’s wrong. We can reduce it with retrieval-augmented generation: we give the model our actual docs at inference time so it answers from real data instead of guessing.”
PM: “Got it. So the model needs context to stay accurate. Can we add that before launch?”
Dev: “Absolutely. That’s exactly why we’re building the embedding pipeline this sprint — once that’s done, hallucination on product questions should drop significantly.”
🏆 Daily Challenge
Your task: Explain what AI hallucination means to a non-tech friend in 30 seconds or fewer.
Rules:
- No jargon (no “LLM”, “token”, “inference”)
- Use an analogy from everyday life
- Make it clear it is a limitation, not a bug that can just be “fixed”
Example answer (30 seconds when spoken):
“You know how someone can be extremely confident and still be completely wrong? AI is like that. It does not look things up — it predicts what words should come next based on patterns it learned from billions of pages of text. Most of the time that’s great. But sometimes it predicts something that sounds right but is totally made up. We call that hallucination. It is like a student who has read a lot of books but is guessing on an exam they haven’t studied for — and they guess confidently.”
Try it yourself: Write your own 30-second explanation in the comments, or say it aloud and record yourself. The goal is to make a non-technical person nod and say “Oh, I get it now.”
🔁 Coming up this evening: Negotiation English — how to push back on deadlines and scope without damaging relationships.