☀️ Tuesday Noon: AI Vocabulary Deep Dive

Welcome to your Noon English session! Today we explore the language of Artificial Intelligence — the terms your teammates use in Slack, the buzzwords in job descriptions, and the concepts you need to explain clearly to non-technical stakeholders.

---

🌟 Word of the Day: Hallucinate

IPA	/həˈluː.sɪ.neɪt/
Vietnamese	Ảo giác (khi AI tạo ra thông tin sai hoặc bịa đặt)
Part of Speech	Verb

Definition: When an AI model generates confident-sounding but factually incorrect or completely fabricated information.

📝 Example Sentences

1. “The chatbot hallucinated a research paper that doesn’t actually exist.”
2. “We need to validate every API response because the model sometimes hallucinates function names.”
3. “Our client was upset when the AI hallucinated product details in the generated report.”

🔗 References

• 📖 Merriam-Webster — hallucinate
• 🎥 What is AI Hallucination? — IBM Technology (YouTube)

---

📚 Vocabulary Table: 5 Essential AI Terms

Phrase	Vietnamese Meaning	Example
inference	Suy luận / quá trình AI đưa ra kết quả	“Inference latency dropped to 200ms after we optimized the model.”
fine-tuning	Tinh chỉnh mô hình	“We fine-tuned the base model on our internal support tickets.”
embeddings	Véc-tơ nhúng (biểu diễn dữ liệu dạng số)	“The search uses embeddings to find semantically similar documents.”
context window	Cửa sổ ngữ cảnh (lượng văn bản AI có thể xử lý cùng lúc)	“GPT-4’s context window is large enough to process entire codebases.”
guardrails	Rào cản / giới hạn an toàn cho AI	“We added guardrails to prevent the model from generating harmful content.”

---

🗣️ Pronunciation Guide

Practice Sentence:

“The model’s inference speed improved after fine-tuning, but we still need stronger guardrails to prevent it from hallucinating.”

Breakdown:

Word	IPA	Stress	Tip
inference	/ˈɪn.fər.əns/	IN-fer-ence	3 syllables, stress on first
fine-tuning	/faɪn ˈtjuː.nɪŋ/	fine-TU-ning	hyphenated compound
guardrails	/ˈɡɑːrd.reɪlz/	GUARD-rails	stress on first syllable
hallucinating	/həˈluː.sɪ.neɪ.tɪŋ/	ha-LU-ci-na-ting	5 syllables

🎧 Audio References

• 🔊 inference — Cambridge Dictionary
• 🔊 hallucinate — Forvo

---

✏️ Exercise 1: Vocabulary in Context

Fill in the blank with the correct word: (inference / fine-tuning / embeddings / context window / hallucinate)

1. “The AI tends to ________ when asked about very recent events it wasn’t trained on.”
2. “We store user queries as ________ in a vector database for semantic search.”
3. “After ________ the model on our domain data, accuracy improved by 30%.”
4. “If the ________ is too small, the model loses track of earlier parts of the conversation.”
5. “Running ________ on a GPU cluster is much faster than on CPU.”

✅ Check Your Answers

1. hallucinate — AI bịa thông tin khi không biết
2. embeddings — lưu trữ dạng véc-tơ để tìm kiếm ngữ nghĩa
3. fine-tuning — tinh chỉnh mô hình trên dữ liệu chuyên biệt
4. context window — cửa sổ ngữ cảnh xác định lượng thông tin AI nhớ được
5. inference — quá trình AI chạy để đưa ra kết quả

---

✏️ Exercise 2: Translation Challenge

Translate these Vietnamese sentences into natural English using today’s vocabulary:

1. “Mô hình này đôi khi tạo ra thông tin sai, chúng ta cần thêm rào cản an toàn.”
2. “Chúng tôi đang tinh chỉnh mô hình trên dữ liệu hợp đồng pháp lý của công ty.”
3. “Tốc độ suy luận là vấn đề lớn nhất trong hệ thống thời gian thực của chúng tôi.”

✅ Sample Answers

1. “This model sometimes hallucinates — we need stronger guardrails.”
2. “We’re fine-tuning the model on our company’s legal contract data.”
3. “Inference speed is our biggest challenge in the real-time system.”

Vocabulary used: hallucinate, guardrails, fine-tuning, inference

---

💡 Idiom of the Day: “garbage in, garbage out”

Vietnamese	Đầu vào rác, đầu ra rác — chất lượng kết quả phụ thuộc vào chất lượng đầu vào
Pronunciation	/ˈɡɑːr.bɪdʒ ɪn, ˈɡɑːr.bɪdʒ aʊt/
Usage	Very common in tech and AI discussions

Usage Examples:

1. “The model’s outputs are unreliable because the training data wasn’t cleaned — garbage in, garbage out.”
2. “Before we blame the AI, let’s check the input quality. Remember: garbage in, garbage out.”

---

🎭 Mini Dialogue: Code Review Discussion

Scene: Two engineers reviewing an AI feature in a pull request on Slack.

---

Priya: Hey, I looked at the PR. The inference latency looks good, but I’m worried the model might hallucinate on edge cases.

Thuan: Yeah, that’s why I added guardrails in the post-processing step — it filters out responses below a confidence threshold.

Priya: Smart. What about the context window? Are we truncating inputs properly?

Thuan: We’re cutting at 3,000 tokens. The fine-tuned model handles that range well based on our evals.

Priya: Perfect. And the embeddings for the search feature — are those cached?

Thuan: Yes, cached at the document level. Regenerating only when content changes.

Priya: Looks solid. Approving this one. 🚀

---

Vocabulary used: inference latency, hallucinate, guardrails, context window, fine-tuned, embeddings

---

🎯 Your 2-Minute Challenge

Right now, do this:

Open a new Slack message or Google Doc and write 2–3 sentences explaining what “AI hallucination” means to a non-technical colleague — as if they just asked you “why did the AI give us wrong info?”

Example starter:

“When an AI hallucinates, it doesn’t know that it doesn’t know something. It generates a confident-sounding answer that’s actually made up. That’s why we always need guardrails to validate outputs before they reach users.”

✍️ Try to use at least 3 words from today’s vocabulary table!

---

📊 Today’s Summary

Term	Quick Meaning
hallucinate	AI tạo thông tin sai/bịa đặt
inference	AI chạy để đưa ra kết quả
fine-tuning	Tinh chỉnh mô hình trên dữ liệu mới
embeddings	Biểu diễn ngữ nghĩa dạng số
context window	Lượng thông tin AI xử lý cùng lúc
guardrails	Rào cản an toàn cho AI

---

🕛 Noon session complete! Come back at 6PM for your evening English practice.

Part of the Daily English Series — designed for Vietnamese tech professionals.