☀️ Tuesday Noon: AI Vocabulary Deep Dive
Master the language of AI — from jargon to plain-English explanations that impress colleagues and clients alike.
🌟 Word of the Day: Hallucinate
| IPA | /həˈluː.sɪ.neɪt/ |
| Vietnamese | Ảo giác / bịa đặt (khi AI tạo ra thông tin sai) |
| Part of speech | Verb |
In AI, when a model hallucinates, it confidently generates false or made-up information.
📖 Example Sentences
- “The chatbot hallucinated a fake research paper and cited authors who don’t exist.”
- “We need to add a validation layer because our LLM sometimes hallucinates API endpoints.”
- “One challenge with generative AI is preventing the model from hallucinating facts in customer-facing responses.”
🔗 References
- 🎬 IBM Technology — AI Hallucinations Explained
- 📖 Merriam-Webster: hallucinate
- 🎓 Cambridge Dictionary: hallucinate
📚 Vocabulary Table: AI Terms You Must Know
| Phrase | IPA | Vietnamese Meaning | Example in Context |
|---|---|---|---|
| inference | /ˈɪn.fər.əns/ | Suy luận / chạy mô hình để dự đoán | ”Inference latency dropped by 40% after we switched to a quantized model.” |
| fine-tune | /faɪn tjuːn/ | Tinh chỉnh mô hình | ”We fine-tuned GPT on our internal documents to improve domain accuracy.” |
| grounding | /ˈɡraʊn.dɪŋ/ | Gắn dữ liệu thực tế vào AI | ”Grounding the model with real-time data prevents hallucinations.” |
| tokenize | /ˈtoʊ.kə.naɪz/ | Chia văn bản thành token | ”The model tokenizes your input before processing — each word may become multiple tokens.” |
| agentic | /eɪˈdʒen.tɪk/ | Có tính tự chủ / hành động như agent | ”We’re moving toward agentic workflows where AI takes multi-step actions autonomously.” |
🗣️ Pronunciation Guide
Practice Sentence:
“The agentic AI was fine-tuned to minimize hallucinations during inference.”
Breakdown:
| Word | Sounds Like | Tip |
|---|---|---|
| agentic | ay-JEN-tik | Stress on second syllable: a-JEN-tic |
| fine-tuned | FYNE-tyoond | Say “fine” + “tuned” smoothly together |
| hallucinations | huh-LOO-sih-NAY-shunz | 5 syllables — stress on “NAY” |
| inference | IN-fuh-runss | 3 syllables, soft ending |
🔗 Audio References
✏️ Exercise 1: Vocabulary in Context
Fill in the blank with the correct word: hallucinate / fine-tune / inference / grounding / agentic / tokenize
- Before the model can understand your message, it must ________ your text into smaller units.
- We plan to ________ the base model on our company’s codebase so it understands our coding patterns.
- The new AI assistant is ________ — it can search the web, write code, and send emails without being told each step.
- Adding RAG (retrieval-augmented generation) is a ________ technique that connects the AI to real documents.
- During ________, the model processes your query and generates a response in milliseconds.
- Without proper safeguards, the model may ________ and give users completely wrong medical advice.
✅ Check Your Answers
- tokenize — Breaking text into tokens is the first step in NLP processing.
- fine-tune — Domain-specific fine-tuning improves model performance on specialized tasks.
- agentic — Agentic AI can plan and execute multi-step tasks autonomously.
- grounding — Grounding anchors AI responses to verified, real-world data.
- inference — Inference is when a trained model generates output from new input.
- hallucinate — AI hallucination is a major safety concern in high-stakes domains.
🌍 Exercise 2: Translation Challenge
Translate these Vietnamese sentences into natural English using today’s vocabulary:
- “Chúng tôi cần tinh chỉnh mô hình này để nó hiểu thuật ngữ y tế.”
- “Mô hình đôi khi bịa đặt thông tin — đó là lý do chúng ta cần kiểm tra đầu ra.”
- “Tốc độ suy luận rất quan trọng trong các ứng dụng thời gian thực.”
✅ Sample Answers
- “We need to fine-tune this model so it understands medical terminology.”
- “The model sometimes hallucinates information — that’s why we need to validate the output.”
- “Inference speed is critical in real-time applications.”
💡 Pro Tip: In tech discussions, prefer “inference latency” over “inference speed” — it sounds more precise and professional.
💡 Idiom of the Day: “Garbage in, garbage out”
| Vietnamese | Rác vào, rác ra — dữ liệu xấu → kết quả xấu |
| Used in AI/Tech | Nói về chất lượng dữ liệu đầu vào ảnh hưởng đến đầu ra |
Usage Examples:
-
“Our model’s predictions were terrible last quarter. Classic garbage in, garbage out — the training data had too many labeling errors.”
-
“Before we blame the AI, let’s check the input quality. Garbage in, garbage out — if users send vague prompts, the output will be vague too.”
🎯 When to use it: Any time poor-quality inputs lead to poor-quality results — in AI, data pipelines, or even team processes.
🎭 Mini Dialogue: Explaining AI to a Stakeholder
Context: A product manager asks a senior engineer about why the AI chatbot gave wrong answers.
Sarah (Product Manager): “A customer complained that the chatbot told them our product has a feature it doesn’t have. What happened?”
Dev (Senior Engineer): “The model hallucinated. It generated a confident-sounding answer that wasn’t grounded in our actual product data.”
Sarah: “So it just… made things up? How do we fix it?”
Dev: “We’re implementing grounding — connecting the model to our product database so every response is backed by real data, not just what the model was trained on.”
Sarah: “Will that slow down inference — the time it takes to respond?”
Dev: “Slightly, but we’re optimizing. We may also fine-tune the model on our product documentation to reduce errors further.”
Sarah: “Got it. Garbage in, garbage out — better data means better answers.”
Dev: “Exactly. You’re speaking AI now! 🎉”
🏆 Today’s Challenge
⏱ 2-Minute Action — Do this RIGHT NOW:
Open Slack, Teams, or your notes app and write one sentence using today’s vocabulary to describe your current AI project or a recent AI tool you used.
Template to get started:
“We’re using [tool] which [fine-tunes / grounds / tokenizes] data to [solve problem], but we sometimes see [hallucination / inference] issues when [situation].”
Example:
“We’re using Claude which grounds responses to our knowledge base, but we sometimes see hallucination issues when users ask about very recent events.”
📤 Bonus: Share it with your team or in a tech community — teaching others cements your own learning!
📊 Quick Summary
| Term | Remember It As |
|---|---|
| Hallucinate | AI confidently making things up |
| Inference | Model generating output from input |
| Fine-tune | Training further on specific data |
| Grounding | Connecting AI to real facts |
| Tokenize | Splitting text for AI to process |
| Agentic | AI that acts autonomously |
| Garbage in, garbage out | Data quality = output quality |
💪 Keep going! Check the Morning Session for today’s warm-up, and come back for the Evening Session for advanced practice.
🔔 Tip: Bookmark this page and review vocabulary before your next technical meeting!