Tuesday Morning: AI & Machine Learning Vocabulary
Tuesday is a great day to dig into domain vocabulary. If you work in tech and need to discuss AI systems with international colleagues, you need the right words — not just to understand what others say, but to speak confidently in meetings, on calls, and in written communication.
Today’s focus: the core vocabulary of AI and machine learning that appears in every conversation about modern software.
Word of the Day
inference /ˈɪn.fər.əns/ — suy luận / suy diễn (trong AI: quá trình mô hình tạo ra kết quả)
Pronunciation breakdown:
- IN — short “i”, like “inn”
- fər — unstressed schwa, the “uh” sound
- əns — ends with a soft “s”
Full word: IN-fər-əns (stress on the first syllable)
In AI, “inference” has a specific meaning: it is the process of running a trained model on new input to get an output. When you send a message to an LLM and it replies, the model is performing inference.
Examples:
“Our API handles about 10,000 inference requests per second at peak.”
“Inference costs have dropped significantly with the new model architecture.”
“The model runs inference locally on the device — no data leaves the phone.”
Vietnamese tip: Đừng nhầm “inference” (quá trình mô hình xử lý) với “training” (quá trình mô hình học). Training tốn nhiều tài nguyên hơn; inference là bước chúng ta dùng hàng ngày.
Practice sentence: Say this out loud three times:
“The inference latency dropped from 800ms to 120ms after we switched models.”
Vocabulary Table: 5 Essential AI/ML Terms
| Term | IPA | Vietnamese | Example in Context |
|---|---|---|---|
| LLM (Large Language Model) | /ˌel.el.ˈem/ | Mô hình ngôn ngữ lớn | ”We’re using an LLM to power our search feature.” |
| prompt | /prɒmpt/ | câu lệnh đầu vào / lời nhắc | ”A well-written prompt can dramatically improve output quality.” |
| token | /ˈtoʊ.kən/ | đơn vị văn bản nhỏ nhất | ”This model has a 128,000 token context window.” |
| RAG (Retrieval-Augmented Generation) | /ræɡ/ | tìm kiếm + tạo sinh kết hợp | ”We use RAG to ground the model on our internal docs.” |
| agent (AI agent) | /ˈeɪ.dʒənt/ | tác nhân AI (tự hành động) | “The agent browses the web, reads the result, and updates the ticket.” |
Pronunciation Guide
These AI terms are often mispronounced by non-native speakers. Here’s what to watch for:
“inference” — NOT “infe-RENCE” (wrong stress). Say IN-ference.
“prompt” — The “o” sounds like “awe” (British) or “ah” (American). One syllable: /prɒmpt/. Don’t add an extra syllable: not “pro-MEPT.”
“token” — TOE-ken. Two clear syllables. The “o” is a long vowel.
“agent” — AY-jent. The first syllable rhymes with “day.” Don’t say “AH-jent.”
“RAG” — Pronounced as a word, like the cleaning cloth. /ræɡ/. Some engineers say “R-A-G” letter by letter — both are fine in context.
Practice drill — read this aloud:
“The RAG agent takes the user’s prompt, fetches relevant tokens from our knowledge base, and passes them to the LLM for inference.”
Exercise 1: Fill in the Blank
Complete these sentences with the correct word from the vocabulary table. Read each completed sentence aloud.
- “We spent two weeks fine-tuning our ___________ strategy before we got consistent results from the model.”
- “The new model performs ___________ locally on the device, which means no latency from network calls.”
- “Our ___________ pipeline searches our documentation, pulls the top three matches, and adds them to the prompt before generation.”
- “At current pricing, each ___________ costs about $0.003 per thousand input ___________s.”
- “We built an ___________ that monitors our Jira board, reads new tickets, assigns labels, and posts a summary to Slack.”
Check your answers
- prompt — “We spent two weeks fine-tuning our prompt strategy…”
- inference — “The new model performs inference locally on the device…”
- RAG — “Our RAG pipeline searches our documentation…”
- LLM / token — “At current pricing, each LLM call costs about $0.003 per thousand input tokens.”
- agent — “We built an agent that monitors our Jira board…”
Exercise 2: Match the Definition
Match each term on the left with its correct definition on the right.
| Term | Definition |
|---|---|
| A. Token | 1. A system where retrieved documents are added to the model’s context before it generates a response |
| B. Inference | 2. The smallest unit of text a language model processes — roughly a word or part of a word |
| C. Prompt | 3. The process of running a trained AI model on new input to produce an output |
| D. RAG | 4. An AI system that can take actions — like browsing, writing, or running code — to complete a task |
| E. Agent | 5. The text instruction you give a model to tell it what to do |
Check your answers
- A → 2 (Token = smallest unit of text)
- B → 3 (Inference = running the model to get output)
- C → 5 (Prompt = instruction text)
- D → 1 (RAG = retrieval + generation)
- E → 4 (Agent = autonomous action-taking system)
Idiom of the Day
”garbage in, garbage out”
Meaning: If you give a system bad input, you’ll get bad output. Originally a programming concept, it’s now widely used in AI discussions.
In context:
“We keep getting weird model outputs. I think it’s a garbage-in, garbage-out problem — the prompts aren’t specific enough.”
“The model is fine; the data pipeline is the issue. Garbage in, garbage out.”
Vietnamese equivalent: “Đầu vào rác, đầu ra rác” — hoặc dùng nguyên văn tiếng Anh vì cụm này rất phổ biến trong cộng đồng tech.
Practice: Use this idiom today when discussing data quality or prompt quality with a colleague.
Recommended Watching
Search YouTube for: “Andrej Karpathy: Intro to Large Language Models” (1 hour)
This is one of the clearest English explanations of how LLMs work, by a former Tesla AI director and OpenAI cofounder. Even if you know the concepts, watching and listening will help your ear get used to how native English speakers explain technical AI topics.
Focus on: how he pronounces “inference,” “training,” “token,” and “embedding.”
Daily Challenge
Write 3 sentences about your current project or team using at least 3 words from today’s vocabulary table. Share them in a team chat, or just say them aloud to yourself.
Example:
“Our team is evaluating two LLMs for our support chatbot. We’re testing inference speed and cost per token. The RAG pipeline pulls from our help center docs.”
The goal is not perfect sentences — the goal is saying these words out loud until they feel natural.