Tuesday Morning: AI & Machine Learning English

AI and ML are reshaping every tech conversation — from job interviews to architecture reviews to stakeholder updates. Today we focus on the words you say in those conversations, not just read.


🌟 Word of the Day: Inference

Pronunciation/ˈɪn.fər.əns/
Vietnamesesuy luận / chạy model để dự đoán kết quả
StressIN-fer-ence (3 syllables, stress on first)

🔊 Hear it: Cambridge · YouGlish AI · 3Blue1Brown YouTube

3 real-world examples:

  1. “Our inference latency dropped from 800ms to 120ms after we switched to a quantized model.”
  2. “We’re running inference on-device to avoid sending user data to the cloud.”
  3. “The cost of inference at scale is the real bottleneck — training is a one-time cost.”

Common mistake: Vietnamese engineers often say “in-FER-ence” (stress on second syllable). The correct stress is IN-fer-ence.


📋 Vocabulary: AI/ML in Real Conversations

PhraseVietnameseExample
fine-tune a modeltinh chỉnh model cho domain cụ thể”We fine-tuned GPT-4o on our support tickets to reduce hallucinations.”
context windowcửa sổ ngữ cảnh — lượng text model xử lý được”The 200k context window lets us pass the whole codebase in one prompt.”
hallucinatebịa ra thông tin sai — model tự “sáng tác""The model hallucinated a function that doesn’t exist in the SDK.”
embeddingsvector biểu diễn ngữ nghĩa của text”We store embeddings in Pinecone for semantic similarity search.”
prompt engineeringkỹ thuật viết prompt để tối ưu output”Good prompt engineering cut our token costs by 40%.”

🗣️ Pronunciation Guide

Practice sentence (read aloud 3 times):

“We’re using retrieval-augmented generation to reduce hallucinations during inference — the embeddings are stored in a vector database.”

Breakdown:

  • retrieval = /rɪˈtriːvəl/ — re-TRIEV-al (3 syllables)
  • augmented = /ɔːɡˈmentɪd/ — aug-MENT-ed
  • hallucinations = /həˌluːsɪˈneɪʃənz/ — ha-loo-si-NA-tions
  • inference = /ˈɪn.fər.əns/ — IN-fer-ence
  • embeddings = /ɪmˈbedɪŋz/ — em-BED-ings

Rhythm tip: This sentence has a technical density — pause after “generation” and after “inference” to let each concept land.


✏️ Exercise 1: Fill in the Blank

Choose from: inference, fine-tune, hallucinate, context window, embeddings, prompt engineering

  1. “We need to _______ the model on our legal documents — the base model doesn’t know our terminology.”
  2. “The model started to _______ when we asked about events after its training cutoff.”
  3. “Shrinking the _______ helped reduce costs, but we lost accuracy on long documents.”
  4. “Our _______ pipeline converts all support articles to vectors for semantic search.”
  5. “With careful _______, we reduced the number of tokens per request by 30%.”
✅ Answers
  1. fine-tune
  2. hallucinate
  3. context window
  4. embeddings
  5. prompt engineering

✏️ Exercise 2: Translate to English

  1. “Model của chúng tôi đang hallucinate khi được hỏi về dữ liệu ngoài training set.”
  2. “Chi phí inference ở production scale cao hơn nhiều so với khi chúng tôi dự đoán ban đầu.”
  3. “Chúng tôi cần fine-tune model với domain-specific data để cải thiện accuracy.”
✅ Suggested Answers
  1. “Our model is hallucinating when asked about data outside its training set.”
  2. “Inference costs at production scale are much higher than we initially projected.”
  3. “We need to fine-tune the model with domain-specific data to improve accuracy.”

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

Vietnamese: Dữ liệu rác vào thì kết quả rác ra — chất lượng output phụ thuộc vào chất lượng input.

In AI conversations:

  • “The embeddings are giving us poor search results. Garbage in, garbage out — the source documents need better chunking.”
  • “No matter how good the model, garbage in, garbage out — our prompt is too vague.”

  1. Andrej KarpathyYouTube — deep technical explanations of neural nets, LLMs, in excellent English
  2. AI ExplainedYouTube — current AI news explained clearly, great for vocabulary in context
  3. Yannic KilcherYouTube — paper walkthroughs, advanced but excellent pronunciation of technical terms

🎯 Daily Challenge

In your next Slack message or PR comment today, use at least one of these words naturally:

  • “inference” (e.g., in a cost or latency discussion)
  • “hallucinate” (when describing a model error)
  • “fine-tune” (when discussing model customization)

Don’t force it — wait for the right moment. When it comes, use it.


Tomorrow: Wednesday Morning — Architecture vocabulary (system design, scalability, distributed systems) 🏗️

Export for reading

Comments