The AI vocabulary has changed faster than any tech vocabulary I can remember. Words that were niche two years ago — agent, RAG, MCP, eval — are now in every product brief, board paper, and procurement doc. Words that felt central — chatbot, prompt engineer — feel quaint.
This glossary is the working vocabulary I use in 2026 with Malaysian corporate teams. It is grouped by where the terms actually come up at work, not alphabetically — which is more useful when you are looking up a phrase from yesterday's meeting.
1. Foundations (the words behind everything else)
1. Artificial Intelligence (AI) — Software that performs tasks normally requiring human reasoning, decision-making, or perception.
2. Machine Learning (ML) — A subset of AI where a system learns patterns from data instead of being explicitly programmed.
3. Deep Learning — Machine learning using large neural networks. The technique behind modern image, speech, and language models.
4. Neural Network — A mathematical model loosely inspired by the brain, made of layered nodes that transform inputs into outputs.
5. Generative AI (GenAI) — AI that produces new content (text, images, code, audio) rather than just classifying or predicting.
6. Foundation Model — A large model trained on broad data, designed to be adapted for many downstream tasks. GPT, Claude, and Gemini are foundation models.
7. Frontier Model — The current generation of largest, most capable foundation models. In 2026 this includes Claude 4, GPT-5, and Gemini Ultra.
8. Narrow AI — AI built for one specific task (e.g. fraud detection). Almost all AI in production is narrow AI.
9. AGI (Artificial General Intelligence) — Hypothetical AI matching general human reasoning across any task. Does not exist yet, despite headlines.
2. Large Language Models (LLMs)
10. LLM (Large Language Model) — A foundation model trained on text. The core engine behind Claude, ChatGPT, Gemini, and others.
11. Token — The unit an LLM reads and writes. Roughly four characters of English text. You pay providers per token.
12. Context Window — How much text the model can hold at once. Claude's 200K and 1M context windows in 2026 mean entire codebases or contracts fit in one prompt.
13. Parameters — The internal weights of a model. More is not always better; well-trained smaller models often beat larger weakly trained ones.
14. Inference — Running a trained model to produce outputs. The thing you pay for on every request.
15. Hallucination — When a model produces a confident answer that is factually wrong. The single biggest risk in deploying LLMs without guardrails.
16. Temperature — A setting (0 to 1+) that controls how varied the model's outputs are. Lower for deterministic tasks, higher for creative ones.
17. Top-p / Top-k — Other sampling controls that shape model output diversity.
18. Fine-tuning — Continuing to train a model on your own data so it specialises for a domain.
19. Distillation — Training a smaller, faster model to mimic a larger one's behaviour. Behind many "small but capable" models.
20. Quantisation — Compressing a model so it runs cheaper, often with minimal quality loss.
3. Prompting & Working with LLMs
21. Prompt — The instruction you give the model.
22. System Prompt — The instruction that defines the model's role, behaviour, and constraints — separate from the user's message.
23. Few-shot Prompting — Including a few worked examples in the prompt so the model learns the pattern.
24. Chain-of-Thought (CoT) — Asking the model to think step-by-step. Improves performance on reasoning tasks.
25. Zero-shot — Asking the model to do a task without examples.
26. Prompt Injection — A security attack where malicious text in the input tries to override the system prompt. Mitigation is now part of any serious AI deployment.
27. Jailbreak — A prompt designed to bypass a model's safety guidelines.
28. Prompt Engineering — The discipline of designing reliable prompts. Less of a job title in 2026, more of a skill.
29. Structured Output — Asking the model to return JSON or another machine-readable format. Critical for production use.
4. Agents & Tool Use
30. AI Agent — An LLM-powered system that can plan, use tools, and take actions toward a goal — not just answer a single question.
31. Tool Calling (Function Calling) — The mechanism that lets an LLM invoke external functions, APIs, or workflow nodes.
32. Multi-Agent System — Multiple specialised agents that collaborate on a task, often coordinated by a planner agent.
33. Orchestration — The pattern of coordinating multiple agents, models, or tools to deliver an outcome. The core of our orchestration training.
34. Planner / Executor — A common multi-agent pattern: one agent decides what to do; others carry out the steps.
35. ReAct — A pattern where the agent alternates between Reasoning and Acting (calling tools).
36. Autonomous Agent — An agent allowed to operate without human approval at each step. Rarely the right answer in production.
37. Human-in-the-Loop (HITL) — Workflows where a human approves or edits the agent's output before it acts.
38. Computer Use — A 2024–2026 capability where models can directly operate browsers and computers.
39. Browser Agent — An agent that drives a real browser to complete tasks (forms, scraping, transactions).
5. RAG & Knowledge
40. RAG (Retrieval-Augmented Generation) — Pattern where the system retrieves relevant documents and gives them to the LLM as context.
41. Vector Database — A database optimised for similarity search over embeddings (Pinecone, Weaviate, Postgres pgvector).
42. Embedding — A numerical representation of a piece of text that captures its meaning. The basis of semantic search.
43. Semantic Search — Search by meaning rather than keywords.
44. Chunking — Breaking long documents into smaller passages for retrieval.
45. Reranker — A second model that re-orders retrieved chunks for better relevance.
46. Hybrid Search — Combining keyword and semantic search. Usually beats either alone.
47. Knowledge Base / Corpus — The collection of documents your RAG system can draw from.
6. Engineering & Operations
48. MCP (Model Context Protocol) — An open protocol for giving models access to tools, files, and APIs. Standardises what used to be custom integration work.
49. API (Application Programming Interface) — How software talks to other software. The backbone of almost any AI integration.
50. Webhook — A URL that another system calls when an event happens. n8n triggers often start as webhooks.
51. Latency — How long the model takes to respond.
52. Throughput — How many requests per second the system can handle.
53. Observability — The ability to see what the model is doing in production — inputs, outputs, costs, latencies.
54. Eval (Evaluation) — A systematic test of an AI system's quality. The thing that separates real engineering from demos.
55. Guardrails — Rules that constrain what an AI system can do or say. Includes content filters, scope limits, and tool allowlists.
56. Vibe Coding — Building software by describing intent to an AI coding assistant rather than writing every line. Behind many of our vibe coding workshops.
7. Risk, Governance, Ethics
57. Bias — Systematic skew in a model's outputs reflecting biased training data or design choices.
58. Drift — When a model's performance degrades over time as the world changes around it.
59. Explainability — Whether and how a system can explain why it made a decision.
60. Model Risk Management — A formal discipline (especially in banking) of governing the lifecycle of AI models. BNM RMiT touches this for Malaysian financial institutions.
61. PDPA — Malaysia's Personal Data Protection Act 2010. Defines lawful processing of personal data, with implications for AI training and inference.
62. Data Residency — The requirement that data stays inside a specific country or region. Often relevant for Malaysian banks and government.
63. Red Teaming — Stress-testing an AI system with adversarial inputs to find failure modes.
64. Differential Privacy — Mathematical technique for releasing data analysis results without exposing individuals.
65. Watermarking — Embedding signals in AI-generated content to make it identifiable later.
8. Tools & Platforms (2026 working set)
66. Claude (Anthropic) — A frontier model family. Strong tool use and long context. The default in many of our deployments.
67. Claude Code — Anthropic's terminal-native AI coding agent. Behind much of the vibe-coding shift.
68. ChatGPT / GPT-5 (OpenAI) — OpenAI's model line. Strong general-purpose performance.
69. Gemini (Google) — Google's frontier model. Strong on multimodal and search-grounded tasks.
70. n8n — Open-source workflow automation platform with native AI nodes. The backbone of most agent deployments we run.
71. LangChain / LangGraph — Frameworks for building LLM applications and agents in code.
72. Cursor / Antigravity — AI-native code editors that pair with frontier models.
73. Hugging Face — Hub for open-source models, datasets, and tools.
74. HRDC (HRD Corp) — Malaysia's Human Resource Development Corporation. AI training fees are claimable under SBL-KHAS for eligible employers.
75. Supern8n — Our marketplace of pre-built n8n workflows for Malaysian use cases.
How to use this glossary
Bookmark it, share it across your team, and update it. If a term in your meetings is missing here, send it to us and we will add it. The point is shared vocabulary — half the difficulty of an AI initiative inside any organisation is people using the same words to mean different things.
If you want to take this further, our AI Vocabulary Alignment programme is a half-day workshop designed to land this vocabulary across a single leadership team — so finance, operations, IT, and HR all leave the room with the same definitions.