You don’t need a computer-science degree to use AI well at work. But you do need to know the language.
Every memo, vendor pitch, and LinkedIn post about AI right now is loaded with terms — LLM, RAG, agent, fine-tuning, hallucination — that get used interchangeably, redefined by marketing teams, or left unexplained. The result: smart professionals nod along in meetings without being sure what anyone actually means.
This glossary fixes that. It covers 30 terms that show up in real conversations about AI at work, with plain-English definitions, concrete examples, and links to where each skill fits in the Lumaire Level 1 — AI Foundations course or Level 2 — AI Productivity if you want to go deeper.
We’ve grouped them into three sections — start at the top if you’re new to AI, skip to the section that matches the conversations you’re having.
Section 1: The Basics — How AI Actually Works
If you’ve ever wondered “what is AI, really?” or felt embarrassed asking what an “LLM” stands for, start here. These are the terms that come up in every introductory article and most vendor pitches.
1. AI (Artificial Intelligence)
The umbrella term for computer systems that perform tasks normally requiring human intelligence — recognizing speech, making decisions, translating languages, identifying objects in images. “AI” covers everything from a chess engine to a chatbot.
In practice: When someone says a product “uses AI,” ask what kind. The honest answer might be a statistical model built in 2012, or it might be a frontier large language model from last week. The label tells you almost nothing about capability.
Related: Machine Learning, Generative AI, Narrow AI, AGI.
2. Machine Learning (ML)
A subfield of AI where computers learn from data instead of being explicitly programmed. Instead of writing rules (“if X then Y”), engineers feed examples to a model and let it discover the patterns.
In practice: Your email spam filter is machine learning. It learned what spam looks like by reading millions of examples, not because someone wrote a rule for every spam pattern. Same idea scales up to everything from fraud detection to medical imaging.
3. Deep Learning
A type of machine learning that uses multi-layer “neural networks” — loosely inspired by how the brain processes information. Deep learning is what made modern AI possible. It’s particularly good at messy, unstructured data: images, audio, natural language.
In practice: When a vendor says “deep learning” they usually mean “modern AI.” Don’t be intimidated — it’s a useful label but not a separate thing you need to understand. The underlying mechanics don’t change what you can do with the tools.
4. Neural Network
The computational structure underlying deep learning. Data flows through layers of interconnected “neurons” (mathematical functions), each layer extracting slightly more abstract patterns from the previous one.
In practice: You will never build one from scratch. You may, however, hear “neural net” used as a stand-in for “AI model.” That’s accurate at a high level; you don’t need the math.
5. Algorithm
A step-by-step procedure for solving a problem. “Algorithm” predates AI by centuries — it’s not a synonym for AI, even though journalists sometimes use it that way. A recipe is an algorithm. So is the procedure your bank’s app uses to sort transactions.
In practice: If someone says “the algorithm did it,” they’re usually deflecting responsibility rather than explaining anything. Ask for specifics.
6. Generative AI (Gen AI)
AI that creates new content — text, images, audio, video, code — rather than just classifying or predicting. ChatGPT, Claude, Midjourney, and Sora are all generative AI. This is the category that’s reshaped work since late 2022.
In practice: Generative AI is what most professionals mean when they say “AI” today. If you’re evaluating tools, the meaningful question is usually: “what can this generate?” — drafting, summarizing, ideating, designing, coding.
7. Large Language Model (LLM)
The dominant type of generative AI: a model trained on enormous amounts of text that can produce coherent, contextually relevant writing, code, and reasoning. GPT-4, Claude, Gemini, and Llama are LLMs. “Large” refers to the number of parameters (weights inside the model), not the size of the company that built it.
In practice: Every modern AI chatbot you’ve heard of is powered by an LLM. When you hear “model” in AI conversations, it almost always means LLM. Knowing which model a tool uses matters less than you think — most leading models are within a few percentage points of each other on common tasks.
8. Natural Language Processing (NLP)
The older, broader field of getting computers to understand and generate human language. LLMs are the most powerful NLP systems ever built, but “NLP” also covers earlier techniques like sentiment analysis, named entity recognition, and machine translation.
In practice: “NLP” is what your doctor’s office used when they automated chart reading ten years ago. “LLM” is what your marketing team uses to draft emails today. The capabilities are vastly different even though both fall under “language AI.”
9. Foundation Model
A large model trained on broad data (text, images, sometimes audio and video) that can be adapted to many downstream tasks. LLMs are foundation models; so are image generators like Stable Diffusion. “Foundation” emphasizes that they’re the base layer that other, more specific tools get built on top of.
In practice: When a vendor says they “built on top of a foundation model,” they mean they took a general-purpose model (often via API) and added their own data, prompts, or workflow on top. This is now the dominant pattern — almost every “AI product” is mostly a thin layer over a foundation model.
10. Transformer
The neural network architecture, introduced in a 2017 Google paper, that makes modern LLMs possible. Before transformers, language models processed text sequentially; transformers process text in parallel and use a mechanism called “attention” to weigh which words matter most in context.
In practice: You almost never need to know this. But if you read any AI technical content, the word “transformer” will appear in the first paragraph. Now you know: it’s the architecture, not a robot in disguise.
Section 2: How AI Systems Behave
These are the terms that describe what AI does — the practical behaviors you’ll encounter when you actually use these tools.
11. Training Data
The corpus of text, images, or other content a model learns from. LLMs are trained on hundreds of billions of words scraped from books, websites, code repositories, and curated datasets. Training data shapes what a model “knows” and what biases it inherits.
In practice: A model’s training data has a cutoff date — typically 6–18 months before release. That’s why older events are confidently answered but last week’s news gets invented (“hallucinated,” see term 19). When in doubt, ask: “What was this trained on, and when?“
12. Token
The unit of text an LLM reads and writes. Tokens are roughly word-fragments — “running” might be one token, “unbelievable” might be three. Models charge by the token, both for input (what you send) and output (what they generate back).
In practice: When you see pricing like “$3 per million input tokens,” that’s how the meter runs. A token is roughly 4 characters of English text, so a million tokens is about 750,000 words — a small library. Most prompts are 100–2,000 tokens.
13. Context Window
The maximum amount of text a model can consider at once. Older models had 4,000–8,000 tokens (a few pages). Today’s frontier models have 100,000 to 2,000,000 tokens — entire books, multi-day conversations, large codebases. The larger the context window, the more the model can “remember” within a single session.
In practice: If you’re trying to summarize a long document or maintain a multi-turn conversation, context window size matters. If you’re asking short questions, it doesn’t. There’s no public data proving which use cases benefit from which size — experiment.
14. Embedding
A numerical representation of text (or images, audio, etc.) that captures semantic meaning. Embeddings let computers compare “how similar” two pieces of content are, even when the words are completely different. “King” and “queen” sit close together in embedding space; “king” and “banana” sit far apart.
In practice: Embeddings are the engine behind semantic search — finding documents that match the meaning of your query, not just the literal words. They’re also the foundation of RAG (term 23) and most enterprise search tools.
15. Vector Database
A database optimized for storing and searching embeddings. Instead of searching by keyword, you search by similarity: “find me the 10 chunks of text closest in meaning to this query.” Tools like Pinecone, Weaviate, and Chroma are vector databases; Postgres now has vector search built in.
In practice: If you’ve ever heard a product team say “we’re using vector search to make our docs smarter,” this is what they mean. The vector database holds embeddings of every document, then finds the closest matches when you ask a question.
16. Fine-Tuning
Taking a pre-trained model and training it further on a smaller, specific dataset to specialize it. Fine-tuning teaches the model a style, a domain, or a task it didn’t handle well out of the box — like training a general-purpose LLM to write in your company’s voice, or to read legal contracts accurately.
In practice: Fine-tuning is more expensive and slower than prompting. Most teams should try RAG (term 23) and prompt engineering (term 22) first. Fine-tune only when you have a clear, repeated failure mode that prompting can’t fix.
17. Parameters
The internal “knobs” a model learns during training. Modern LLMs have anywhere from a few billion to over a trillion parameters. More parameters generally means more capable (but also more expensive to run). GPT-3 had 175 billion; GPT-4’s exact count is undisclosed but rumored to be in the trillions.
In practice: When vendors brag about parameter counts, take it with a grain of salt — capability depends on data quality, architecture, and training method, not just size. The era of “bigger is always better” is over; today’s gains come from better data and smarter training.
18. Inference
Running a trained model on new input to get a prediction. “Training” is the learning phase; “inference” is the using phase. Every time you ask ChatGPT a question, that’s an inference call.
In practice: Inference is what’s expensive at scale. If you’re building a product on top of an LLM, your cost is roughly (number of users) × (average tokens used per request) × (price per token). Inference costs have fallen by ~10x per year since 2023 — designs that were uneconomical last year are now viable.
19. Hallucination
When a model confidently generates information that isn’t true. LLMs don’t “know” anything; they predict plausible-sounding text. Sometimes the plausible text is wrong — wrong facts, made-up citations, fictional case law. This isn’t a bug; it’s an inherent property of how these systems work.
In practice: Treat every AI output as a draft, not a source of truth. Verify facts, citations, and numbers before they go anywhere consequential. The 2023 Mata v. Avianca case — where a lawyer submitted ChatGPT-fabricated case law — is the canonical cautionary tale.
20. Guardrails
Technical and policy measures that constrain what a model will and won’t do. Guardrails include content filters (refusal to discuss violence), output format constraints (“always respond in JSON”), and tool restrictions (“don’t send emails without approval”). They’re how vendors balance model capability with safety.
In practice: When a model refuses to do something you asked, you’re hitting a guardrail. They’re imperfect — too tight and the model is unhelpful; too loose and it’s dangerous. If you’re building a product, you’ll spend significant time tuning them.
Section 3: Working with AI Day to Day
These are the terms you’ll actually use in conversations about getting things done with AI tools. If you only read one section of this glossary, read this one.
21. Prompt
The instruction you give an AI model. In the simplest form: a question. In practice: instructions, context, examples, output format, role assignment, constraints. The prompt is the entire interface — everything the model knows about what you want comes from what you wrote.
In practice: A useful prompt typically includes: (1) the role or context, (2) the task, (3) relevant background, (4) the desired output format. Compare “Summarize this” to “You are a product manager. Summarize this customer feedback in 3 bullets, focusing on bugs. Output as markdown.” The second prompt produces dramatically better results.
22. Prompt Engineering
The discipline of writing prompts that reliably get the output you want. Prompt engineering is closer to specific technical writing than to “coding” — it’s about clarity, structure, examples, and iteration. This is the single highest-leverage skill for most professionals using AI today.
In practice: Good prompt engineers work in iterations: write a prompt, evaluate 10 outputs, identify the failure pattern, adjust the prompt, repeat. They don’t try to write the “perfect prompt” — they write systems that handle variation. Many of Lumaire Level 2’s modules are built around this skill.
23. Retrieval-Augmented Generation (RAG)
A pattern where the model is given access to a knowledge base (usually a vector database) and retrieves relevant documents before generating an answer. RAG reduces hallucinations and lets a model answer questions about your company’s data without retraining the model.
In practice: If you’ve used a customer-support chatbot that knew about your account, or an AI tool that could summarize your internal docs, you’ve used RAG. It’s the dominant pattern for “make our AI know our stuff” in 2026.
24. AI Agent
A system that uses an LLM to plan and execute multi-step tasks, often calling tools (search, code execution, APIs) along the way. A simple chatbot answers questions; an agent can book flights, file tickets, or research a topic across multiple websites.
In practice: “Agent” is the most overused word in AI for 2025–2026. Some vendors use it for any tool with a slightly autonomous loop; others reserve it for systems that genuinely plan and recover from errors. Ask: “What tools can it use? Can it take action or just suggest? Does a human need to approve each step?“
25. Multimodal AI
A model that handles more than one type of input or output — text, images, audio, video. GPT-4o, Claude, and Gemini can all read images and respond in text; some can generate audio or video. Multimodal is the direction the entire field is moving.
In practice: For most office work, multimodal means “I can paste a screenshot of a chart and ask questions about it.” That alone is a major productivity unlock — no more transcribing visuals into text descriptions.
26. API (Application Programming Interface)
A way for software systems to talk to each other. In AI context, “the OpenAI API” or “the Anthropic API” is the programmatic door into a model — you send text in, get text back, programmatically. APIs are how products are built on top of LLMs.
In practice: You don’t need to use an API to use AI. But if you’ve ever wondered how a “custom AI workflow” works under the hood, it’s almost always: send a prompt to an API, parse the response, do something with it. APIs make AI composable.
27. Copilot
A category label, originally popularized by Microsoft’s GitHub Copilot, for AI assistants that work alongside you in an existing tool — your code editor, your email client, your document. The brand name has spilled into generic use: every major SaaS product now has a “copilot” feature.
In practice: Microsoft’s “Copilot” trademark doesn’t extend everywhere — Google has “Gemini in Workspace,” Notion has “Notion AI,” Salesforce has “Einstein Copilot.” Functionally, they all do roughly the same thing: AI assistance embedded in tools you already use.
28. AI Governance
The policies, processes, and oversight structures an organization puts around AI use. Governance covers data privacy, model risk, audit trails, bias mitigation, regulatory compliance, and human review. It’s the “boring” term that CFOs and legal teams actually care about.
In practice: If you’re deploying AI in a regulated industry — finance, healthcare, legal, government — governance is what makes the difference between a pilot and a production rollout. Most organizations underestimate it. AI courses for managers (Level 4 on Lumaire) typically include a governance module for this reason.
29. Alignment
The effort to make AI systems do what their designers intended, including following human values and intentions. “Alignment” is a loaded term — it spans technical research (“how do we build models that pursue the right goals?”), philosophical debate (“whose values?”), and practical concern (“will this model do what we want?”).
In practice: Most consumer products are “aligned” in the narrow sense (they don’t help with bioweapons, they don’t lie deliberately). The deeper alignment questions — about long-term AI safety, about whose values are encoded, about autonomy and control — are real research areas, but they don’t directly affect most day-to-day AI use yet.
30. AGI (Artificial General Intelligence)
Hypothetical AI that matches or exceeds human capability across virtually all cognitive tasks — not just narrow domains. AGI is the goal some labs are explicitly pursuing (OpenAI, DeepMind) and the philosophical horizon that frames much of AI policy and long-term safety discussion.
In practice: Nobody has built AGI. Some researchers think we’re decades away; some think it could happen in the next 5–10 years. As a working professional, AGI isn’t directly relevant to your day-to-day AI strategy — but it shapes the regulatory and competitive landscape you operate in. Mention it only when you want to start (or end) a long argument.
What’s Next
You don’t need to memorize all 30 of these to use AI well. The 8 terms that matter most for day-to-day work: LLM, prompt, prompt engineering, hallucination, context window, token, RAG, and agent. If you understand those and can ask smart questions about the rest, you’re ahead of 90% of professionals.
When you’re ready to move from vocabulary to skill, the natural progression is:
- AI Foundations (Level 1) — the basic mechanics, how to think about AI as a working tool, and the prompt patterns that show up everywhere.
- AI Productivity (Level 2) — practical workflows: writing, research, brainstorming, building custom GPTs and AI workflows you can use tomorrow.
Both courses are self-paced, lifetime-access, and built for people who don’t write code. No PhD required.
Have a term you’d add to this list? Tell us — we’re planning follow-up posts on prompt engineering, AI ROI, and the AI security stack, and your question might drive one of them.