AI Glossary for Beginners: The Words You Need and How They Fit Together

AI has its own language. After one minute of reading about it, you meet words like LLM, token, RAG, agent and MCP. Nobody explains them. Everybody seems to know them already.

This article changes that. It explains the main words and abbreviations of AI in simple language. More important, it shows how the words connect. A single word is hard to remember. A word that has a place in a picture is easy to remember.

Here is how to read it. Each section teaches one idea and ends with a box of key words. Seven diagrams show the ideas in motion. They play by themselves when you scroll to them. You can pause them, or move step by step with the buttons. At the end you find a cheat sheet with all the abbreviations, and a simple plan for your first steps.

You do not need to know how to code. You do not need to know math.

The big picture: AI, machine learning and generative AI

Start with the biggest word. Artificial intelligence, or AI, means computer systems that do things we call smart: understanding language, recognizing images, making decisions. The word is old. A chess program from the 1990s is AI. So is the filter that sorts spam out of your email.

Inside AI is a smaller idea: machine learning (ML). In normal software, a person writes the rules. In machine learning, nobody writes the rules. The computer looks at many examples and finds the patterns by itself. Show it a million emails that people marked as spam, and it learns what spam looks like.

Inside machine learning is deep learning. It uses a neural network: a large web of small calculation units, arranged in layers. "Deep" means many layers. Deep learning is behind most of the AI news you hear today.

Inside deep learning is generative AI. Classic AI mostly sorts or predicts: spam or not spam, cat or dog. Generative AI makes new things: a text, an image, a song, a piece of code. An image generator is generative AI. So is a chatbot.

The last ring is the large language model (LLM). It is generative AI that works with text. When people say "AI" in daily talk, they most often mean a chat app that runs on an LLM.

How a model learns: data, training and inference

A model is the part that does the work. Think of it as a very big formula with many adjustable numbers inside. These numbers are called parameters. Some people say weights. Before training, the numbers are random, and the model gives nonsense. After training, the numbers hold what the model learned.

Training is how the numbers get good. You give the model a huge amount of data: examples. The model makes a guess and checks how wrong it was. Then it changes its numbers a tiny bit. It does this again and again, and slowly the mistakes get smaller.

Training is the expensive part. It needs a lot of computing power, usually from special chips such as GPUs (graphics processing units). A GPU can do many calculations at the same time. That is why training the biggest models costs so much.

A large model usually learns in two stages. First comes pre-training: it learns from a huge pile of general data. Later comes fine-tuning: it gets extra training on a smaller set of examples, so it is better at one job or one style.

When training is finished, the model is fixed. Now you can use it. Using a trained model is called inference. Each time you type a message and press Send, inference happens. The model does not learn from your message while you chat. If an app seems to remember you, that is a feature of the app. It is not the model changing.

How a chatbot writes: tokens, context and prediction

Now the part that surprises most beginners. An LLM does one simple thing: it predicts the next piece of text. It reads the text so far, guesses what comes next, adds it, and guesses again. It has learned so much from its training data that the result looks like a real conversation.

The pieces are called tokens. A token is a small part of text: a short word, a part of a long word, or a punctuation mark. In English, one token is about four characters, or three quarters of a word. This is a rule of thumb, and it differs from model to model. Tokens matter because models count, limit and charge by tokens.

A model can only see a limited amount of text at once. This space is the context window. It holds your prompt (the text you send), the earlier messages of the chat, and the answer that is being written. If a conversation is longer than the window, the app must cut or shorten the oldest text.

Apps often add a hidden instruction at the start. This is the system prompt. It sets the role and the rules, for example "You are a friendly support assistant." You usually do not see it, but it changes how the model answers.

Another word you will meet is temperature. It controls how adventurous the model is when it picks the next token. A low temperature gives safe, predictable text. A high temperature gives more variety, and also more risk.

Be careful with one thing: hallucination. The model predicts likely text. It does not have a built-in sense of what is true. So it can write something that sounds sure and is wrong, for example a book that does not exist. Check important facts. This is the most useful habit for a beginner.

You will also see model names. GPT stands for Generative Pre-trained Transformer. It is the name of the model family from OpenAI. "Transformer" is the design of the neural network behind most LLMs. Claude, Gemini and Llama are other model families. A model that works with more than text, such as images, audio or video, is called multimodal.

Giving a model your own knowledge: embeddings and RAG

An LLM only knows what it saw in training. It does not know your price list, your notes or yesterday's news. There are two common ways to fix this.

The first is fine-tuning, which you know from above. It is mostly used to change the style or behavior of a model, not to add many facts.

The second is retrieval-augmented generation, or RAG. The idea is simple. Before the model answers, the app searches your documents. It finds the best passages and puts them into the prompt. Then the model answers with these passages in front of it.

But how does the app find the right passages? It cannot search for exact words only. A question about a "refund" should also find a text about "getting your money back". For this, RAG uses embeddings. An embedding turns a piece of text into a long list of numbers. The numbers work like a place on a map. Texts with a similar meaning get places that are close together. "Puppy" lands near "dog" and far from "truck".

A vector database stores these places and finds the closest ones very fast. "Vector" is the math word for a list of numbers. Now you can see the whole RAG path: search, find, add to the prompt, answer.

From chatbot to agent: tools, MCP and skills

A chatbot only talks. An agent can also act. It uses an LLM as its brain, and it works in a loop: think, act, observe the result, and repeat until the goal is done.

To act, an agent uses tools. A tool is a small function, such as "search the web", "read a file", "check a calendar" or "send an email". The model does not run a tool by itself. It writes a request, for example "call the calendar tool for Friday". The app runs the tool and sends the result back to the model. This is called tool calling, or function calling.

If every app connects tools in its own way, it is a lot of work. MCP (Model Context Protocol) is an open standard that gives AI apps one common way to connect to tools and data. Think of a USB port: one plug shape, many devices.

Skills are written instructions for one specific task. An agent loads a skill only when it needs it, so its context stays small. I wrote a longer article about this, with a real example: What's the difference between an AI skill and an AI agent?

One more word that every builder meets: API. It means Application Programming Interface. It is the way one program talks to another. When a developer puts an AI model into an app, the app calls the AI company's API. To use an API, you get an API key. It is a secret password. Keep it secret. I learned this the hard way: Someone pulled an API key out of my old app.

Words about quality and safety

Some words are not about how AI works, but about how well it works. They come up in news and in product pages, so it helps to know them.

How it all fits together

You now know the parts. Here is one request from start to end. Say you ask an assistant app: "Find a quiet cafe near my office that is open until 8."

You type the message in a chat app. The app sends it to a model through an API. It adds the system prompt and the earlier chat. Together, this is the context. If the app has your documents, it uses RAG first and adds the best passages.

Then the agent reads the context and thinks: "I need fresh facts." It asks for a tool to search, connected with MCP. The result goes back into the context. At last, the LLM writes the answer, token by token, on GPUs in a data center. The app shows it to you.

Every word in this glossary names one step on that path. The diagram shows the layers. Each layer uses the one below it.

Cheat sheet: the abbreviations

Come back to this part when you meet an abbreviation and do not remember it.

• AIArtificial IntelligenceComputer systems that do tasks that seem smart.
• MLMachine LearningAI that learns patterns from examples.
• DLDeep LearningMachine learning with neural networks that have many layers.
• NLPNatural Language ProcessingThe field of teaching computers to work with human language.
• LLMLarge Language ModelA model for text. The engine inside chatbots.
• GPTGenerative Pre-trained TransformerA family of language models from OpenAI.
• RAGRetrieval-Augmented GenerationSearch your documents first, then let the model answer with them.
• RLHFReinforcement Learning from Human FeedbackPeople rate answers, and the model is tuned to give better-rated ones.
• MCPModel Context ProtocolAn open standard to connect AI apps to tools and data.
• APIApplication Programming InterfaceThe way one program talks to another program.
• GPUGraphics Processing UnitA chip that does many calculations at once. Used to train and run models.
• AGIArtificial General IntelligenceA future AI that could do any thinking task a person can.
• STTSpeech-to-TextTurns spoken words into written text.
• TTSText-to-SpeechTurns written text into a spoken voice.

Your first steps

You do not need a course to begin. This path works for most people. Take as much time for each step as you need.

1. Use an assistant every day. Pick one chat app, such as ChatGPT, Claude or Gemini. Use it for real tasks: summarize a long email, explain a topic, plan a trip. You learn most by using it.
2. Learn to write a good prompt. Say who the answer is for. Give context. Say the format you want. Show an example. If the answer is not right, tell the model what to change. These four habits help more than any trick.
3. Check the answers. Ask for sources and open them. Remember hallucination.
4. Give it your own text. Upload a document and ask questions about it. This is close to RAG: the model answers from your text, not only from memory.
5. Try a tool or an agent. Use a feature where the assistant searches the web or works with files. Watch the loop: think, act, observe.
6. Build one small thing. If you want to code, call a model through an API from a small script. Keep the API key out of your app's public code.

Keep this page open in a tab. When a new word appears, find its box or its card. After a few weeks, you will not need it anymore.