Also available in: Português (Brasil)
AI glossary for beginners
Artificial intelligence has filled the news with strange words: token, embedding, RAG, hallucination, fine-tuning. This glossary explains the 30 most essential terms in plain language, each with a short definition and an everyday example. No formulas, just enough for you to follow any conversation about AI without getting lost.
| Term | What it is | Everyday example |
|---|---|---|
| Token | It is the chunk of text a model actually processes: it can be a word, part of a word or a punctuation mark. Language models do not read loose letters, they read this sequence of tokens. In English, one token roughly equals 4 characters. | The word dog may become 1 token, while a long, rare word is usually split into several. |
| LLM (large language model) | A model trained on huge amounts of text to predict the next word and, from that, write, summarize and answer. ChatGPT, Gemini and Claude are examples of it. | When you ask for a summary of an email and get a ready paragraph, an LLM generated it. |
| Prompt | It is the instruction or question you send to the model. The quality of the answer depends heavily on how clear the prompt is, and it can include context, examples and the desired format. | Writing list 5 dog names in alphabetical order is a prompt. |
| System prompt | A behind-the-scenes instruction, set by whoever builds the app, that guides the model before your message. It fixes the role, the tone and the rules, and you usually do not see it. | A bank's virtual agent has a system prompt telling it to stay formal and never ask for your password. |
| Context window | It is the maximum amount of tokens the model can consider at once, counting your message plus the answer. Anything beyond that limit gets cut or forgotten. Recent models have windows of hundreds of thousands of tokens. | Pasting a whole book can overflow the context window of a smaller model. |
| Parameters (weights) | The internal numbers tuned during training that store what the model learned. More parameters usually means more capability, and more cost too. Large models have billions of them. | Saying a model has 70 billion parameters is describing the size of its numeric brain. |
| Neural network | The computing structure, loosely inspired by the brain, made of layers of connected units that turn numbers into numbers. It is the base of almost all modern AI and learns by adjusting its connections. | Your phone's face unlock runs on a neural network. |
| Transformer | The neural network architecture that has ruled language AI since 2017, introduced in the paper Attention Is All You Need. It processes text in parallel using the attention mechanism. The T in GPT comes from it. | Modern machine translators use transformers to keep the meaning of a sentence. |
| Attention (self-attention) | The mechanism that lets the model weigh which words in a sentence matter most to each of the others. It is what allows it to tell who a pronoun refers to. It is the heart of the transformer. | In the cat climbed up because it was hungry, attention links it to cat. |
| Embedding | The representation of a text as a list of numbers, a vector, that captures its meaning. Similar texts end up with nearby vectors. This is how a machine compares meaning, not just identical words. | A meaning-based search finds automobile when you look for car because the embeddings are close. |
| Vector and vector database | A vector is the list of numbers of an embedding. A vector database stores millions of them and quickly finds the ones most similar to a query. It is the memory behind meaning-based search. | An assistant that remembers your documents usually keeps them in a vector database. |
| Pre-training | The first and most expensive phase, when the model reads a gigantic amount of text to learn the language in a general way. At this stage it is not specialized in anything yet. Fine-tuning only comes later. | It is like someone who read half a library before choosing a profession. |
| Fine-tuning | Training a ready-made model a little more on specific data to make it better at a task or a style. It costs far less than training from scratch and adapts the model to your domain. | A firm fine-tunes a model on its contracts so it writes in the house style. |
| LoRA | A lightweight fine-tuning technique that changes only a small part of the model instead of all of it. It saves memory and time and is very popular for customizing image generators. | Creating your own illustration style without retraining the whole model usually uses LoRA. |
| Inference | The moment when the already trained model is used to produce an answer. Each answer consumes computation and therefore costs money. Unlike training, it happens every single time you use the model. | Every message you send to a chatbot triggers an inference. |
| Temperature | A control over the randomness of the answer, starting at 0. Near 0 the text is more predictable and repetitive; higher up it gets more creative and unpredictable. It tunes how bold the model is. | For a poem you raise the temperature; to generate an ID number you lower it. |
| Top-p (nucleus sampling) | Another randomness control that limits the choice to the most likely words that together add up to a fraction p of the probability. With a low p the model stays conservative. It often works alongside temperature. | A top-p of 0.9 keeps the plausible options and drops the bizarre ones. |
| Hallucination | When the model answers something that sounds convincing but is false or made up. It happens because the model predicts likely text, not verified truth. That is why important facts are worth checking. | A model citing a book that does not exist is a hallucination. |
| RAG (retrieval augmented generation) | The system first searches for relevant passages in a trusted source and only then asks the model to answer based on them. It reduces hallucination and lets the model use up-to-date data. | A support bot that answers by quoting the product manual usually uses RAG. |
| Zero-shot | Asking the model to do a task without giving any example, relying only on what it already learned. It is the most common everyday use and works well for simple tasks. | Asking translate this into Spanish, with no examples, is zero-shot. |
| Few-shot | Placing a handful of examples inside the prompt to show the pattern you want before asking for the result. It usually improves accuracy and format, and a few examples already help a lot. | Showing 3 emails in your tone before asking for a new one is few-shot. |
| Quantization | Reducing the precision of the model's numbers, for example from 16 to 4 bits, so it takes less memory and runs faster. It loses a little quality in exchange for lightness and lets big models run on modest machines. | Running a model on your laptop is usually only possible with the quantized version. |
| Distillation | Training a small model to imitate a large one, transferring much of its capability into a cheaper package to run. The small one becomes the student of the teacher. It powers fast, economical versions. | The mini and flash versions of famous models are often distilled. |
| GPU | The graphics card, a chip that does many calculations in parallel. It was built for games but became the engine of training and running AI. Without GPUs, today's models would be unfeasible. | The scramble for GPUs grew precisely because of the AI race. |
| Diffusion (diffusion models) | The technique behind most image generators. The model starts from random noise and cleans it up step by step until the requested image forms. It is now used for video and audio too. | Asking for an astronaut cat in watercolor and getting the finished art uses diffusion. |
| Multimodal | A model that understands or generates more than one type of data, such as text, image, audio and video together. It can look at a photo and answer about it, which greatly widens its uses. | Sending a photo of your fridge and asking for recipes with what is inside needs a multimodal model. |
| AI agent | A system that uses a model to plan and carry out several steps on its own, calling tools such as search, a calculator or email. It goes beyond answering: it acts toward a goal. It still makes mistakes and needs supervision. | Asking it to research flights and build an itinerary can be an agent's job. |
| Tools and function calling | The model's ability to call external functions, such as querying a database or an API, when it sees the need. The model decides when to call and with what data. It is what connects AI to the real world. | A chatbot that shows the real tracking of your parcel is using tools. |
| Overfitting | When the model memorizes the training examples instead of learning the general pattern and, because of that, does poorly on new cases. It is a common risk when tuning with little data. The sign is acing training and failing in real life. | A model that only recognizes the cats from photos it already saw has overfit. |
| Benchmark | A standardized test used to compare models on tasks like math, code or general knowledge. It helps measure progress but does not tell the whole story of real use. A high score does not always become practical value. | When a launch claims it beat the competitor, it is almost always quoting a benchmark. |
How these definitions were checked
Every entry was verified against at least two reliable technical sources, among the documentation of OpenAI, Google Cloud and Hugging Face, plus the papers that founded the field, such as Attention Is All You Need (Vaswani and colleagues, 2017), which introduced the transformer, and the retrieval augmented generation paper (Lewis and colleagues, 2020), which gave RAG its name. The goal was to keep the explanation simple without sacrificing accuracy.
Last updated: · Methodology and sources