AI glossary
AI concepts explained simply: machine learning, deep learning, LLMs, NLP, RAG, embeddings, fine-tuning, MCP…
The basics
- Artificial intelligence (AI)
- The set of techniques that let a machine perform so-called intelligent tasks: understanding text, recognizing an image, making a decision.
- Machine learning
- A branch of AI where the machine learns from examples instead of being programmed rule by rule. For instance, from thousands of emails labelled "spam" or "normal", it learns to sort them on its own. Classic methods (regression, decision trees) are mostly used on tabular data.
- Deep learning
- A branch of machine learning that uses neural networks with many layers. It needs a lot of data and compute, but excels at text, images and sound. All recent models (LLMs, image generators, speech recognition) rely on deep learning.
- Neural network
- A mathematical model inspired by the brain, made of connected layers of "neurons". By adjusting the weights of these connections during training, the network learns to turn an input (text, image) into an output (an answer, a category).
- Model
- The result of training: a file of parameters (the weights) that can perform a task. You download it, call it through an API, or run it yourself. See the guide →
The main fields
- NLP (natural language processing)
- Everything about text: writing, summarizing, translating, classifying, extracting information, answering questions.
- Computer vision
- Everything about images and video: recognizing objects, classifying a photo, detecting faces, generating images.
- Multimodal
- A model that handles several kinds of data at once: describing an image in text, answering a question about a photo, generating a video from text.
- Generative AI
- Models that produce new content (text, image, video, sound, code) rather than only classifying or predicting.
Language models
- LLM (large language model)
- A deep learning model trained on huge amounts of text to predict how a text continues. It powers Claude, GPT, Mistral or Llama. See the guide →
- Transformer
- The neural network architecture behind almost all current language models. Its attention mechanism lets it look at all the words of a text at once and link each one to the others, instead of reading them one by one.
- Classifier
- A model that puts an input into a category: spam or not, positive or negative review, type of customer request. It does not write text, it picks a label from a list.
- Encoder
- A model (or part of a model) that reads a whole text and turns it into numbers that sum up its meaning. Embedding models and many classifiers are encoders, such as BERT. See the guide →
- Decoder
- A model (or part of a model) that writes text one token at a time, each time predicting the next token. LLMs such as Claude, GPT, Mistral or Llama are decoders. See the guide →
- Token
- The piece of text an LLM reads, often part of a word. Model limits (context size) and prices are counted in tokens.
- Context window
- How much text (in tokens) a model can read at once: the question, instructions, documents and answer.
- Hallucination
- A wrong answer given confidently by a model that makes things up instead of saying it doesn't know. RAG and guardrails reduce it.
- Inference
- Using an already trained model to get an answer, as opposed to training. See the guide →
- Quantization
- Lowering the precision of a model's numbers (e.g. from 16 to 4 bits) so it uses less memory and runs on smaller hardware, at a slight cost in quality. See the guide →
- Tokenizer
- The tool that splits a text into tokens before giving it to the model. Each model family has its own: the same text does not count the same number of tokens everywhere.
- Base model
- A model fresh out of pre-training: it can continue a text but does not follow instructions well yet. It is then tuned into an instruct model. See the guide →
- Instruct model
- A model tuned to follow instructions and hold a conversation. It is the version used in an assistant. See the guide →
- Open weights
- A model whose parameter files can be downloaded: you can run it yourself and tune it. Its license sets what you are allowed to do with it. See the guide →
- Latency
- The waiting time before the model answers, especially before the first token appears. See the guide →
- Throughput (tokens per second)
- How fast a model writes its answer. It depends on the model, the hardware and how many requests are served at once. See the guide →
Using a model
- Prompt
- The text sent to the model to ask for something. A reusable prompt contains variables, such as {{product}}. See the guide →
- Context (system instruction)
- Instructions and examples given to the model before the question, to set its role, tone and rules, without retraining it. See the guide →
- Few-shot
- Giving a few examples (input and expected output) in the context to show the model what you want. See the guide →
- RAG (retrieval-augmented generation)
- Having a model answer from your documents: relevant passages are retrieved, then given to it with the question. See the guide →
- Embeddings
- Lists of numbers (vectors) that represent the meaning of a text. Texts with similar meaning get close vectors: the basis of semantic search and RAG. See the guide →
- Structured output
- Forcing the model to answer in a precise format, usually JSON described by a schema. See the guide →
- Temperature
- The setting that controls randomness in answers: low, the model answers in a stable, predictable way; high, it varies more. See the guide →
- Zero-shot
- Asking the model for a task without giving it any example, with the instruction alone. See the guide →
- Chain-of-thought
- Having the model reason step by step before its final answer, which improves calculation and logic tasks. See the guide →
- Function calling (tool calling)
- A model's ability to request a function call (search, calculate, query a database) with structured parameters, then use the result. See the guide →
- Chunking
- Cutting documents into pieces before indexing them for RAG. The size of the pieces and their overlap change the quality of the answers. See the guide →
- Vector database
- A database that stores embeddings and finds the texts closest to a question. It is the search engine behind RAG. See the guide →
- Reranker
- A model that re-sorts the passages found by the search to keep only the most relevant ones before giving them to the LLM. See the guide →
Acting and tooling
- Agent
- An AI assistant that chains several steps and uses tools to reach a goal, instead of answering in one go. See the guide →
- MCP (Model Context Protocol)
- An open standard to plug tools and data (files, databases, software) into an AI assistant, compatible with Claude, Cursor and others. See the guide →
- Workflow
- An automated chain of steps where AI takes part, built in a tool such as n8n or LangGraph. See the guide →
Training and adapting
- Dataset
- A set of examples used to train, adapt or evaluate a model. See the guide →
- Pre-training
- The first training phase of a language model: it reads huge amounts of text and learns to predict the next word. It is very expensive and only model publishers do it; you then start from this pre-trained model to specialize it. See the guide →
- Fine-tuning
- Retraining an existing model on your own examples to specialize it in a domain. See the guide →
- LoRA
- A low-cost fine-tuning method that trains only a small part of the model. QLoRA applies it to a quantized model. See the guide →
- Evaluation (benchmark)
- A set of tests to measure the quality of a model, a prompt or an agent. See the guide →
- SFT (supervised fine-tuning)
- The most common form of fine-tuning: the model is shown questions and the expected answers so it learns to answer the same way. See the guide →
- QLoRA
- A variant of LoRA that loads the model quantized to 4 bits during tuning, to train a large model on a single graphics card. See the guide →
- Epoch
- One full pass over all the examples of the dataset during training. Too many epochs lead to overfitting. See the guide →
- Synthetic data
- Training examples produced by a model rather than written by people. Quick to get, they must be reviewed and filtered. See the guide →
- Distillation
- Training a small model to imitate the answers of a larger one, to get a cheaper, faster model for a given task. See the guide →
- RLHF
- Tuning from human feedback: people rank answers, and the model learns to produce the ones they prefer. See the guide →
Reliability and compliance
- Guardrail
- A rule that controls what goes into or out of an AI assistant: blocking personal data, refusing off-topic requests. See the guide →
- Prompt injection
- An attack that hides instructions in a message or document to turn the model against its rules. See the guide →
- Red teaming
- Deliberately attacking an AI system to find its weaknesses before users do. See the guide →
- AI Act
- The European regulation on artificial intelligence: it sets transparency, documentation and oversight obligations depending on the system's risk level. See the guide →
- GDPR
- The European regulation on personal data protection, which also applies to data sent to an AI or used to train it.
- LLM as a judge
- A model used to score another model's answers against written criteria, when an exact comparison is not possible. See the guide →
Resource terms
- JSONL
- A text file with one example per line, each line a JSON object. It is the standard format for fine-tuning datasets. See the guide →
- System prompt
- The message placed before the conversation that sets the model's role, tone and rules. In a chat dataset, it is the message with the “system” role. See the guide →
- Harness
- The setup around a coding assistant (Claude Code, Codex, Cursor…): its instructions, what it may do without asking, what it is forbidden to do, and its automatic checks. See the guide →
- Hook
- A command run automatically at a given point in an agent's work, for example the tests after each file change. See the guide →
- GGUF
- A file format for quantized models, read by llama.cpp, Ollama and LM Studio to run a model on a computer. See the guide →
- VRAM
- The graphics card's memory. It limits the size of model you can run: a model too large for the VRAM will not load or becomes very slow. See the guide →
- DPO (direct preference optimization)
- A tuning method that shows the model, for the same question, a preferred answer and a rejected one, to steer it toward the wanted style. See the guide →
- Overfitting
- When a model memorizes its examples instead of learning the rule: excellent on the training examples, poor on new cases. A separate test set reveals it. See the guide →
- Test set
- A share of the examples set aside and never shown during training, to measure whether the model really improved. See the guide →