The dictionary of artificial intelligence

AI system able to act autonomously to reach a goal.

An AI agent combines a language model with tools (web search, code execution, API access) to carry out multi-step tasks without human intervention. Unlike a simple chatbot, it plans, acts, observes the result and self-corrects.

Making AI behaviour match human values and intent.

AI alignment is the discipline of ensuring AI systems do what we actually want, not just what we technically asked for. It covers training methods (RLHF, DPO), benchmarking, red-teaming and policy. It's a hard, open research problem — especially as systems become more capable.

How much text a model can read at once (input + output).

The context window is the maximum number of tokens a model processes in a single call. Recent models reach 200k–1M tokens. A wider window enables long-document analysis, complex agents, and large-codebase reasoning — at the cost of latency and price.

A vector representation of meaning.

An embedding is a list of numbers (e.g., 1536 dimensions) that captures the meaning of a piece of text. Similar texts produce similar vectors. Embeddings power semantic search, RAG, and recommendation systems.

Adapting a base model on your own dataset.

Fine-tuning continues training a pretrained model on domain-specific data so it specialises (tone, format, vocabulary). Modern variants (LoRA, QLoRA) are cheaper and only update a fraction of weights. Often less needed than a good RAG + system prompt.

When a model produces a confident but false answer.

An LLM hallucinates when it generates plausible-sounding content that isn't grounded in reality (false references, made-up facts, wrong code). Mitigation: RAG, source citations, output validation, lower temperature on factual tasks.

Large language model trained on massive text corpora.

An LLM is a neural network with billions of parameters trained to predict the next token. By scale and training methods, it acquires useful general capabilities (writing, reasoning, coding). Examples: GPT-4o, Claude 4.5, Gemini 2, Llama 3.

A model that handles multiple modalities (text, image, audio, video).

A multimodal model can ingest and/or produce text, images, audio and video. Examples: GPT-4o (text + image + voice), Claude 4.5 (text + image), Gemini (full multimodal). It enables use cases like vision Q&A, image generation from text, voice agents.

The instruction sent to a model.

A prompt is the text input you provide. The quality and structure of the prompt have a major impact on the response: be specific, give examples (few-shot), specify the role and the output format. Prompt engineering is the discipline of optimising prompts.

Attack where malicious content overrides the system prompt.

Prompt injection happens when a user (or untrusted document) crafts text that tells the model to ignore its instructions or leak data. It's a major risk for agents that read external content (websites, emails, files). Defences: input filtering, sandboxing tools, separation of trust contexts.

Augmenting an LLM with retrieved external knowledge.

RAG combines a retrieval step (vector or keyword search over your documents) with generation: the LLM answers using only the retrieved context. It reduces hallucinations, supports up-to-date information without retraining, and enables source citation. The most common pattern in production GenAI.

Training a model on human preferences.

RLHF is a fine-tuning method where humans rank model outputs, and the model is trained to favour preferred outputs. It's a key technique behind ChatGPT, Claude and Gemini. Modern variants include DPO (Direct Preference Optimization), simpler and often as effective.

Sub-word units a model processes (and bills you for).

A token is a chunk of text (word, sub-word, character) the model processes. Models bill in tokens (input + output). Rough rule: 1 token ≈ 4 characters in English ≈ 0.75 word. Watch out: French, code, and JSON often use more tokens per word.

Neural network architecture behind nearly all modern LLMs.

Introduced by Google in 2017 (paper « Attention Is All You Need »), the transformer uses an attention mechanism to weigh each part of the input against the others. It scales massively and is the foundation of GPT, Claude, Gemini, Llama and most modern LLMs.

A database optimised to store and search embeddings.

A vector DB indexes embeddings to find the nearest ones to a query (semantic search). Examples: Pinecone, Weaviate, Qdrant, ChromaDB, pgvector (PostgreSQL extension). The backbone of RAG.