The AI Glossary: 60+ Terms Explained

Terms are grouped alphabetically. Where an article explains a term in depth, the definition links to it.

A – C

Agent: A system in which an AI model uses tools in a loop — taking actions, observing results and deciding what to do next — until a task is complete. Read more.
AGI (Artificial General Intelligence): A hypothetical AI able to match or exceed humans across essentially all cognitive tasks. There's no agreed definition or test, and experts disagree on if and when it will arrive.
Artificial Intelligence (AI): The broad field of building computer systems that perform tasks normally requiring human intelligence, such as understanding language, recognizing images or making decisions.
Alignment: Research and techniques aimed at making AI systems behave in line with human intentions and values — helpful, honest and harmless.
API (Application Programming Interface): A way for programs to talk to each other. Most AI models are used through an API: your code sends text, the provider's servers return the model's response.
Attention (self-attention): The mechanism that lets each token in a transformer gather information from other tokens, so meaning depends on context. Read more.
Benchmark: A standard test set used to compare models, e.g. on reasoning, coding or knowledge. Useful but imperfect — always test on your own tasks.
Bias: Systematic unfairness in AI outputs, often learned from biased training data. Read more.
Chain-of-thought: Having a model work through a problem step by step before answering, which improves accuracy on complex tasks. Read more.
Chunking: Splitting documents into smaller passages so they can be embedded and retrieved precisely in RAG systems. Read more.
CLIP: A model that maps images and text into a shared embedding space, enabling image search by description. Read more.
Context window: The maximum amount of text (in tokens) a model can consider at once — its entire working memory for a request. Read more.
Cosine similarity: A measure of how similar two vectors are, based on the angle between them. The standard way to compare embeddings. Read more.

D – F

Deep learning: Machine learning using neural networks with many layers. The approach behind nearly all modern AI.
Diffusion model: A type of generative model that creates images (or video, audio) by starting from random noise and removing it step by step, guided by a prompt. Read more.
Embedding: A vector of numbers representing the meaning of text, images or other data, so similar meanings have nearby vectors. Read more.
Evaluation (evals): Systematic testing of an AI application's outputs against defined criteria, to measure and track quality. Read more.
Few-shot prompting: Including a few examples of the desired input and output in a prompt so the model can follow the pattern. Read more.
Fine-tuning: Further training a pre-trained model on specific examples to change its behavior, style or skills. Read more.
Foundation model: A large model trained on broad data that can be adapted to many tasks — such as today's large language models.
Function calling: See Tool use.

G – L

Generative AI: AI that creates new content — text, images, audio, video or code — rather than only classifying or predicting.
GPU (Graphics Processing Unit): A processor that performs many calculations in parallel, making it ideal for training and running AI models. Read more.
Grounding: Basing a model's answer on specific provided sources rather than its general training, to improve accuracy and allow verification.
Guardrails: Rules and checks around an AI system that prevent unwanted inputs, outputs or actions.
Hallucination: When a model generates confident but false or made-up information. Read more.
HNSW: A graph-based index for fast approximate nearest-neighbor search, widely used in vector databases. Read more.
Hybrid search: Combining vector (semantic) search with keyword search so both meaning and exact terms are matched.
Inference: Running a trained model to produce outputs — as opposed to training it.
Large Language Model (LLM): A very large neural network trained on huge amounts of text to predict the next token, which lets it understand and generate language. Read more.
Latency: How long a response takes. For streaming models, often measured as time to first token and tokens per second.
LLM-as-judge: Using a language model to grade another model's outputs against a rubric. Read more.
LoRA (Low-Rank Adaptation): A parameter-efficient fine-tuning method that trains a small number of extra weights instead of the whole model, making fine-tuning much cheaper.

M – O

Machine learning (ML): Building systems that learn patterns from data instead of following hand-written rules.
MCP (Model Context Protocol): An open standard for connecting AI applications to external tools and data sources. Read more.
Model: The trained mathematical function — a large set of learned parameters — that turns inputs into outputs.
Multimodal: Able to work with more than one type of data, such as text, images and audio. Read more.
Neural network: A model made of layers of simple connected units whose connection strengths (weights) are learned from data.
Open-weight model: A model whose trained weights are publicly released so anyone can download and run it, subject to its license. Read more.
Overfitting: When a model (or a prompt) performs well on the examples it was tuned on but poorly on new data.

P – R

Parameters (weights): The learned numbers inside a model. Model size is usually given as a parameter count, such as "8B" for 8 billion.
Pretraining: The first, largest training stage, where a model learns from massive general data by predicting the next token. Read more.
Prompt: The input you give a model: instructions, context, examples and questions.
Prompt engineering: The practice of designing and iterating on prompts to get reliable, high-quality outputs. Read more.
Prompt injection: An attack where instructions hidden in content an AI reads try to override its intended behavior. Read more.
Quantization: Storing model weights (or vectors) with fewer bits to save memory and speed up computation, with a small loss of accuracy. Read more.
RAG (Retrieval-Augmented Generation): Retrieving relevant information from your own data and adding it to the prompt so the model can answer from it. Read more.
Reasoning model: A model trained to "think" through problems — often producing internal reasoning before its answer — improving results on complex tasks.
Reranker: A model that re-scores retrieved results against the query for higher precision, typically applied to the top results of a first search.
RLHF (Reinforcement Learning from Human Feedback): A training method that uses human preferences between responses to make models more helpful and safe. Read more.

S – Z

Semantic search: Searching by meaning rather than exact words, usually with embeddings. Read more.
Structured output: Constraining a model to return data in a defined format, such as JSON matching a schema.
System prompt: Standing instructions that set a model's role, rules and context for a whole conversation. Read more.
Temperature: A setting that controls randomness in generation: low is focused and consistent, high is varied and creative. Read more.
Token: A chunk of text — a word or part of a word — that a model processes. Pricing and limits are counted in tokens. Read more.
Tool use: A model's ability to request calls to external functions (search, databases, APIs), which your code executes. Also called function calling. Read more.
Training data: The data a model learns from. Its content, quality and cutoff date shape what the model knows.
Transformer: The neural network architecture, introduced in 2017, behind almost all modern language models; built around attention. Read more.
Vector: An ordered list of numbers. In AI, vectors represent data such as embeddings.
Vector database: A database that stores embeddings and quickly finds the ones most similar to a query. Read more.
Workflow: A fixed, code-controlled sequence of AI steps — more predictable than an agent. Read more.
Zero-shot: Asking a model to do a task with no examples, relying only on instructions.