AI Glossary
From tokens and context windows to MCP and reasoning models — the key AI terms, explained simply.
33 terms
Core concepts
- Artificial Intelligence (AI)Artificial Intelligence
- The field of computer science that enables machines to perform tasks requiring human-like intelligence — such as understanding language, recognizing images and making decisions. Full explanation
- Machine Learning (ML)Machine Learning
- An approach to AI where a model learns patterns from data instead of being explicitly programmed, improving as it sees more examples. Full explanation
- Deep LearningDeep Learning
- A subset of machine learning using multi-layer neural networks; it drove the leap in image, speech and language quality and underpins every modern LLM. Full explanation
- Large Language Model (LLM)Large Language Model
- A massive neural network trained on vast text corpora that can understand and generate language; ChatGPT, Claude and Gemini are examples. Full explanation
- TokenToken
- The small unit of text a model reads and generates — roughly ¾ of an English word or 1–2 Persian words. Both context length and API pricing are counted in tokens. Full explanation
- Context WindowContext Window
- The maximum amount of text (in tokens) a model can "see" at once in a conversation — your prompt, files and the whole chat history included. The larger it is, the more the model remembers per session. Full explanation
- ParameterParameter
- The learned weights inside a model where its knowledge and abilities are stored; their count (e.g. 700 billion) indicates the model’s scale. Full explanation
Architecture & models
- TransformerTransformer
- The neural architecture that processes whole text in parallel via the attention mechanism — the foundation of every large language model since 2017. Full explanation
- Mixture of Experts (MoE)Mixture of Experts
- An architecture that splits a model into several "experts", activating only the relevant ones per input — yielding much larger models at lower inference cost. Full explanation
- MultimodalMultimodal
- A model that understands or generates images, audio and video in addition to text — like Gemini viewing a chart and explaining it. Full explanation
- EmbeddingEmbedding
- A numeric (vector) representation of text or images that preserves meaning; similar concepts get nearby vectors — the basis of semantic search and RAG. Full explanation
- Vector DatabaseVector Database
- A database built for fast similarity search over embeddings, enabling "most-similar" content retrieval — usually the external memory of a RAG system. Full explanation
- Small Language Model (SLM)Small Language Model
- A lightweight language model (a few billion parameters) that runs on laptops and phones; cheaper and faster than an LLM for narrow tasks. Full explanation
- Diffusion ModelDiffusion Model
- The image-generation architecture that starts from noise and step-by-step denoises it into the final picture — underlying tools like Midjourney. Full explanation
Training & tuning
- Pre-trainingPre-training
- The costly initial stage where the model trains on vast public text and acquires general language and knowledge — after it, the model is still a generalist. Full explanation
- Fine-tuningFine-tuning
- Additional training of a pre-trained model on specialized data so it performs better in a specific domain or style — e.g. one brand’s customer support. Full explanation
- RLHFRLHF
- A technique where humans rank the model’s outputs and it learns to give more helpful, accurate and safe answers — the technical core behind ChatGPT’s and Claude’s behavior. Full explanation
- LoRALoRA
- A cheap fine-tuning method: instead of training all parameters, a few small matrices are added — enabling model customization at a fraction of the cost. Full explanation
- DistillationDistillation
- Transferring knowledge from a large model to a small one: the small model learns from the large model’s outputs to behave similarly at much lower cost. Full explanation
- QuantizationQuantization
- Reducing the precision of a model’s internal numbers (e.g. 16-bit to 4-bit) so its size and memory needs shrink dramatically and it runs on ordinary hardware — with minor quality loss. Full explanation
Agents & usage
- AI AgentAI Agent
- A system that takes a goal and autonomously plans and executes the steps — calling tools, writing code, reading results and correcting its course. Full explanation
- Agentic AIAgentic AI
- The 2025–2026 shift: models don’t just answer, they "do work" — chaining multiple tools and steps over minutes or hours; Devin and Claude Code are examples. Full explanation
- MCP (Model Context Protocol)Model Context Protocol
- An open standard for connecting agents to tools and data — "the USB-C of AI". Write one MCP server and any compatible agent can use it. Full explanation
- RAGRAG (Retrieval-Augmented Generation)
- A pattern where relevant documents are retrieved from a knowledge source and fed to the model before answering — grounding responses in real, current data instead of static training memory. Full explanation
- Prompt EngineeringPrompt Engineering
- The craft of writing effective instructions for a model: providing the right context, examples and output format to get more accurate, reliable results. Full explanation
- InferenceInference
- Running the trained model — the moment you send a prompt and tokens are generated; the per-message cost of ChatGPT is inference cost. Full explanation
- Vibe CodingVibe Coding
- The new natural-language style of programming: instead of writing code, you describe what you want and the agent builds, tests and fixes it — the outcome of the 2025–2026 coding-agent wave. Full explanation
Reasoning, safety & quality
- Reasoning ModelReasoning Model
- A model that "thinks" before answering — producing a chain of internal reasoning and far higher accuracy on hard math, code and logic problems; slower and costlier in return. Full explanation
- Test-time ComputeTest-time Compute
- Giving a model more compute while answering (longer thinking) in exchange for better results — the phenomenon that moved the quality frontier of 2026 reasoning models. Full explanation
- HallucinationHallucination
- When a model confidently produces false information — a wrong source, non-existent reference or invented number. Mitigations: RAG, reasoning models and citing checkable sources. Full explanation
- GuardrailsGuardrails
- Control layers around a model that block harmful, inappropriate or off-policy outputs — separate from the model itself, usually as input/output filters. Full explanation
- AlignmentAlignment
- Training models to stay in step with human values and intent — helpful and honest rather than deceptive or harmful; a core concern of frontier labs. Full explanation
- AGIAGI (Artificial General Intelligence)
- A hypothetical AI that matches or beats humans across all cognitive tasks — not just specific domains. Not yet achieved, and its exact definition is disputed among labs. Full explanation