AI glossary

AI terms in plain English

Tokens, context windows, RAG, MCP, fine-tuning — every AI term you keep meeting, defined in plain English. Searchable, free.

83 terms·8 categories·Updated as the field shifts

How to use this glossary

Search for a specific term, or click a category pill to focus on one area. Click any term to read the full definition with examples and links to related compares and skills.

Terms are defined the way someone shipping AI work uses them — not the textbook version. If a term is contested or vague, we say so.

Foundations

13 terms

Core LLM concepts: tokens, context windows, prompts, hallucinations, frontier models.

AGI / ASI

AGI (Artificial General Intelligence) is AI matching humans across most cognitive tasks. ASI (Artificial Super Intelligence) exceeds humans across all of them. Both terms are contested and there's no agreed definition of when either has arrived.

Context window

The maximum amount of text (measured in tokens) a model can consider at one time — the prompt plus its response.

Frontier model

The leading-edge models at any given moment — defined by capability, not parameter count. In April 2026 that's GPT-5, Claude Opus 4.7, Gemini 3 Pro, and a handful of others.

Hallucination

When an LLM produces fluent, confident-sounding output that is factually wrong or fabricated — fake citations, invented APIs, made-up case law, plausible-but-false biographies.

LLM (Large Language Model)

A neural network trained on huge amounts of text to predict the next token. The engine behind ChatGPT, Claude, Gemini, and most modern AI tools.

Open source AI

Per OSI's 2024 Open Source AI Definition: weights, training code, and enough information about training data for someone to reproduce the model. Most 'open' models in 2026 do not actually qualify.

Open weights

A model whose trained weight files are publicly downloadable, so anyone can run it locally or fine-tune it. It does not mean the training data or training code is public.

Prompt

The input you send to an LLM. In a chat interface, it's your message; in an API call, it's the full text the model sees, including any system instructions and conversation history.

System prompt

A persistent instruction that defines the model's persona, rules, and behavior for an entire conversation — separate from the user's individual messages.

Temperature

A sampling parameter controlling randomness in an LLM's output. 0 is deterministic and conservative; ~1 is creative and varied; above 1 starts to incoherent.

Token

The atomic unit an LLM reads and writes — usually a sub-word fragment. "The" is one token; "tokenization" is three.

Tokenizer

The component that splits raw text into tokens before the model sees it. Different models use different tokenizers, so the same word can cost more on one model than another.

Top-p (nucleus sampling)

An alternative or complement to temperature: sample only from the smallest set of tokens whose cumulative probability exceeds p (e.g. 0.9). Cuts off the long tail of unlikely tokens dynamically.

Architecture

7 terms

How models are built: Transformers, attention, MoE, diffusion, multimodal.

Attention

The mechanism that lets a model decide which parts of its input matter most for each token it produces. Every Transformer is built around it.

Decoder-only model

A Transformer that just predicts the next token, autoregressively. The architecture used by GPT, Claude, Llama, Gemini, and basically every modern chat model.

Diffusion model

A generative model that learns to reverse a noising process. Start with pure noise, denoise it step by step, end up with an image. The mechanism behind Stable Diffusion, Flux, Midjourney, and most modern image generators.

Encoder-decoder

The original Transformer architecture: an encoder that reads the input and a decoder that generates the output, connected by cross-attention. Still used for translation and some summarization, mostly extinct for chat.

Mixture of Experts (MoE)

A sparse Transformer variant that routes each token to only a few "expert" sub-networks instead of running the whole model. Lets you scale total parameters way up without scaling cost proportionally.

Transformer

The neural network architecture behind every modern LLM. Introduced in 2017, it processes sequences using self-attention instead of recurrence.

Vision-Language Model (VLM)

A model that processes both images and text in a unified way. Show it a screenshot, a chart, or a photo, and it can describe, analyze, or answer questions about what it sees.

Training & adaptation

11 terms

Pre-training, fine-tuning, RLHF, DPO, LoRA, quantization, distillation.

Constitutional AI

Anthropic's alignment technique where the model critiques and revises its own outputs against a written set of principles, reducing the need for human preference labels.

Distillation

Training a smaller 'student' model to mimic a larger 'teacher' model, getting most of the teacher's quality at a fraction of the size and cost.

DPO (Direct Preference Optimization)

A simpler alternative to RLHF that trains directly on preference pairs without needing a separate reward model or reinforcement learning loop.

Fine-tuning

Fine-tuning is continued training on a pre-trained model with a smaller, curated dataset, adapting it to a specific task, domain, or style.

LoRA (Low-Rank Adaptation)

A parameter-efficient fine-tuning method that trains tiny adapter layers and freezes the base model, cutting fine-tuning cost by 100x or more.

Post-training

The umbrella term for everything done to a model after pre-training — SFT, RLHF, DPO, safety tuning, persona work — where most modern model differentiation actually happens.

Pre-training

The first and costliest training phase: the model learns to predict the next token over trillions of tokens, gaining knowledge but not assistant behavior.

Quantization

Reducing the numerical precision of a model's weights — for example FP16 down to INT8 or INT4 — so it runs on smaller, cheaper hardware.

RLAIF (RL from AI Feedback)

RLHF where the human ranker is replaced (or augmented) by an AI judge, making preference training cheap enough to run at massive scale.

RLHF (Reinforcement Learning from Human Feedback)

RLHF (reinforcement learning from human feedback) is a post-training technique that uses human rankings of model answers as the reward signal for reinforcement learning. It is the method OpenAI used to turn a GPT-3.5 model into ChatGPT.

Synthetic data

Training data generated by AI models rather than collected from humans — increasingly the dominant data source in modern post-training.

Inference & reasoning

8 terms

Chain of thought, reasoning models, test-time compute, ReAct, in-context learning.

Chain of thought (CoT)

A prompting technique that asks the model to reason step-by-step before giving a final answer, dramatically improving accuracy on math, logic, and multi-step problems.

Few-shot / Zero-shot / N-shot

Prompting styles defined by how many examples you include before the actual task: zero examples (zero-shot), a few (few-shot), or N specific examples (N-shot).

In-context learning

The phenomenon where an LLM "learns" a new task from examples shown in the prompt, without any gradient updates or training — the model adapts purely from context.

ReAct (Reasoning + Acting)

An agent pattern where the model interleaves thinking and tool use in a loop: reason about what to do, take an action, observe the result, reason again.

Reasoning model

A class of LLM trained to think for an extended period — sometimes minutes — before producing a final answer, trading latency for much higher accuracy on hard problems.

Self-consistency

Running the same prompt multiple times with sampling and taking the majority answer — a simple way to boost reasoning accuracy by trading more compute for better results.

Test-time compute

The principle that you can make a model smarter by spending more compute at inference time — not just at training time — by having it think longer, sample more, or search more.

Tree of thought (ToT)

A reasoning technique where the model branches into multiple candidate thoughts at each step, evaluates them, and prunes — exploring a tree of reasoning paths instead of one chain.

RAG & retrieval

8 terms

Vector databases, semantic search, chunking, reranking, hybrid search.

Chunking

Splitting source documents into smaller pieces before embedding so each chunk is small enough to retrieve precisely and fit in the prompt.

Embedding model

A model that turns a piece of text (or image, or audio) into a fixed-length vector, where similar inputs produce similar vectors. Different from the LLM that generates answers.

Hybrid search

Combining semantic (embedding) search with keyword search (BM25) and fusing the results — the production-default retrieval strategy in 2026.

Knowledge cutoff

The date after which a model has no training data. Anything after the cutoff, the model either does not know or hallucinates with confidence.

RAG (Retrieval-Augmented Generation)

RAG (retrieval-augmented generation) is a technique that gives an LLM the most relevant passages from your own documents at question time, so it answers from that text rather than from memory. No fine-tuning is required: change the documents and the answers change with them.

Reranking

A second-pass model that takes the top-K results from initial retrieval and reorders them by true relevance to the query.

Semantic search

Search by meaning instead of keyword overlap — uses embedding models to find documents that are conceptually similar to the query, even with no shared words.

Vector database

A vector database stores embeddings, the numeric vectors that represent text, images, or audio, and returns the entries closest in meaning to a query vector. It is the retrieval layer most RAG systems search, answering in milliseconds across millions of documents.

Agents & tools

16 terms

AI agents, tool use, MCP, computer use, Skills, custom GPTs, memory, multi-agent.

.cursorrules / repo instruction files

Plain-text files (.cursorrules, AGENTS.md, CLAUDE.md) at the repo root that tell coding agents how your codebase works — conventions, dependencies, gotchas.

AI agent

An AI agent is a system in which a language model directs its own work: it picks a tool, acts, reads the result, and repeats until the task is done. The model, not a fixed script, decides each next step and when to stop.

AI memory

Persistent context an AI assistant carries across separate conversations — your preferences, past projects, recurring facts — so you do not have to re-explain yourself every time.

AI team topology

The pattern of using multiple AI models for different roles in a single workflow — typically architect, builder, reviewer — instead of one model doing everything.

Artifact

Claude's term for generated content rendered as a separate, persistent object — code, a document, a chart, a webpage — instead of just inline chat text.

Build-more, verify-more

A practical principle for AI-assisted work: as builds get faster with AI, verification has to scale with them — otherwise quality degrades silently.

Computer use

Computer use is an AI model's ability to operate a computer through its screen: the model reads screenshots, then moves the mouse, clicks, and types to finish tasks in ordinary desktop and browser apps.

Cross-model handoff

Passing context, decisions, and intermediate work from one AI model to another — the connective tissue in a multi-model workflow.

Custom GPT

OpenAI's user-built assistants — a base GPT model plus custom instructions, knowledge files, and optional API actions, packaged as a shareable bot.

Function calling / Tool use

Function calling (also called tool use) is the ability of an LLM to reply with a structured JSON call to a function you defined, instead of plain text. Your code runs the function and returns the result, which is how a text model searches the web, queries a database, or calls any API.

MCP (Model Context Protocol)

MCP (Model Context Protocol) is an open standard for connecting AI applications to external tools, data, and services: you build an integration once as an MCP server, and Claude, ChatGPT, Cursor, or any other MCP client can use it. Anthropic released it in November 2024 and donated it to the Linux Foundation's Agentic AI Foundation in December 2025.

Model verification

Using a second AI model to check the first model's output — a structural defense against hallucinations and confident-wrong answers.

Multi-agent system

A setup where multiple specialized LLM agents collaborate — one plans, another codes, another reviews — instead of one model trying to do everything.

Orchestration

Coordinating multiple LLM calls, tools, retrievals, and conditional logic into a working AI pipeline — the plumbing layer between "raw model" and "shipped product."

Skill (Claude Skills)

A Claude Skill is a folder containing a SKILL.md instruction file, plus optional scripts and references, that teaches Claude one specific, repeatable task. Unlike a prompt you paste into every chat, a skill loads only when your request matches its description, and it needs no fine-tuning or running server.

Workflow (in AI)

A pre-defined sequence of LLM calls and tool steps with fixed control flow — the developer decides the order, not the model. Distinct from an agent, which lets the model wander.

Modalities

6 terms

Text-to-image, text-to-video, ControlNet, image LoRA, latent space, OCR.

ControlNet

A Stable Diffusion technique that lets you condition image generation on a structural input — a depth map, pose skeleton, edge map, or scribble — alongside the text prompt.

Image LoRA

A small fine-tuning file (typically 50-300MB) that bends a base diffusion model toward a specific style, character, or subject. Stack them at inference to combine effects.

Latent space

The compressed numerical representation that diffusion models actually work in — not pixels, but a smaller learned embedding. The 'Latent' in Stable Diffusion's name.

OCR (Optical Character Recognition)

Extracting machine-readable text from images of text — scanned documents, photos of receipts, screenshots, PDFs. Increasingly handled by general vision-language models instead of dedicated OCR engines.

Text-to-image

Generative AI that turns a text prompt into an image. The category covers tools like Midjourney, DALL-E, Stable Diffusion, Flux, and Ideogram.

Text-to-video

Generative AI that turns a text prompt (or an image) into a short video clip. The leading systems in 2026 are Sora, Runway, Pika, Veo, and Kling.

Safety, eval & ops

14 terms

Benchmarks, jailbreaks, prompt injection, alignment, red teaming, latency, pricing.

Alignment

The technical problem of getting AI systems to actually do what humans want — follow intent, respect values, and avoid harmful behavior.

API (in AI context)

Programmatic access to a model — the HTTP endpoint you POST a prompt to and get a completion back. This is what you actually integrate.

Benchmark

A standardized test of model capability — same questions, same scoring, run across many models so you can rank them.

Eval / evaluation

The broader practice of measuring model output quality on tasks you actually care about — usually a custom test suite specific to your app.

Guardrails

Runtime safety filters wrapped around an LLM's input and output to catch bad prompts and bad responses before they escape.

Inference provider

A third party that hosts open-weight models behind an API so you do not have to run GPUs yourself. Often cheaper than self-hosting at small or medium scale.

Jailbreak

A prompt or technique that bypasses an LLM's safety training to get outputs the model would normally refuse.

Latency

Time from sending a request to receiving the response. For LLMs, split into time-to-first-token (TTFT) and total completion time.

MMLU (Massive Multitask Language Understanding)

A 57-subject multiple-choice benchmark covering everything from US history to abstract algebra. The most-cited general knowledge test of the LLM era.

Prompt injection

Prompt injection is an attack in which text an LLM reads, typed by a user or hidden in a webpage, email, or file, overrides the instructions its developer gave it. It is ranked #1 (LLM01:2026) in the 2026 OWASP Top 10 for LLM applications, published in August 2026.

Rate limit

A provider-imposed cap on how many requests or tokens you can use per minute or day. The thing that breaks production at scale.

Red teaming

Proactive adversarial testing of an AI system before release — paid attackers try to break it so the lab can fix issues.

Self-hosting

Running open-weight models on your own infrastructure instead of calling a hosted API.

Token cost

How API providers price model usage — dollars per million tokens, with input tokens cheaper than output tokens.