AI Glossary

Every AI term you'll hear in 2026 — explained simply.

AGI: Artificial General Intelligence — AI that matches human reasoning across any task.; AGI is the long-term goal of building AI that can learn and reason at human level on any task, not just narrow ones. No system has reached AGI yet.
Agent: An AI that takes actions on your behalf, not just answers questions.; Agents loop: plan → call a tool → observe → repeat, until the goal is done. Examples include ChatGPT Operator and Cursor background agents.
Context window: How much text an AI model can read at once.; Measured in tokens. GPT-5 handles ~400k, Claude 4 handles 200k, Gemini 2.5 Pro handles 2M. Larger windows let you paste long docs.
Embedding: A numeric fingerprint of text used for search and memory.; Embeddings map text into vectors so similar meanings sit close together. They power retrieval, search and recommendations.
Fine-tuning: Training a base model on your own data for a specific task.; You start with a pretrained model and continue training on labelled examples. Cheaper than training from scratch, more accurate than prompting alone.
Hallucination: When an AI confidently makes something up.; Hallucinations happen because models predict likely text, not verified truth. Reduce them with grounded sources, lower temperature, and tools like Perplexity.
LLM: Large Language Model — the engine behind ChatGPT, Claude and Gemini.; LLMs are trained on huge text datasets to predict the next token. Everything you say to ChatGPT goes through an LLM under the hood.
Multimodal: AI that handles text, images, audio and video together.; GPT-4o, Gemini and Claude can all read images. Multimodal models can describe a photo or analyse a chart in one prompt.
Prompt engineering: The craft of writing prompts that get the best answers.; Good prompts give context, examples, format and constraints. It is a real skill — small tweaks change quality a lot.
RAG: Retrieval-Augmented Generation — letting AI read your documents at answer time.; RAG searches a knowledge base for relevant passages, then feeds them to the model with the question. Used by NotebookLM and Perplexity.
Token: A chunk of text — roughly 0.75 words.; Models read and bill by token. 1,000 tokens ≈ 750 words. Long prompts and long answers both cost tokens.
Temperature: A dial for AI creativity vs predictability.; 0 = deterministic, safe. 1+ = creative, surprising. Use low for code and facts, higher for creative writing.