AI Glossary: 50+ Terms You’ll Actually Use

AI writing is full of shorthand. This glossary keeps it plain. Each term gets a one-line definition — no math, no marketing.

Core concepts

LLM (large language model) — a model trained on vast text to predict the next token. See What Is an LLM?
AGI (artificial general intelligence) — a hypothetical AI that matches human ability across most tasks. Doesn’t exist yet; the label is heavily contested.
Token — a chunk of text (~0.75 of a word) that models read and write.
Parameters / weights — the numbers a model learned during training; a rough measure of capacity.
Context window — how many tokens a model can hold in memory at once.
Inference — running a trained model to produce output (as opposed to training it).
Prompt — the input you give a model; prompt engineering is the craft of writing it well.

Transformer — the neural-network design behind nearly all modern LLMs (2017).
Pretraining — the initial, expensive phase where the model learns language from raw text.
Fine-tuning — further training on curated data to specialize behavior.
RLHF — reinforcement learning from human feedback; humans rank outputs to steer preferences.
Temperature — a dial controlling randomness; low = deterministic, high = creative.
Hallucination — a confident but false output (the model made it up).

RAG (retrieval-augmented generation) — look up relevant documents first, then generate an answer grounded in them.
Fine-tuning vs. RAG — teaching vs. giving the model a reference sheet.
Quantization — shrinking model weights to lower precision so they fit on weaker hardware.
Embeddings — turning text into vectors of numbers so “similar” means something mathematically.
Zero-shot / few-shot — asking a model to do a task with zero or a few examples.
Chain-of-thought — prompting the model to “think step by step” before answering.

Open-weight — the trained weights are public (Llama, Mistral, DeepSeek); closed — they aren’t (GPT, Claude, Gemini).
Frontier model — the most capable models at any given moment.
Agent — an LLM wired to tools (search, code, browsers) that can act, not just answer.
Multimodal — a model that handles text, images, audio, and/or video.

MMLU — a broad knowledge test; HumanEval — coding ability; GSM8K — grade-school math.
Perplexity — how surprised a model is by text; lower is better.
Latency / throughput — speed (time per response) and volume (tokens per second).

For the companies behind these terms, see Nvidia vs OpenAI. To run a model on your own machine, see Run LLMs Locally.