The words people use

Most AI writing is not difficult because the ideas are hard. It is difficult because it uses a private vocabulary without explaining it. Here is that vocabulary.

The core

Model — the trained thing that produces answers. Sometimes called an LLM (large language model). It is a file, or a service, not a person.

Parameters — the numbers inside a model. Billions of them, and they are what training adjusts. "8B" means eight billion parameters.

Weights — another word for those parameters, used more when talking about downloading a model. "Open weights" means the file is available.

Training — the process that produces a model, by adjusting its parameters against enormous amounts of text.

Inference — using a trained model to produce an answer. Training is expensive and happens once; inference is cheap and happens every time you type.

Cutoff — the point in time where a model's training data ends. It knows nothing after it.

Token — the unit models actually work in. Roughly three-quarters of a word, so 1,000 tokens is about 750 words. It matters because context limits and prices are both measured in tokens.

Context window — how much text a model can consider at once — your prompt plus the conversation so far. Exceed it and the earliest parts fall out of view.

Prompt — what you type. System prompt — instructions set in advance that shape every answer.

Hallucination — fluent, confident output that is not true. Not lying: there is no intent, and that is what makes it dangerous.

Multimodal — can handle images or audio as well as text.

The marketing terms

Frontier model — the most capable and most expensive models, made by the largest labs.

Reasoning model — a model that works through a problem in steps before answering. Slower and more expensive, better at problems with a right answer.

Agent — a model given tools and allowed to take actions in a loop, rather than just answering.

Harness — the program that drives a model like an agent. The model is the engine; the harness is the car.

Open weights — the model file is downloadable. Not the same as open source. See open weights are not open source.

MoE (mixture of experts) — a design where only part of a large model runs for each token. More capability for the same memory; usually faster too.

The running-it-yourself terms

Local — running a model on your own machine. Hosted — using someone else's, over the internet.

Quantisation — compressing a model so it fits in less memory, at some cost in quality. See what Q4 costs you.

VRAM — memory on your graphics card. This is the number that decides which models you can run.

Runtime — the program that loads and runs a model. Ollama, LM Studio, llama.cpp. See runtimes.

GGUF — a common file format for quantised models you run yourself.

The retrieval terms

Embedding — turning text into numbers so that similar meanings sit near each other. The basis of search over your own documents.

Vector database — somewhere to keep those numbers so they can be searched quickly.

RAG (retrieval-augmented generation) — giving a model relevant passages from your own documents before it answers, so it works from real material instead of memory. The standard fix for hallucination.

Fine-tuning — further training a model on your own examples so it behaves differently. Powerful, and often the wrong first answer — try prompting and RAG first.

Benchmarks

Benchmark — a standard test used to score models. Useful, frequently misleading. See why the numbers do not describe your machine.

Arena — a site where people compare two model answers blind and vote. The most honest ranking available.

Where to go next

• What is a model, really? — the concept behind all of it
• Try it today — three ways to start, free