Tokens, Training and Next-Word Guessing: AI Jargon in Plain English

Every AI conversation now assumes you know a dozen borrowed words. Nobody stops to define them, so everyone nods along. Here are the twelve terms that actually matter, in plain English, with zero maths. Bookmark this; send it to the colleague who's been nodding.

The Core Six

LLM (Large Language Model)

The engine behind ChatGPT, Claude and Gemini. A program trained on enormous amounts of text until it becomes extremely good at one task: predicting the next word. Everything it does — answering, coding, summarising — is that one task, repeated. (Watch a miniature one do it in our free Micro LLM.)

Token

The chunks a model actually reads and writes — usually word-pieces, not whole words. "Unbelievable" might be three tokens: un-believ-able. Why care? Because AI pricing, speed and limits are all counted in tokens. A token is roughly ¾ of an English word.

Training

The counting phase: the model reads mountains of text and adjusts itself millions of times to get better at predicting what comes next. Training happens once, in a data centre, at vast cost. When you chat with a model, it is not training on the fly — it's frozen, using what it learned.

Parameters (or "weights")

The adjustable dials inside the model — billions of them — whose settings are everything it learned. "A 70-billion-parameter model" just means 70 billion dials. More dials generally means more capability and more cost, not automatically more accuracy.

Prompt

Everything you give the model before it responds. Not a magic spell — just the context that shapes its predictions. Better input, better predictions; that's the whole discipline covered in The Anatomy of a Perfect Prompt.

Hallucination

When a model states something false, fluently and confidently. Not a bug being fixed next quarter — a direct consequence of "predict the plausible next word". When truth and plausibility diverge, the model follows plausibility. This is why you verify anything that matters.

The Useful Six

Context window

The model's working memory: how much conversation and documentation it can consider at once, measured in tokens. Exceed it and the earliest material effectively falls out — the real reason long chats "forget" the beginning.

Fine-tuning

Taking a trained model and giving it extra training on specific material — legal documents, your company's style — so it specialises. Expensive-ish, permanent-ish. Most teams get what they need with better prompting instead.

RAG (Retrieval-Augmented Generation)

The trick behind "chat with your documents": search your files for relevant passages first, paste them into the prompt invisibly, then let the model answer using them. Search plus prompting, wearing a fancy acronym.

Inference

The running phase — the model generating output for you. When people say "inference costs", they mean the electricity and computing spent answering, as opposed to training.

Agent

A model wired to tools and given a goal, so it can take multiple steps on its own — search, write, run code, check the result — rather than just answering once. The word doing the most hype-lifting in 2026, but the concept is real.

Open-source model

A model whose weights you can download and run yourself — free, private, on your own hardware — versus a closed model you access through a company's service. Both have their place; the free tier of our directory is full of the former.

The Sentence That Organises All of It

A model is trained (adjusting parameters) to predict the next token; you steer it with a prompt inside its context window, it sometimes hallucinates, and wiring it to tools makes it an agent.

Twelve terms, one sentence. And if you want the intuition underneath the vocabulary — the thing that makes it stick — spend ten minutes training your own tiny model in Micro LLM, then read How LLMs Actually Work.