AI vocabulary is unusually bad at explaining itself. Half the terms are metaphors, a third are marketing, and several mean different things depending on who is speaking.
Here are the ones that actually come up, in plain language.
The basics
LLM (large language model) — a model trained on enormous amounts of text to predict what comes next. The thing underneath ChatGPT, Claude and Gemini.
Token — the unit a model works in: roughly a word fragment. Common words are one token, unusual ones several. Pricing is per token, and it is why models miscount letters — they never see letters.
Prompt — what you type. Prompt engineering — writing prompts deliberately, which mostly means giving material, audience, length and constraints.
Context window — how much text the model can consider at once, including your conversation and anything you pasted. Exceed it and the earliest parts fall out, which is why long chats seem to "forget" the beginning. They have not forgotten; that text is no longer in front of them.
Inference — running a model to get an answer, as opposed to training it.
Parameters — the numbers inside a model, usually quoted in billions. Bigger correlates with more capable, imperfectly. A newer small model routinely beats an older large one.
Behaviour
Hallucination — stating something false with confidence. Not a malfunction: the model produces plausible text, and plausible is all you get when it lacks the information.
Temperature — how adventurously the model samples its next word. Low is repetitive and predictable, high is varied and eventually incoherent. It is why the same prompt gives different answers.
Grounding — restricting answers to documents you supply. The most effective defence against hallucination.
RAG (retrieval-augmented generation) — a system that searches your documents, puts the relevant parts into the prompt, then answers. How most business chatbots avoid inventing your refund policy.
Fine-tuning — further training on your own data to change how a model behaves. Different from RAG: fine-tuning changes the model, RAG changes what it sees.
System prompt — hidden instructions set by the product, shaping tone and limits before you type anything.
Capability
Multimodal — handles more than text: images, audio, video as input or output.
Reasoning model — tuned to work through problems step by step before answering. Slower, dearer, better on hard problems.
Chain of thought — the model writing out its intermediate steps. Improves accuracy on multi-step problems and lets you check the working.
Agent — a model given tools and the ability to act: reading files, calling APIs, running commands, iterating toward a goal. The word is heavily overloaded in marketing.
MCP (Model Context Protocol) — an emerging standard for connecting models to external tools and data.
Models and access
Open weights — you can download and run the model. Open source would also mean training data and pipeline, which almost nobody releases. The distinction matters.
Quantisation — compressing a model to lower numerical precision so it fits in less memory. Small quality cost, large practical gain. What makes local models possible.
API — the interface a provider opens so other software can use its model. You send a request, you get a result.
Distillation — training a smaller model to imitate a larger one. How the fast, cheap tiers get good.
Safety and evaluation
Alignment — making a model behave as intended. RLHF — the main technique: humans rate outputs, the model learns from the ratings.
Guardrails — restrictions on what a model will produce. Jailbreak — getting around them.
Benchmark — a standardised test. Treat published tables sceptically: test questions leak into training data, and margins are often noise. More on that.
Red teaming — deliberately trying to make a model misbehave, before users do.
Words to be sceptical of
AGI — artificial general intelligence. No agreed definition, which makes claims about its arrival unfalsifiable.
"Reasoning" — the models produce reasoning-shaped text that improves accuracy. Whether that constitutes reasoning is a philosophical argument, not a technical one.
"Understands" — it predicts. Whether prediction at this scale amounts to understanding is exactly the disputed question, so treat the word as a claim rather than a description.
"Trained on your data" — could mean fine-tuning, or RAG, or just a long prompt. Ask which.
Common questions
What does LLM stand for? Large language model — a model trained on text to predict what comes next.
What is a token in AI? The unit models work in, roughly a word fragment. Pricing is per token.
What is a context window? How much text a model can consider at once. Exceed it and the earliest content drops out.
What is the difference between RAG and fine-tuning? RAG changes what the model sees at question time. Fine-tuning changes the model itself.
What is an AI agent? A model with tools and the ability to act rather than just reply. The term is used very loosely.
Are open weights the same as open source? No. Open weights means you can run the model; the training data and pipeline are generally not released.
What is temperature? How adventurously the model picks its next word. It is why the same prompt gives different answers each time.