Glossary

Every term this site uses, in plain language. Nothing here needs a technical background.

chatbot	A program you type to, which types back. The ones this site is about are built on a language model.
language model	The part that produces the words. It was trained on a very large amount of text, and it works by predicting what could plausibly come next.
prompt	Whatever you type in. Also used as a verb: to prompt is to give it something to work from.
token	A fragment of text the model works in. Sometimes a whole word, often part of one. “unhelpful” might be three tokens.
context window	How much of the current conversation the model can take into account at once. Finite. Older turns fall out of it.
training	The one-off process, finished before you arrived, that set the model's internal values. Talking to it does not repeat this.
hallucination	The common name for the model stating something untrue in a confident tone. Not lying: it has no separate sense of what is true.
weights	The numbers inside the model that training set. They do not change while you are talking to it.
retrieval	Looking something up in a store of documents. Some tools can do this in addition to generating text; generating text is not itself retrieval.
system prompt	Standing instructions the tool's makers put in front of your conversation, which you usually cannot see.