Next-token prediction

The model is just guessing what comes next.

The model predicts one token at a time — short text chunks (a word, sub-word, or punctuation).
A learned stop token (<|END|>) signals end-of-turn.

System: / User: / Assistant: are training-time conventions. The model sees a flat string and continues it.
Fully stateless — every call re-sends the full history. No memory or learning between turns.

Tool definitions are added to the prompt; the model emits a tool call as text in a special format.
The runtime watches for that, executes the tool, and feeds the result back as a new turn.
This is the foundation of an agent.

The company is fictional, but the assistant is already committed partway into a real-looking URL (…sec.gov/news/press-release/) — the only plausible continuation is a specific release number, so the model invents one.
Every token is statistically likely; the result is a confident, official-looking link the model never looked up. Paste it in a browser: it either 404s, or resolves to a real press release about a completely unrelated company — both prove the point.
This is why you ground facts — attach the source document or give the model a search tool, rather than asking it to recall.
The grounded example is identical — same question, same pre-seeded reply ending at …press-release/ — but a search_web call and its result are now in the window. The model completes the URL from the tool response (2025-187-northridge-…), every time. Ungrounded it invents a different number on every run.