1 / 5

Lesson · 12 min

Next-token prediction

The engine under the chat UI is not a database and not a mind.

A large language model writes the next token, then the next, using everything in the context as the prefix. That is the whole trick. It is an extraordinary trick — it can look like reasoning, retrieval, even taste — but it is not looking up a row in a table of true facts. It is continuing a pattern.

This is why Grok can write a sonnet and also invent a plausible citation. Both are 'what usually comes next' in language that looks like that. Truth is a property you enforce with tools, retrieval, and checks — not a guarantee of the decoder.