Large language models — the AI behind ChatGPT, Claude, Gemini, and others — are software systems trained to work with text. They don't search a database. They don't look things up. They generate. And the way they generate is stranger than most people realize. The whole operation runs on a single repeated question: given everything I've seen so far, what word comes next?

The model never holds the complete sentence in mind. It has no plan for where the sentence is going. It just looks at what's already there and asks: what token is statistically most likely to follow this? The coherence you read isn't the result of understanding — it's the result of patterns so deeply learned that local decisions (this token, then this one, then this one) produce globally sensible text. Most of the time. When it fails — when it confidently states a wrong fact, or a sentence drifts off-topic — that's not a malfunction. That's the system working exactly as designed, just landing on a plausible-but-wrong token.

Where do the probabilities come from? Training. Before you ever typed a prompt, the model processed hundreds of billions of tokens of text — books, websites, code, forums — and adjusted billions of internal numerical weights to get better at predicting what comes next. The weights encode something like: after the token sequence 'the capital of France is', the token 'Paris' should have very high probability. Not because the model knows geography. Because it saw that pattern enough times that predicting 'Paris' became the right move. The knowledge is real. The understanding isn't. And that distinction matters more than the AI industry usually admits.