How an LLM Works

A language model processes your prompt as tokens — small chunks of text, not whole words. For each position it assigns scores to every token it knows, picks one, appends it to what it is writing, and repeats the whole cycle. The answer grows left to right, one piece at a time.

Training and answering are two different jobs. Training adjusts the model's internal numbers using vast amounts of example text — that is expensive and happens once. Answering (inference) only uses those learned numbers; it does not retrain anything. This is why a model can sound authoritative about something that happened after its training: it cannot learn from your conversation the way you learn from it.

The key insight: the model is a predictor of likely text, not a database of facts. It learned which words tend to follow which other words, and it is very good at continuing patterns. Fluency comes free; truth does not.

Try completing this sentence in your head: “The cat sat on the ___.” Several endings fit — mat, rug, couch. Your brain picked the most likely one for a generic sentence. A language model does the same thing, over and over, to write whole paragraphs.

Now notice: the most likely ending is not a fact about any real cat. It is just the most common continuation. That gap — between “likely text” and “true statement” — is the entire reason AI hallucinations exist, and it is exactly what the animation above shows: the model picks likely next tokens, nothing more.

Fluent text is not proof. This page shows the shape of generation, but the real machinery — tokenizers, attention, sampling, billions of weights — is deliberately collapsed into six boxes. The simplification is honest but wide.

A model can explain, summarize, or draft with remarkable skill, but important claims still need checking against real sources. Nothing about the animation above measures what any commercial model would actually output for a given prompt — two models, or two settings, can answer the same prompt differently.