Yesterday we learned that LLMs "probabilistically predict the next token." But a model is a computer. Computers don't understand letters. They only understand numbers. So how does a sentence like "The weather is nice today" become numbers that go into the model, and how does it come back out as letters?
Today we'll trace how text is processed inside an LLM using four core concepts: Token → Embedding → Context Window → Inference. These four terms appear directly on the AIF exam and are also the keys to understanding cost and performance. Our goal is to understand without formulas—like drawing a picture.
LLMs don't receive sentences all at once