A language model is a prediction engine. It was trained on a vast corpus of text — books, articles, code, conversations — and learned to predict: given these words, what word is most likely to come next? That process, repeated billions of times across billions of text sequences, produced a model that can complete sentences in ways that feel like understanding.
When you ask Claude a question, it isn't looking up the answer in a database. It isn't searching the internet. It's generating the most likely continuation of your conversation, token by token, based on patterns absorbed during training. The output feels like retrieval because human writing — which Claude learned from — tends to follow predictable patterns of question and answer.
The model doesn't know things the way you do. It has no beliefs, no memory of previous conversations, and no understanding in the philosophical sense. What it has is an extraordinarily rich map of how language patterns connect — which is often indistinguishable from understanding, and sometimes isn't.
Understanding token prediction explains many things that otherwise seem mysterious: why Claude can write convincingly about topics it gets wrong (fluency and accuracy are different things), why it performs better when given more context (more tokens means a clearer prediction signal), and why it can be confidently wrong (the most probable continuation isn't always the correct one).
Before continuing, write down one thing Claude has done that confused or surprised you. By the end of this module, you should be able to explain it in terms of what you've just learned about token prediction.