From autocomplete to assistant

Here's the one idea in this lesson. Pretraining is where the model gets what it knows. Post-training is where it learns how to behave. Everything below is those two sentences, slowed down.

A base model continues documents. That's all pretraining taught it. So if you type a question, it asks itself: what kind of page starts like this, and what comes next on that page? Often the answer is “more questions”. Here's the same prompt sent to a base model and to an assistant.

Both are good next-token guesses. Only one is useful to you. Turning the left into the right is called post-training.

First, the model is shown a large set of example conversations: a person asks, a helpful assistant answers. It's the same next-token exercise as pretraining, but now the text is always a chat, so the model learns to continue one as the assistant. Each example has the same shape you'll meet next lesson: the person's turn, the assistant's turn, and often a hidden note up top saying how to behave.

Then it's polished by preference. The model writes several answers to the same prompt. People rank them, and later an AI judge does most of the ranking against written principles. The model is nudged towards the answers that rank well, over and over, so being helpful, honest and careful becomes its habit. Anthropic's version of that judge-with-principles is called Constitutional AI, and the principles are published as Claude's constitution.

One more thing current assistants learn: to think before they answer. Given a hard question, Claude can write out its working first and the reply after. That working is produced the same way as everything else, one token at a time. It's just text the model writes to itself.

Underneath all of this, it's still predicting the next token. It's just predicting what a helpful assistant would say. Which is why most of what Claude knows comes from pretraining, and most of how it acts comes from post-training.