Lesson 1 of 7
From rambler to assistant
10 min read
Straight out of training, a model given “What's the capital of France?” is about as likely to reply “What's the capital of Italy? What's the capital of Spain?” as to answer. It has read half the internet — so why won't it just answer you?
The machine that only autocompletes
The model that comes out of pretraining — the base model — has exactly one skill: continue text. It learned by playing guess the next word across a huge slice of the internet, so that's all it does. Ask it a question and it doesn't see a request to answer; it sees a piece of text to continue — and online, a question is often followed by more questions, or a list, or a tangent. Flip the switch below between the raw model and the finished assistant on the same prompt.
The base model isn't broken — it's doing exactly what it learned: completing text. It was never taught that a question is a request for an answer. That behaviour has to be added.
A few thousand good examples
You add it with a second, much smaller training stage: finetuning. Instead of the whole internet, you show the model a curated pile of example conversations — a person asks something, and an ideal answer follows — hand-written or carefully checked. It keeps playing guess the next word, but now on this tidy set of good behaviour, so it learns the shape of a helpful reply: address the question, be clear, and stop when you're done.
Finetuning barely adds knowledge — the model learned almost everything it knows back in pretraining. What it adds is behaviour: the habit of answering as a helpful assistant instead of rambling on.
Curate the example pile
So what's actually in that curated pile? Each example quietly teaches one habit: this is how you answer, this is when you stop, this is how you admit you're unsure. Build the pile yourself below — switch example types on and off, and watch the same base model's reply change. Turn answer the question off and it slides back to rambling; add behaviours and each one shows up in the reply. And keep an eye on the knowledge bar underneath — it never budges.
The assistant's whole manner — whether it answers, stops, admits doubt, sounds warm — is assembled from the example types you include. Nothing new is learned about the world; only how to behave.
Why one base model can become anyone
This is why a single base model can be turned into a terse coder, a patient tutor, or a playful companion — just by changing the pile it copies. The manner is chosen at training time. It also hints at something trickier: because the model is trained to sound like a confident helper, it can end up preferring a confident-sounding wrong answer over stepping out of character to say “I don't know.” A polished manner is picked in training; being right is not guaranteed by it.
Under the hood this stage is supervised finetuning (SFT), the first step of post-training — everything done to a base model after it reads the internet. It's cheap next to pretraining: a small, high-quality set of examples, not another pass over the whole web. The rest of this course is about the training that comes after this step.
Recap
- —A base model can only continue text — ask it a question and it may just autocomplete more questions
- —Finetuning on a small set of example conversations teaches it to answer as a helpful assistant
- —The example types you include compose the assistant's manner — answering, stopping, warmth, admitting doubt
- —This stage adds behaviour, not knowledge — the same model underneath, taught how to respond
A team takes a fresh base model and, with no finetuning, asks “How do I reset my password?” It replies with a list of *other* support questions instead of an answer. What's going on?
Continue in the app
Take the whole How AI Learns to Think course — tracked
Get your personalized path, progress and streaks in the app — this lesson and every next one, in order.