The Gap: Completion Is Not Conversation
Lesson 1 of 3 in From Base Model to Assistant.
Take a Base model straight out of pre-training — the artifact the last domain spent seven modules building — and send it a question: What is the capital of France?
A well-trained base model may not answer. It may instead produce more questions: What is the capital of Spain? What is the capital of Italy? — continuing what looks like a geography worksheet. Or it may append (a) Paris (b) Marseille (c) Lyon, because your question also resembles the start of an exam item. Or it may write a paragraph about the question, forum-style. All of these are classic base-model behaviors, and none of them is a malfunction.
It is the objective working perfectly. Pre-training taught the model exactly one skill: given a prefix, predict what comes next in documents like the ones it has read — Next-token prediction over the pre-training Data mixture. And in that corpus, a lone question is at least as often followed by more questions — worksheets, FAQs, exam papers, quiz lists — as by a clean, direct answer from a helpful expert. The model gives you the most plausible continuation of the document you started. Your question was a document.
| Prompt | Base model — a plausible continuation (illustrative) | Assistant — a trained response (illustrative) |
|---|---|---|
“What is the capital of France?” | “What is the capital of Spain? What is the capital of Italy? What is the capital of Germany?” — the prompt reads like a worksheet, so the model extends the worksheet | “The capital of France is Paris.” |
“Write a haiku about the ocean.” | “Write a sonnet about the mountains. Write a limerick about a river.” — a list of writing prompts is a common document shape | A haiku about the ocean, followed by an offer to try another style |
“My code throws IndexError on line 3.” | “I’ve tried everything and nothing works. Please help — this is due tomorrow!! EDIT: still stuck.” — it continues your side of a forum post | An explanation of what raises IndexError and a request to see the code around line 3 |
So the base model is right by its own rules and wrong for the product. The two jobs are genuinely different. Pre-training builds a model of what text exists: a distribution over documents, learned from everything in the corpus. A product needs a helpful response from one consistent participant in a conversation: answer the question, stay in role, decline some requests, stop when done. Nothing in the pre-training objective ever asked for that second thing.
Here is the clue that the gap is narrower than it looks. You can rent assistant behavior from a base model with prompting alone. Frame the input as a transcript — The following is a conversation between a helpful expert and a user. User: … Expert: — and the model, predicting what such a document would contain, produces something answer-shaped. The trick is fragile: the model may happily generate the user’s next turn too, drift out of character, or imitate an expert’s style without an expert’s care. But it proves something important: the capability was already in there. The base model knows what good answers look like; it just assigns them no special status among all the continuations it could produce.
The distribution view of the gap
Pre-training fits one object: a distribution over text, whose conditional slices p(next Token | prefix) are everything the model can do. Ask a base model a question and you are sampling from p(continuation | a document that begins with this question) — averaged over every genre where such text appears. Worksheets, forums, exams, and expert answers all contribute mass, weighted by how often each genre produced that prefix in the corpus. The “more questions” behavior is just the mode of that mixture.
The transcript trick works because conditioning is steering: prefixing a dialogue frame moves you to a slice of the distribution where an expert’s answer is the likely continuation. Post-training does the same move permanently — it reshapes the conditional distribution so that dialogue-shaped context concentrates mass on helpful, in-role responses by default, no frame required.
Hold on to three predictions this framing makes, because the rest of the domain confirms them. First, post-training can be small: moving probability mass between continuations a model already represents is far cheaper than creating capability from scratch. Second, the base distribution never disappears: post-training lowers the probability of unwanted continuations rather than deleting them, which is one mechanism behind jailbreaks. Third, the methods will keep the reshaped model anchored near a reference — the KL penalty you will meet in the RLHF module exists precisely to keep the redistribution from wandering.
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.