It was never asked to answer you

Look back at the training loop and the behaviour stops being strange.

Every pass through it asked one question. Given this text, what comes next? Every correction pushed the weights toward the word that actually followed. Billions of times.

Nothing in that loop ever asked the model to be useful. Nothing asked it to answer a question, tell the truth, or decline. Those ideas were not in the task. A model that answered your question helpfully and a model that continued it with three more questions were both just producing likely text, and on the training data, the second is often likelier.

So the base model is not a broken assistant. It is not an assistant at all. It is exactly what the loop selected for, and the loop selected for plausible continuation.

Which leaves a real problem. You have a machine that has absorbed an enormous amount about language and the world, and no way to ask it for any of that directly.

The fix is not a better base model. It is a second, much smaller round of training laid on top of the first, with a different question at its centre. Not what usually comes next but what should this thing say.

That second round is called post-training, and everything you think of as talking to an AI comes from it.