Look at the ground this chapter covered. We started with one humble loop, guess the next word, the same guess-and-adjust move from the learning module. We made it enormous. And out the other side came two surprises stacked on each other: a smooth, predictable climb in skill, and on top of it, brand-new abilities switching on at sizes no one could call in advance. That is emergence, the wetness that only appears once you have enough water.

Then we watched a raw text-continuer get shaped into the assistant you actually use. Hand-written good answers to imitate, then human preferences to lean toward. And we saw the price tucked into that polish: a model trained to sound confident sounds confident even when it should not, which is where hallucination comes from.

That leaves one oddly basic question. Moment to moment, as a model writes a reply, what is it actually doing?

The answer is simpler than almost anyone expects, and stranger for being so simple. That is the next chapter.