text completion engine

One of the clearest ways to picture an LLM is as a very powerful sentence finisher. Give it the beginning of anything — a story, an email, a list, a line of code — and its core job is to produce what plausibly comes next. Chat, translation, and summarization are all this same engine in disguise: the conversation or instruction is just framed as text waiting to be completed.

Seeing it this way demystifies a lot. The model is not consulting a hidden answer; it is asking, given this text, what continuation is most likely, and producing that. It explains why phrasing matters, why the model sometimes drifts off topic, and why it can continue a falsehood as smoothly as a truth — a completion engine optimizes for plausible, not for correct.