Hallucination

The model produces plausible language. Plausible is not the same as true. When the plausible continuation of a question is a confident, specific, wrong answer, the model produces it with the same fluency it uses for correct ones. Nothing inside the prediction loop checks the claim against the world.

Hallucination is therefore not a bug that a future version will simply remove. It is the visible edge of a tension between generating plausible text and guaranteeing truth. You reduce it by giving the model access to facts, by asking for output that can be checked, and by deciding, for each use, what happens when it is wrong.