Research.

Large language models write with total confidence, whether they are right or inventing facts from thin air. In 14 short, visual pages, this explorable builds up an intuition for what hallucinations are, why they are a structural consequence of how these models work, and how researchers, engineers, and everyday users can navigate them.

A language model can get dramatically better answers by thinking longer at answer time — not only by being retrained or made larger. This explorable walks through how inference-time scaling works and traces it back to the training-time loop that makes it possible.