FrontierMath, a new benchmark from Epoch AI, challenges advanced AI systems with complex math problems, revealing how far AI still has to go before achieving true human-level reasoning.