“Misha Laskin, who led reward model development for Gemini, believes the primary bottleneck in scaling reinforcement learning is the "reward problem"—the difficulty of creating accurate reward models for arbitrary tasks.”