“A new "reasoning paradigm" for AI training, focused on math and coding with reinforcement learning, has emerged in the past year and shows a faster, non-decelerating rate of improvement.”