“DeepSeek trained its DeepSeekMath model using a natural language verifier, which was itself trained by a meta-verifier to ensure it gave good feedback on problems from "The Art of Problem Solving" website.”