“The methodology behind DeepSeek's R1 model is the canonical recipe for training a reasoning-only model.”