“Andrej Karpathy asserts that LLMs from frontier labs are trained in large-scale reinforcement learning environments using verification-based rewards.”