“Amjad Masad believes the key technical breakthrough enabling long-horizon reasoning in foundation models was reinforcement learning, specifically RL from code execution.”