“Recent improvements in AI model capabilities are primarily coming from advances in reasoning and reinforcement learning, not pre-training.”