“Recent improvements in AI model capabilities are primarily driven by a new scaling curve based on reasoning and reinforcement learning, rather than pre-training.”