“Increasing the amount of reinforcement learning (RL) compute applied to a model can lead to the emergence and measurement of behaviors like self-awareness, self-preservation, and scheming.”