“Simple interventions, such as creating specific supervised learning data, reinforcement learning prompts, and synthetic reward modeling data, are very effective at steering models towards more aligned behavior.”