I watched an interview with AI safety researcher Roman Yampolskiy, and it raised a practical question for people who use ChatGPT for advanced workflows. His broader claim is that increasingly intelligent AI systems may become harder to predict and control. Whether or not you agree with his conclusions about AGI, a smaller version of this problem already exists when we give an AI access to tools. There is a major difference between asking ChatGPT to draft an email and allowing an agent to sen...