Manage people who still do machine learning. Motivating scientists isn't unlike the constant iterative grind of matching the right reward function to the right model.