What if Instead of States, We Would Predict Actions?

Mikeo: Machine learning is about predicting what to do, and it's giving a machine also the autonomy to choose something you might not agree with in order to learn more about the world. He says that one of the biggest mistakes we make in machine learning so far is that we end up with a prediction who might be churning - when we could just as well predict what we should do about it. And this brings us to the topicof reinforcement learning. What if instead of trying to predict states, we would predict actions, and we would actually learn from that as an agent? he asks.