The subfield of AI for handling decisions which are context dependent and which yield rewards (or minimize costs) is reinforcement learning. Even a shallow fami

The subfield of AI for handling decisions which are context dependent and which yield rewards (or minimize costs) is reinforcement learning. Even a shallow familiarity with this family of techniques raises a number of reasons why automating a CEO is deeply challenging: 1. Specifying the problem to be solved can be very subtle. The article throws in an example of scheduling maintenance in a transit system. That's a case where someone can specify some utility function around service disruptions, cost, etc. Companies are complex. 2. The space of actions available to a company at any point is extremely broad, and also hard to specify. 3a. An agent needs to learn from something. This can take a couple directions. Roughly, "off-policy learning" learns from the actions and outcomes of _another_ agent with different internal rules etc. Perhaps we have incomplete data of this sort (business school case studies etc leave much information out). But you need a _lot_ of data, and it needs to cover the same sorts of situations that your agent would get around to. And some actions probably the software agent _cannot_ copy (e.g. doing an interview on Sunday news program). 3b. "On-policy" learning is improving your agent from the past actions/outcomes of its own decision-making. Normally this means we need some simulation environment in which to let the agent learn. How do you simulate the whole economy, and market forces with which a company interacts? After learning in a simulated environment, you likely need to further improve (and test) in the real world. In this domain, I think that means you have to be prepared to let a computer program run many companies into the ground. 4. Without a decent model of the world around them, RL agents have a harder time attributing outcomes to specific past actions. You beat your earnings estimate, but which of the bajillions of decisions made over the past several quarters is most responsible for that? I think human CEOs actually do a questionable job of this, but an automated one would be worse.