Conversation

AI doesn't have to be self-improving to get smarter than humans, most of the (narrow) AI's that have ever existed, from calculator to AlphaGo, very quickly became superhuman (calculator is better at math than you are).

The real question is how to build a general(as opposed to narrow) intelligence, flexible algorithm that can solve a wide variety of problems. The fact that it will be better at that than us is a given.