We've already crossed that threshold:

https://www.wired.com/2017/05/googles-alphago-levels-board-g...

But then, about a year ago, DeepMind redesigned the system. In essence, they built the new AlphaGo without help from human moves. They trained it entirely from games where the machine plays against itself—part of a continuing progression toward AI techniques that truly learn on their own. “AlphaGo has become its own teacher,” says David Silver, the project’s lead researcher.

Answer: The machines just train themselves.

Yes, but it doesn't take as many people to improve those models as it does to have a thinking person at every single end point of a system. If it did, programming wouldn't be profitable. And it's obviously highly profitable to automate things.

And that's a good thing. Less work for humans is a great thing, but we have to have new economic models to keep up with it.

It might be more accurate to say not that the ML models are trained off [decisions of] people but that ML models are trained off past experience, the results of decisions that were taken.

When introducing a new ML system, it's generally trained off of people, but in most (though not all) contexts you get much, much more data once the system starts getting used in practice and you get feedback on the decisions that the machine made.

The human knowledge is generally needed as a seed to make the model sufficiently not-bad so that we'd allow it to be widely used in practice and learn on its own.

>How do we make new better models?

Who knows? Just because you can't imagine the step after the one I describe doesn't make the one I describe less valid. Maybe it'll become self-improving? We already need AIs to help build CPUs so doesn't seem far-fetched.