Whenever anybody says "we don't know how to make software that learns", my bullshit detectors go off overtime. Various forms of machine learning (including unsupervised) with varying degrees of effectiveness have been around for a very long time, with the first successes dating back to shortly after the introduction of electronic computers. And, we should never forget that we dedicate nearly twenty years of careful, difficult, and expensive labor (with heavy government subsidy no less) to turning human beings from useless blobs of flesh that can't even walk into something approaching functional adults capable of making good decisions, with a relatively poor success rate.
What we can't do is just let an AI build itself in terms of knowledge. Sure, we direct our kids when we put them into K-12 and beyond schooling, but the first 2 to 3 years they're with their parents watching, interacting, messing up, and so on. Kids just learn on their own. Some of that is clearly instinct, but at some point how do we inject such an instinct into a neural net? What would that look like to an AI? I think that's the problem here.