As stated in the article, these systems (we should not call them "AI") seem to be useful in limited contexts, like AlphaFold in predicting protein structure (though there are debates/limitations to this too). In other words, if the domain is specific enough, *we can get useful results with them.
