He spoke about a project from the 2000's looking to apply an artificial neural net to a medical diagnosis problem. What they did was spend many years of work with domain experts to identify the best features for the net to use. The professor then proceeded to explain that nowadays feature engineering approaches became so much more powerful, that significantly better results could be achieved in a more recent project much faster and without almost any domain expertise.
Except, you know, for the expertise that generated the data in the first place.
It can be argued, since we don't know the ultimate cause of the laws of our universe, that we also rely in abductive reasoning and guide ourselves through models created by correlations based in our experience. But that has happened through millions of years of evolution and we still don't know how to build AI supervisors that can challenge that.