Problem is that it's impossible to know whether its own outputs are good or not. How could a model which is dumber know that the outputs of a smarter model are

Problem is that it's impossible to know whether its own outputs are good or not. How could a model which is dumber know that the outputs of a smarter model are better?