How gradient descent finds a good model

So far we have said a model learns by adjusting its numbers. But how does it know which way to adjust them? For a huge share of modern machine learning, including neural networks, the answer is an idea called gradient descent. Once you understand it, much of the field becomes less mysterious.