The guide
Models output guesses, and a guess without a sense of how sure it is cannot be acted on. Probability is the language for that, and most of deep learning is written in it whether or not it says so.
Cross entropy, the loss almost every classifier uses, is a statement about distributions: how many extra bits it costs to encode the truth using your predictions. KL divergence is the same quantity with the irreducible part subtracted. Entropy is the floor. Getting these three straight makes a lot of later reading stop being mysterious.
Bayes' theorem earns its own problem because the intuition is genuinely bad. People reliably get the disease-test question wrong, including people who know the formula.