Probability: "With zero being a complete guess". Is this right? I think zero probability is certainty of non-occurrence.

Thanks for pointing this out! This gets into probability theory and statistics, which are deep topics beyond the scope of an overview course. My goal was to highlight the probabilistic nature of ML models while still prioritizing the compact nature of the lessons, and sometimes that's challenging to do without generalizing or betraying the proper semantics of some explanations.

You are correct that a probability of zero indicates the impossibility of an event occurring, or non-occurrence. Using the example of a single die again, the sample space of outcomes can be expressed as {1,2,3,4,5,6} and so the probability of rolling a zero is, also, zero, or non-occurrence.

Most performance analysis of ML models stems from the fact that training sets are finite and therefore any new example might not perfectly match an existing example. It's common to impose probabilistic bounds on a model. For more on this, see this probability calibration page from the scikit-learn documentation.

In the case of a classifier, probability is used as a measure of confidence in the prediction. Put another way, we're not necessarily asking about accuracy, but rather we're asking something to the effect of, "What is the probability that this answer is accurate?"

Hopefully that makes sense. Again, I genuinely appreciate this type of feedback. It's hard to break down complicated concepts in an overview course, and sometimes I don't always get it right or make explanations as clear as they need to be. I am by no means an expert in probability and statistics - I'm a programmer and creator - so I welcome any additional questions or corrections you might have. :)