Every row of your dataset is a point in a geometric space, and every machine learning model is a line, curve or surface fitted through those points.
Why you shouldn't encode categories as integers, and how one-hot encoding plus softmax turns a multiclass problem into a set of binary classifiers.
Learning means tuning parameters like guitar strings: the error surface, gradient descent, and how it maps back to the model in input space.
A regressor outputs a number, a classifier outputs a category: the geometric difference, the sigmoid, and the decision boundary.
Features and target vs features only: when to use supervised learning, when clustering helps, and where reinforcement learning fits in between.
The simplest machine learning algorithm: look at who is nearby in the input space. How KNN works, its pros, and why it does not scale.