Machine learning is reverse engineering formulas from evidence.

When an outcome depends on many variables and the mechanism is unknown — fraud, house prices, disease risk — building the model from first principles becomes practically impossible for humans. There is no physics to derive it from, and no way to hand-pick the right weight for every variable.

But we have something else: historical records of inputs and their actual outcomes. So we flip the direction. Instead of deriving a formula and checking it against reality, we start from reality and search backwards for the formula that would have produced it. An algorithm like gradient descent adjusts the weights step by step, keeping the model's predictions as close as possible to the recorded outcomes — minimizing the error.

The goal is not to memorize the past. It is to generalize beyond it: a good model predicts cases it has never seen.