What AI Is Great At — and What It Isn't

A neural network is a mathematical function organized as layers of simpler functions called neurons. Despite the biological name, it's just arithmetic — but applied at massive scale. Each neuron takes a set of numbers as input, multiplies each one by a weight (the strength of that connection), adds them together, and applies an activation function that decides whether the neuron "fires" (produces a significant output).

Neural networks are organized in layers. The input layer receives raw data — pixel values in an image, word tokens in a sentence, transaction amounts in a fraud case. The output layer produces the result — "this is a cat," "sentiment is positive," "fraud probability 94%." In between are hidden layers where the real learning happens. Each hidden layer learns to detect increasingly abstract features. Early layers might detect edges and colors. Later layers detect shapes, then objects, then context.

Training is the process of finding the right weights. You start with random weights (the model knows nothing). You feed it an example and it makes a prediction. You compare the prediction to the known correct answer and calculate the error. Then you run an algorithm called backpropagation which calculates how much each weight contributed to the error and adjusts them slightly in the direction that reduces it. Repeat this millions of times across millions of examples, and the model's predictions become accurate.

You don't need to write or understand any of this math to use modern AI effectively. But understanding that weights are the "knowledge" stored in a model — that they encode patterns learned from data — helps you understand why models fail when tested on data that looks very different from what they were trained on.