What is a neural network?

Every headline capability of contemporary AI — from face recognition to ChatGPT — rests on a mathematical construction with a biological name: the neural network, loosely inspired by the brain: artificial “neurons” (small computing units) organized in layers, linked by connections with adjustable weights. The mechanism, reduced to its essence: during training the network is shown examples with known answers (this photo is a cat; this email is spam; after these words, this word probably follows), and each time it errs, the connection weights are nudged slightly in the direction that would have reduced the error — repeated millions or billions of times, until the network generalizes: it recognizes cats it has never seen, patterns nobody described to it explicitly. Nobody programs the rules — the network absorbs them from data; which is exactly why it works in domains where explicit rules always failed (images, language, sound) and exactly why its individual decisions are hard to explain (the knowledge is dissolved across millions of weights, not written in any readable list — the “black box” problem). Size and architecture separate the generations: “deep” networks with many layers unlocked computer vision, and the Transformer architecture, scaled to hundreds of billions of parameters and trained on the whole internet, produced today’s large language models. For a decision-maker, this minimum of mechanics is practical, not academic — from it flow the properties you meet in every AI project: why these systems need data (they are sculpted from examples), why they answer probabilistically and can err with confidence (they approximate the plausible rather than consult truths), why they inherit the biases of their training data (they learned the world as the data described it, crooked parts included — legally relevant when decisions concern people) and why you never need to build one from scratch: pre-trained models are used and specialized, so the competence worth buying is applying them with judgment, not their mathematics.