A trained neural network is not a program in the traditional sense. Nobody wrote down what it knows. Nobody specified the rules. What remains after training is a large file of numbers — weights — that together encode whatever pattern the network extracted from its training data.

That's genuinely strange. The knowledge is real: the network can recognize faces, classify images, understand language. But the knowledge doesn't exist anywhere you can point to or read. It's distributed across billions of numbers, none of which mean anything alone.

The same properties that make this powerful — the ability to generalize, to find patterns nobody anticipated — are what make it hard to inspect and easy to fool. There's no rule list to audit. When the network is wrong, you often don't know why.

The next module is about what finally made it possible to go deeper: more layers, more data, and the hardware that made it all practical. That's where things start moving fast.