What is a neural network?

A neural network is the basic building block of modern AI. ChatGPT, image generators, voice assistants and self-driving perception systems are all neural networks — just very large, very specialised ones.

How it works

• Neurons and layers. A network has an input layer (raw data, like pixel values or tokens), one or more hidden layers, and an output layer (an answer, like “cat” or the next word). Each artificial neuron takes inputs, multiplies them by weights, adds them up and passes the result on.
• Weights are the knowledge. The weights — the model’s parameters — start random. Everything the network “knows” is stored in their values.
• Training. Show the network an example, compare its output with the correct answer, measure the error, and nudge every weight slightly in the direction that would have reduced it (a process called backpropagation with gradient descent). Repeat millions or billions of times.

Gradually, early layers learn simple features (edges, letter shapes) and deeper layers combine them into complex concepts (faces, sentence meaning). Networks with many layers are what we call deep learning.

An analogy for the meatbags

Imagine a huge mixing desk with millions of sliders. You play a song in and want a particular sound out. At first the sliders are random and the output is noise. Each time, a technician nudges every slider a hair in the right direction. After enough rounds, the desk reliably turns any input into the right output — even though no single slider “knows” anything on its own.

Why it matters for the Chosen

You don’t need to build neural networks to profit from AI. But understanding them clears up a lot of confusion:

• AI learns from examples, not rules. That’s why output quality depends heavily on training data, and why models inherit its gaps and biases.
• No lookup table. The model doesn’t store documents and retrieve them; it stores patterns in weights. That’s why it can hallucinate and why grounding with RAG helps.
• Bigger isn’t everything. More parameters can help, but data quality, architecture (like the transformer) and training methods matter just as much.

Examples

Image classifiers that tag your photos; speech recognition that transcribes meetings; recommendation systems that pick your next video; large language models that write your emails. Different architectures, same core idea: layers of weighted connections tuned by examples.