Inside an AI's head
Neural networks as a relay race, LLMs as supercharged autocomplete — and why YOU are the fact-checker.
17 min 8 steps 13 activities Up to 265 XP
- Describe a neural network as a relay race, and say what weights are
- Explain an LLM as supercharged autocomplete — one predicted token at a time
- Name what AI gets wrong and use the verify habit to guard against it
Inside a neural network — a relay race
Time to look inside the machine. In this lesson you'll see how a neural network passes an answer along like a relay race, how a chat AI writes one word at a time — and why it can sound completely confident and still be wrong. You'll leave with the one habit that protects you.
A neural network is the engine behind most modern AI. The honest picture, with no maths: think of a relay race where each runner doesn't just repeat what they were handed — they refine it a little before passing it on.
Feed in a photo of a handwritten "6." The first "runner" (layer) notices only simple things — edges and lines. The next combines those into curves. The next combines curves into shapes. The last runner recognises the whole shape and calls out: "That's a 6."
Recognising a handwritten "6"
Put the relay-race layers in the order they work, from first runner to last:
Each runner also has a habit of trusting some clues more than others — those trust-levels are called weights, and training (the bicycle idea from the last lesson) is really just the process of adjusting millions of them until the whole relay team gets good answers. Stack many of these layers and you get deep learning — the technique behind almost every AI breakthrough since 2012.