One question, six levels down
You type "Why is the sky blue?" and press enter. A second later, words start appearing. Between those two moments, a surprising amount happens, and most explanations pick one level of it (servers, or neural networks, or maths) and skip the rest.
This one doesn't skip. The picture below zooms in by itself, each level sitting inside the middle of the one before it. You start at the chat box and end at the arithmetic underneath. Want to stop and think? Pause holds it right where it is. Missed something? Level up steps back one level and carries on; Restart plays it from the top.
The same trip, in words
1 · The chat window is a front door
Everything you can see (the text box, the send button, your list of past chats) is just the app. None of it is doing the thinking. Pressing send packages up your message and ships it off.
2 · A server gets the whole conversation
What arrives at the data centre isn't only your latest line. It's the entire conversation so far, every time, because the model keeps no memory between messages. The first thing that happens is tokenising: the text is cut into chunks called tokens. Common words become one token; rarer ones get split into pieces. A token is the unit the model reads and writes.
3 · The model is a stack of layers
The model itself is a tall pile of near-identical layers, often around a hundred in large models. Your tokens enter the bottom, pass through every layer in turn, and come out the top as a prediction of one next token, such as "Because". That token is appended to the text and the whole thing runs again for the token after it. A full answer is this loop repeated hundreds of times.
4 · Inside one layer: two halves
By the time a token reaches a layer, it's no longer a word. It's a list of thousands of numbers (a vector) that encodes what the token means in this context. Each layer has two parts:
- Attention lets each token pull in information from the others.
- Feed-forward transforms each token on its own, applying patterns the model picked up during training.
Every layer nudges the vectors a little, and meaning sharpens as they rise through the stack.
5 · Attention: words looking at words
Attention is how context gets in. For each token, the model scores how relevant every other token is, then blends in information in proportion to those scores. In "why is the sky blue", the token "blue" leans hardest on "sky", which is how the model knows this is a question about the sky's colour rather than a mood or a paint swatch. Those scores are the grid of shaded squares in level five.
6 · All the way down: numbers
At the bottom there is no understanding or reasoning written anywhere as such. There is multiplication and addition: a grid of learned numbers (the weights) times a list of numbers (the token's vector) gives a new list. The panel shows a two-by-two version; real ones are thousands wide. Billions of these multiply-adds run for every token the model produces, which is why this work runs on GPUs built to do exactly that, very fast, in parallel.
Why the zoom matters
Holding all six levels at once explains a lot of everyday AI behaviour:
- Long chats get slower and pricier because level 2 resends the whole conversation, and level 5 compares every token with every other.
- It "forgets" things you said long ago because there is a limit on how many tokens fit through that trip at once.
- It can sound sure and still be wrong because level 3 picks a likely next token, not a checked one.
- It struggles to count letters in a word because it never sees letters, only tokens.
None of that is magic, and none of it is obvious from the chat box. It only makes sense once you've zoomed through the layers.
Where to go next
For the part that sits around the model (the tools, memory and loop that let it do real work), read The AI Harness. To see the same ideas built up one step at a time, start the free AI Roadmap.