How AI works, from scratch

The plan

Two tracks, one machine

An AI chatbot is two things working together. The LLM (large language model) is the program: a huge pile of numbers that turns your message into an answer. The chips are the hardware that does the arithmetic those numbers need. The series follows both, and switches between them whenever one side needs the other to make sense.

Read along the line from left to right. A filled dot is published; a hollow one is coming.

How to read it

The same colours mean the same things

Every diagram in every part uses one colour key, so after a part or two you can read the drawings without a legend.

Facts carry a tag. public means a published paper, spec sheet or plain arithmetic. example means a made-up but consistent number used to show how something works. not public means the company has not said. Each part also has Go deeper boxes. They are optional: skip them on a first read and nothing later breaks.

The reading list

Every study, in order

Start at part 1. The early parts need no background; the later ones assume you have read the earlier ones.

Why switch tracks?

Each side explains the other

Training a model (part 6) is just an enormous amount of multiplying, which only makes sense once you know why a graphics card is good at multiplying (part 5). Why a model is slow to answer long questions (part 14) comes down to memory (part 8). So the order follows the questions you would naturally ask, not one subject at a time.

The bonus study, Inside LLM serving, is where the two tracks meet: one model on a group of chips answering millions of people at once. It is written to be read last, when every word in it is familiar.