AI Spaceship

AI Concepts

0 XP0 days
Learnhowmachinesread.
5 tokens: Learn / how / machines / read / .

That sentence, as a model receives it. Short lessons built on real tools — and where something is a simplification, the lesson says so.

Every lesson is self-contained, so you can start anywhere. For how the lessons are built and where the models are deliberate simplifications, see the AI Concepts documentation.

01

Tokenisation

How a model actually reads your text
0/4
  1. 1.1What a token isThe unit a model actually readsStart
  2. 1.2Why words splitWhat decides the cost of a wordStart
  3. 1.3Case, digits and spacesThe details that quietly change the countStart
  4. 1.4What tokens costBytes, languages and doing the sumsStart
02

Embeddings

Turning tokens into meaning
0/3
  1. 2.1From IDs to meaningWhy a number isn’t enoughStart
  2. 2.2Measuring similarityAngle, not distanceStart
  3. 2.3Direction and contextAnalogies, and the catchStart
03

Attention

How tokens read each other
0/3
  1. 3.1Why words need neighboursA word on its own is ambiguousStart
  2. 3.2What “it” points atOne adjective changes the answerStart
  3. 3.3The cost of lookingWhy long inputs get expensive fastStart
04

Output logits

How the next token gets picked
0/3
  1. 4.1What comes outA score for every token there isStart
  2. 4.2Temperature and samplingWhy the same prompt variesStart
  3. 4.3How good is the guess?Loss and perplexityStart
05

Context windows

Why models forget
0/3
  1. 5.1What the window holdsEverything, all at once, every timeStart
  2. 5.2When it overflowsWhat falls out, and what gets missedStart
  3. 5.3Why it’s finiteThe cost behind the limitStart
Token counts come from the real o200k_base vocabulary — the one GPT-4o reads with. The meaning map in chapter 2 is a hand-built stand-in, and says so.