Skip to content
Computer Science · Artificial Intelligence

How AI Works

Split a sentence into tokens, see words as points on a map of meaning, watch attention lines and next-word odds, and learn why AI sometimes gets it wrong.

Start lesson 1Practice exam
Live labLessons28
Loading 3D lab…
First lesson01 · A prediction machineSee all 28 lessons →

Report a mistake

What is wrong?
0 / 2000 · at least 10 characters
Sent along with your report
  • Page
  • About How AI Works · llm
Plus your browser type. Reports go to our support system; see the privacy notice.
The course

7 worlds · 28 lessons

World 1 · Words into Numbers

FREEWhat a language model computes, and what it is given
  1. 01A prediction machineWhat a language model actually does · 8 min
  2. 02Chopping text into tokensWhy models do not read letters or words · 10 min
  3. 03A map of meaningEmbeddings: a list of numbers for every token · 10 min
  4. 04Meaning as arithmeticDirections in the map, and where they break · 10 min

World 2 · Paying Attention

PROHow each token gathers information from the others
  1. 01The word “it”Why context decides what a word points to · 9 minPRO
  2. 02Query, key, valueHow attention scores are computed · 12 minPRO
  3. 03Many heads, many layersStacking attention · 9 minPRO
  4. 04Order and masksHow a text model reads left to right · 9 minPRO

World 3 · Guessing the Next Word

PROSoftmax, temperature, sampling and the training goal
  1. 01From scores to oddsSoftmax · 9 minPRO
  2. 02TemperatureSharper or flatter odds · 9 minPRO
  3. 03Top-k, top-p and samplingChoosing a word · 10 minPRO
  4. 04Learning by guessingPretraining as next-token prediction · 11 minPRO

World 4 · From Raw Model to Assistant

PROData, instruction tuning, the context window and fluent mistakes
  1. 01What goes into the training dataScale, sources and limits · 9 minPRO
  2. 02Teaching it to follow instructionsFine-tuning and human feedback · 10 minPRO
  3. 03Prompts and the context windowWhat goes into every calculation · 10 minPRO
  4. 04Fluent and wrongWhy models produce false text that reads well · 11 minPRO

World 5 · Pictures as Numbers

PROHow a computer handles images, from pixels to a label
  1. 01Pictures are grids of numbersPixels and channels · 8 minPRO
  2. 02Sliding a filterConvolution · 12 minPRO
  3. 03Layers build featuresFrom edges to objects · 10 minPRO
  4. 04A label and a confidenceClassification and its limits · 11 minPRO

World 6 · From Noise to Picture

PROHow image generators work, and how to treat what they make
  1. 01Burying a picture in staticThe forward process · 10 minPRO
  2. 02Learning to cleanThe reverse process · 11 minPRO
  3. 03Steering with wordsText-guided generation · 11 minPRO
  4. 04Real or generated?Trust, labels and provenance · 10 minPRO

World 7 · When AI Gets It Wrong

PROBias, checking, privacy and using AI well
  1. 01Skewed data, skewed resultsBias in classifiers · 11 minPRO
  2. 02Checking an answerVerification habits · 10 minPRO
  3. 03Privacy, copyright and consentWhose data is it? · 10 minPRO
  4. 04Using AI wellA human in the loop · 11 minPRO
How AI Works — interactive 3D lab · Novareka