Part 1 · What is AI
1.1
Following rules vs. figuring it out
Take on a machine that guesses animals. You'll probably win.
H6 Beat the machine
1.2
How does AI learn?
Nobody writes down the answer for it. So how does it learn? Turn one weight and you'll see.
P0 Setting weights
1.3
Seventy years to get here
AI didn't show up out of nowhere. It's a seventy-year story with two big disappointments in it.
H7 Predictions that missed
Part 2 · AI that sees and hears
2.1
A photo is numbers
Zoom into a photo far enough and what do you get? Squares — and every square has a number on it.
P1 Pixel magnifier
2.2
How a computer finds features
Lay two pictures on top of each other and a small tilt ruins it. So a computer looks at the pieces instead.
H1 Matching by overlapP2 Filter playgroundP3 Looking layer by layer
2.3
AI that recognises people
Cameras can draw a box around your face because of one method published in 2001.
H11 The 2001 face detectorP4 Webcam recognition
2.4
AI that sees sound
A computer doesn't listen to sound. It turns sound into a picture and looks at it.
H12 The machine that only knew digitsP5 Make a spectrogram
Part 3 · AI that learns
3.1
Teach it yourself
You're about to teach a machine without writing a single rule. And when you're done, one strange thing is left over.
P6 Your own classifier
3.2
Inside a neural network
Time to open it up and see where the weights are attached. And to meet the problem that stopped the world for twenty years.
P7 Neuron playground
3.3
When AI learns the wrong thing
The machine got it wrong. But maybe the machine isn't the problem — maybe what we gave it was.
P8 Lopsided dataP9 When you add weights
3.4
Carrot and stick
How do you teach something nobody knows the answer to? Reward what works, penalise what doesn't. That's enough.
H4 The matchbox machineP11 Driving range
Part 4 · AI that talks
4.1
Cutting text into pieces
An AI doesn't read text as words. It cuts it into pieces first — and how it cuts depends on the language.
P12 Cutting text into pieces
4.2
The next-word game
Today's AI does exactly one thing — guess the next piece. Make it very good at that and the rest follows.
H5 Guessing the next word by countingP13 Guessing the next word
4.3
Choosing where to look
A long sentence has dozens of words behind it. A model doesn't weigh them all the same. It picks out the ones it needs.
P14 Choosing where to look
4.4
Confidently wrong
The old chatbot gave itself away. Today's chatbot doesn't. Which of those is more dangerous?
H2 The counsellorP15 A chatbot like today's
4.5
Asking well
Same AI, very different answers, depending on how you ask. Knowing why tells you how to ask.
P16 Prompt labP17 Getting an AI to write code
Part 5 · AI that makes things
5.1
From mess to picture
Picture-making AI doesn't use a brush. It looks at a mess, guesses what's in there, and takes away whatever doesn't fit.
P18 From mess to pictureP19 Ordering a picture
5.2
AI that composes
Does music-making AI use some new method? No. It's the one you already saw in part 4.
P20 Melody maker
5.3
Convincing fakes
Fake photographs are not new. What changed isn't that fakes exist — it's that making one now costs nothing.
H9 How long a fake tookP21 Learned from, or made?
Part 6 · Living with AI
6.1
What AI can't do
You've seen plenty of what AI does. Time to look at the other side.
H10 Person, or machine?P22 Who's responsible for this?
6.2
The AI that knows you
The AI you spend most time with isn't a chatbot. It's the one choosing what to show you.
P23 My feed
6.3
Growing up alongside AI
This won't tell you which jobs have a future. There's a different question it can answer.
6.4
Your own rules for AI
You've come a long way. One thing left — how will you use it?
P22 Who's responsible for this?