🤖 AI Lab
Five machines you can take apart. Every one of these runs right here in the page — nothing is sent anywhere, and there is no chatbot hiding behind them. What you see is the whole machine.
Machine learning
Teach the Machine
Most AI is not given rules to follow. You show it examples instead, and it works out the pattern by itself. Here you are the teacher — and the machine will only ever be as good as what you show it.
This creature
What the machine learned
- a Zib you taught it
- a Zog you taught it
- the creature on the left
The shaded background is its answer for every creature that could exist.
Language models
The Next-Word Machine
A chatbot does not know what it is going to say. It picks one word, looks at what it just wrote, and picks the next. This machine works exactly the same way — just with a much smaller memory.
What could come next
Tokens
How a Machine Reads
An AI never sees letters the way you do. Your sentence gets chopped into chunks called tokens, and each chunk becomes a number. Type something and watch it get chopped.
Inside the box
See One Neuron
A giant AI is millions — sometimes billions — of these wired together. One on its own does something simple: it draws a line to separate two groups. Drag the sliders and try to split the dots yourself, then let it learn on its own.
Sweet or sour?
The neuron's three numbers
It gets right
AI that gets it wrong
Bias Detective
An AI can be completely wrong while being completely confident. It is usually not broken — it just learned from examples that never showed it enough of the world. This time you choose every example it gets to see.
Missions
Step 1 · Teach it
Tap a card to change it: not used → this one is a match → this one is not. The machine only ever sees what you put in.
Pick at least one of each above and the machine will get to work.
Step 2 · What it worked out
These are the averages it built from your examples. The longest bar is the thing it is really going on — its shortcut.
Step 3 · Test it on things it has never seen
Six new items. It has to answer Is it an apple? for each one on its own.
The percentages are a rough measure of how far each answer sat from the alternative, not a real probability. Being sure and being right are different things — which is rather the point of this one.