See how AI actually works
Four hands-on demos of the ideas behind ChatGPT and other modern AI. Type, click, drag the sliders and watch what happens. Everything runs in your browser.
Tokenization: how AI reads text
Language models never see letters or whole words. They see tokens: common words stay whole, rare words get chopped into pieces. Type below and watch your text split up.
What to notice
- Common words like the and is are one token each.
- unbelievable becomes un + believ + able, so the model learns word parts.
- Rare names like Karachi split into many small pieces.
- A space usually belongs to the next token (shown as a faint dot).
Why it matters
AI pricing and context limits are counted in tokens. English averages about 4 characters per token; Urdu and other languages often need more tokens for the same meaning. Hover a token to see its ID number, which is what the model really receives.
This is a simplified tokenizer with a small vocabulary, built to show the idea. Real ones learn ~50,000–200,000 pieces from data.
Word embeddings: meaning as location
AI turns each word into a list of numbers, a point in space. Words with similar meanings end up close together. Real models use hundreds of dimensions; here we squash it to 2 so you can see it. Click any two words to compare them.
Word arithmetic
Directions in the space carry meaning. The step from man to woman is the same as from king to queen.
Try these
- king − man + woman
- Karachi − Pakistan + England
- puppy − dog + cat
Where it's used
Search engines, recommendation systems and RAG chatbots all use embeddings to find text with similar meaning, not just matching keywords.
Neural network: learning a boundary
A neural network learns by adjusting its weights a tiny bit at a time to reduce its error. Pick a dataset, press Train, and watch it learn to separate the teal points from the purple ones.
Experiments
- XOR with 1 neuron can't succeed: one neuron draws only a straight line.
- Raise neurons to 4+ and it bends the boundary to solve XOR.
- Set the learning rate very high and watch the loss jump around.
- The spiral needs many neurons and patience. Deep learning stacks many layers for problems like this.
Temperature & sampling: how AI picks the next word
A chatbot writes one token at a time. For each step it gives every possible next word a probability, then picks one. Temperature controls how adventurous that pick is.
What to notice
- Low temperature (0.1): almost always the top word. Safe and predictable, good for facts and code.
- High temperature (2.0): odd words like "banana" get a real chance. More creative, but riskier.
- Top-k = 1 is "greedy" decoding: no randomness at all.
This is also why the same question can get different answers from a chatbot: the model is sampling, not looking up a fixed reply.