Learn how AI models work
A learning path in plain language: start with the basics, work up to the advanced machinery, then step into the model lab and play with real (tiny) models in your browser. Read in order or jump to the one you came for.
Play with a real model now
The model lab holds tiny models that run in your browser. You can train one yourself and watch it work, starting with the three below.
The lab course
The ten labs in the order they build on each other. Weights come first because everything else needs them; after that, vision, training, and language run in parallel. Follow an arrow or jump anywhere you like.
Model vs. model
Direct comparisons of the big-name models now live in the guides, along with the best-for-your-task picks.
The model lab
Tiny models small enough to see through, running in your browser. Meet Doodle-15, Doodle-64, Doodle-64R, Doodle-525, Doodle-918, Parrot-43, Parrot-2D, and Finch-4, then scale the same ideas up to billion-parameter LLMs.
How do neural network weights work?
The tiny dials that make neural networks learn anything.
How do neural networks classify images?
Meet Doodle-64, a 64-parameter classifier that tells 3 from E and works exactly like the big models.
What happens if you set the weights yourself?
Meet Doodle-15. This model is small enough that you can set all fifteen of its weights by hand, and then watch training find its own.
How do AI models generate images?
Meet Doodle-64R. You can ask it to draw a 3 or an E, and it builds the picture by running the classifier backward.
How do neural networks recognize digits?
Meet Doodle-525: draw any digit and watch its hidden layer find the strokes before it names the number.
What is gradient descent?
Choose the starting noise, train Doodle-64 in your browser, then test the weights it learned.
How do you train a neural network?
You are the teacher: label the drawings, then watch the model turn your answers into weights.
How do LLMs predict the next word?
Meet Parrot-43, a language model trained on nine sentences. Watch it predict the next word and see exactly where every prediction comes from.
How do word embeddings predict the next word?
Meet Parrot-2D. This is a model that gives every word two numbers you can read, one for how friendly it is and one for whether it is a person or a verb. You can watch it predict the next word right on a map of those two meanings.
How do position and attention make language models grammatical?
Meet Finch-4. This is a model that adds a position signal and a small attention head to Parrot-2D, so it can build a subject, verb, object sentence.
Why do neural networks need more than one hidden layer?
Meet Doodle-918. This is a model that can see loops and curves, and can interpret these into digits. You can draw a digit and watch how the different shapes become an answer.