Free browser game · no signup · plays in 2 minutes
Grad School
Neural networks, by hand.
Getting good at this game and understanding the mathematics are the same activity. There is no way to win by pattern-matching, because the winning move is the algorithm.
▶ Play nowNothing to install. Nothing to sign up for. Works on your phone.
What you actually do
Route the gradient yourself
Tap an edge, type what the gradient gets multiplied by. Not multiple choice — you produce the number, which is the difference between recognising a rule and knowing one.
Then watch it learn
Use the gradients you computed to move real weights, and watch the loss fall in front of you: 1.96 → 0.0035. Nothing is simulated. It is the arithmetic every training run performs.
Get told why, never just what
Every rule is derived from something ordinary — meshed gears, a square’s area, repeated addition — and links to the exact second of the lecture where it is proved.
The payoff
You compute it. The machine learns.
Thirty-six levels teach you to find a gradient. This is what a gradient is for: you drive five steps of training by hand and the neuron gets it right.
Every day
One new graph, the same for everyone.
A fresh board every day at midnight UTC. No timer, no single attempt, and missing a day costs you nothing but the streak. There is no fail state anywhere in this game.
Five courses, unlocked in order
Built on Andrej Karpathy’s Neural Networks: Zero to Hero. Each one finishes with a cup.
Send a gradient back through an expression graph, one edge at a time.
15 levelsUse your own gradients to train a neuron and watch the loss fall.
5 stepsMatch a query to a key and watch a pronoun find its noun, under a causal mask.
8 levelsThe keys are fixed. You build the query that makes the right token win.
4 levelsA network arrives sick. Read the vitals, prescribe the fix, watch it recover.
8 casesBackpropagation takes about two minutes to start.
▶ Play Grad SchoolFree. No signup. Sign in later only if you want your score to follow you.