An AI that taught itself a puzzle.
A small neural network learned the 15-puzzle from nothing: no human games, no rules of thumb, only trial and error. It now runs entirely in your browser. Scramble the board and watch it think.
Tap a tile next to the gap to play yourself. The AI's estimate updates after every move.
How it learned, without a single example.
Nobody showed it a solved game. It started with random guesses about how far each board is from solved, then corrected itself over and over: a board is one move further away than its best neighbour. Thousands of rounds later, those guesses became good judgement.
Scramble backwards
Start from the solved board and slide tiles at random. The further back it walks, the harder the boards it practises on.
Look one move ahead
For each practice board, it checks every possible next move and asks its older self how far each one is from solved.
Correct itself
Best neighbour plus one becomes the new answer. It nudges its numbers toward that and repeats, slowly widening the scrambles.
Think with it
In your browser, a search uses that judgement to decide which boards are worth exploring, and keeps going until it finds the way home.
Honest notes: solutions are short but not proven to be the shortest possible. Training ran once on one laptop processor, with no graphics card and no cloud. Nothing you do here leaves your device, and the voice uses your device's own speech.