Let's play 2048

Every three seconds the model is sent the board, the score and the moves that are legal right now, and is asked which move is best in this game of 2048. It answers with a probability for each of those moves, and the one it rates highest is played.

That is all it gets. It is not told the rules, how to play well or what any move would lead to, and it sees no earlier moves and remembers nothing, so every move is a fresh decision. These are decision models: rather than writing text, they return a probability for each option they are offered.

When only one move is legal it is played without asking. After every move a 2 (nine times in ten) or a 4 appears in a random empty cell, and the game ends when no move changes the board.

The request behind the latest live move