BLOCKWAVE · live, in your browser

An agent that never sees the score

Watch the pretrained agent play, see what the pixel agent sees, or play it yourself.

Downloads about 12.5 MB the first time: 8.1 MB of Python 3.13 and NumPy (Pyodide 0.29.5, from jsDelivr, cached by your browser afterwards), 4 MB of policy weights and 170 KB of the repo's Python. The pixel agent's 6.8 MB loads only if you open its tab.

decisions: 0
pieces: 0
lines: 0
games done: 0

Its policy right now

Empowerment reward per placement

What it was trained to maximize: how many distinct boards its next two pieces can still reach, minus the same count on an empty board. Zero while the board has room, negative as options run out. A top-out is charged about −607 (marked red): zero control for every placement the dead game would have had.

It acts on two 24×10 grids (locked cells, active piece) and the next two pieces. Score, lines and level are in the picture for you; none of them reach the agent.

Lines per piece on held-out seeds, from the README

never hard-drops ("drift")0.0018
pixel agent, 50M steps (pretrained/)0.0941
board-state agent, 150M steps (pretrained/)0.1629
scripted heuristic0.393
board-state agent, this page

The four grey rows are the README's results table (complete games, seeds never used in training; 0.4 is the ceiling). They were not measured here. Each run is one seed, and a second seed of a pixel configuration scored 42% differently, so small gaps are not results. The purple row, once a game finishes, is the agent on this page.