Tiny AI Arena makes model battles watchable, not a benchmark

Tiny AI Arena is a spectator site for turn-based matches between four AI models. Each chooses to move, attack or wait on an 8x8 grid; the site lets viewers step through a finished game frame by frame, inspect the log and see an Elo-style leaderboard. The maker says the public site serves exported, static replays, not a live game you can create there. The source is on GitHub; running new games locally needs your own OpenRouter key.

A replay makes AI decision-making easier to inspect than a one-number ranking. You can see a model wait, move toward gold, attack and survive, then look back at the exact turns. The Show HN thread drew active questions about what the results mean and bug reports about mobile scrolling. If you are deciding which model to use for real work, this is a useful prompt to ask about the rules and repeatability before believing a leaderboard.

Fun game, weak evidence of general intelligence. Four-player combat has random turn order, random obstacles and small match counts; the maker told Hacker News that a legitimate benchmark would need many more games and an estimate of how much the result is random. An Elo score here summarizes these matches under these rules, not writing, coding or other tasks. We watched a public replay and read the repository, but we have not run a new match, audited its model labels or tested reliability. Public replays are static; local games require an OpenRouter key and could incur model charges. The maker credits third-party pixel art and sound. Do not treat footage or game art as freely licensed for reuse without permission.