What the 25 rungs represent

Twenty-five rungs make up the training ladder, from rung 1 to rung 25. Each has a specific Elo target, spaced a hundred points apart: 100 for the first, 2500 for the last — the function that calculates this target in the game code simply multiplies the rung number by a hundred. These are not measured ratings: they are goals set in advance to give newcomers to ultimate tic-tac-toe a clear ladder to follow.

The page itself says so plainly: beside the currently selected bot, the ‘Provisional Elo’ label stands in for a measured rating until calibration games have been played. That is a deliberate choice, not a box ticked lightly. Twenty-five bots, plus two reference benchmarks, await a measurement campaign before their displayed target becomes an observed Elo rating. This honesty is a strength, not an embarrassment to hide: it states exactly what the page knows and what it does not yet know, instead of dressing an assumption up as a certainty.

What actually changes from one rung to the next

What actually distinguishes rung 1 from rung 25 can be checked in the bots’ configuration file, not in a marketing promise. Two settings change across the ladder: the number of search simulations rises from 2 to 64, and the temperature falls from 1.6 to 0.1. The first number says how many sequences of moves the algorithm explores before settling on a move: 2 at rung 1, up to 64 at rung 25, or thirty-two times as much exploration at the top of the ladder as at the bottom.

The second setting, temperature, controls how closely the final choice follows the option rated highest by that exploration: a high temperature allows more varied moves, while a low temperature almost always narrows the choice to the highest-rated candidate. These settings change how the bot decides — not yet the quality of the judgment behind that decision: the network that evaluates each position has not yet been trained, as the next section explains.

How the engine is designed

The engine is designed as a neural network coupled with MCTS (Monte Carlo Tree Search): the network will suggest a position evaluation and promising moves, and the search tree will test these intuitions by simulating sequences of moves. Its architecture is already fixed — a trunk of 4 blocks with 32 channels, a policy head over the super board’s 81 cells, and a value head with 64 hidden units — but for now its weights come only from a random number generator.

All twenty-five rungs currently point to that same starting point: a single set of weights shared across the ladder while training is pending. The plan is training through self-play, with the network playing against itself to generate its own training data, followed by a promotion arena where each candidate faces the previous best version, before a calibration campaign measures where each rung really sits on the Elo scale. None of this has happened yet. The network has not learned to play: it is still waiting for its first training game.

Choose your rung and improve

You do not need to know your level in advance to choose a rung. If you simply want to play tic-tac-toe against an AI without waiting for a human opponent to connect, start at rung 1: it is the lowest entry point on the ladder, and you can replay it as often as you need before moving up. The rung list is always available: you do not have to play in order, but following the ladder gives you a path if you do not know where to start.

The page’s progress bar shows how many rungs you have already beaten, and a check mark identifies each one in the list. Once you beat a rung, it is better to face the next one than to skip several at once: each rung makes only a small change to how the bot searches for moves, and advancing one at a time lets you feel the difference. Choosing your side — first player, second player or random — before starting the game also changes the experience, without changing the rung you face.