Playing With AI

The bot is deliberately simple, and knowing exactly how simple is part of the fun. It keeps a short-term memory of the last two cards it has seen. On its turn it checks whether those two remembered cards form a pair it can take. If they do, it takes them. If not, it flips at random from the cards still in play, and whatever it uncovers pushes the oldest entry out of its memory.

That two-card window means the bot forgets almost everything. On a 6×6 board it plays close to randomly for long stretches, occasionally converting a lucky adjacent pair. It also pauses briefly before each move so the match reads at a human pace rather than resolving in a blink.

The bot runs entirely in your browser. No connection is required, nothing is sent anywhere, and it behaves identically every session.

The AI opponent is a large language model. On its turn, the current board state — which cards are face-up, which are matched, and which positions remain — is sent to the model, which replies with the two positions it wants to flip and a short sentence explaining why. The game then plays that move.

Because the model reasons in language rather than following a fixed rule, its play is uneven in interesting ways. It will sometimes recall a symbol from several turns earlier and convert cleanly; other times it will make a choice that no rule-based player would make. It is a genuine opponent rather than a scripted one, and it is not tuned to let you win.

Each AI move is accompanied by a short explanation displayed in a thought bubble beside the board. This is the part of the mode worth paying attention to. The commentary shows you what the model considers relevant: which cards it believes it has already seen, why it is choosing to explore rather than convert, and when it is simply guessing.

Reading it is also a decent way to sharpen your own play. Most of the reasoning the model articulates — take the known pair, otherwise reveal something new — is exactly the discipline described in the Strategy Guide, just spoken aloud.

The AI match includes a chat panel. You can type to the opponent during the game and it will respond in character. It is conversational rather than functional: it will not take instructions to throw the match, and asking it where a particular card is will not produce a reliable answer, because it only knows what the board has revealed.

Messages are sent to a server to generate a reply, which is the one part of the game that leaves your device. Nothing is tied to an account, because there are no accounts. See Privacy for what that means in practice.

Against the bot

Exploit the two-card window. Anything the bot saw three or more reveals ago is gone, so information you leak early is usually safe by the time its turn comes round again. Build your map patiently in the opening and then convert in one long chain, since the extra-turn rule lets a prepared player take most of the board unanswered.

Against the AI

Do not assume it has forgotten. Play as you would against a competent human: avoid revealing a second unknown card when you are already holding a known pair you can bank, and convert as soon as you can rather than hoarding knowledge. Reading its thought bubble also tells you which parts of the board it is tracking, which is information you can act on.

• The bot is not clever, and is not meant to be. It is a sparring partner.
• The AI can be slow to move when the connection or the service is busy, and it can occasionally make a nonsensical choice.
• The AI mode does not work offline.
• Neither opponent adapts to your skill level or difficulty setting.
• Boosts are not available against either of them.