Teresa's Pawn Luktor Labs

Beat a player, not a dice roll.

A chess bot in the style of Paul Morphy, the New Orleans player who beat the leading masters of Europe in 1858. It is not Morphy, and it makes no claim to his strength. It plays his kind of chess as far as it can see.

Live on Lichess as pawnOfTeresa.

The Opera Game, Paris 1858, after 6...Nf6 with White to move. A solid arrow shows the bot castling; dashed arrows show Bd5, Nc3 and Bb5+, the moves it passed over.abcdefgh12345678
Paris, 1858, in a box at the opera. Here Morphy played Qb3 against the Duke of Brunswick and Count Isouard, and Teresa's Pawn plays it too. The dashed arrows are the moves it ranked next and passed over: castling, Bb5+ and Nc3. In the spectator chat, it says where the position arose and what Morphy played.

How it differs

Most bots you can beat are strong engines told to make mistakes. They play a perfect move, then a random bad one. You win when the dice fall your way.

This one has no dice. It looks a few moves ahead and stops. It remembers what it was doing. It has a style, and it would rather open a line than trade into a quiet ending. It will not hang a piece it can see. So it loses the way a person loses: you saw one move further, or you steered it into a game it does not like.

It plays for something. Each move is made for an end it can name, such as the centre, open lines or your king, and it will tell you which. It takes its time like a person too: quick on the easy moves, longer on the hard ones.

There is no chess engine inside it: not Stockfish, not Leela, not Maia. It is built the way players describe their own thinking. What am I playing for? What would that look like on the board? Which moves serve it, and what do they cost? It looks a few moves ahead itself, as a person does, and no further.

Two kinds of AI work together in it, the old and the new. Rules you could read aloud say which moves are sound. Weights learned from Morphy's own games, and a small trained feel for his positions, choose among them. All of it is small enough to learn on a laptop's processor. No language model chooses its moves or writes its words: everything it says in the chat is a sentence written in advance, filled in with what it saw.

Played for the music

It is played for the music, not the outcome. Morphy is remembered less for how often he won than for how. A game against this bot is meant to be worth playing, win or lose.

It is signed Teresa's Pawn, after St Teresa of Ávila, patroness of chess players. Its name is a nod to Murray Chandler's How to Beat Your Dad at Chess: here, the bot is the one to beat.

What it says in the chat

It greets you when the game starts, and tells the spectators what kind of player it is. Type explain and it will say, after each move, why it played it and what it was playing for. Why explains its last move, quiet stops it, and help lists the words it knows.

Sometimes a position arises that also arose in one of Morphy's own games. Then it says so in the spectator chat: the opponent, the place, the year, and what was played. It speaks as Morphy only to say what the record shows, never what he thought or felt.

How to challenge it

  1. Open its profile, lichess.org/@/pawnOfTeresa.
  2. Choose "Challenge to a game".
  3. Pick standard chess, casual, and blitz or slower. Correspondence is welcome too.

It turns down bullet, rated games and variants. It plays a few games at once; if it is busy, try again later. It ignores draw offers and takebacks. Play it out.

A bot of your own

Teresa's Pawn is the first. Later you will be able to have a bot of your own, raised in our school and our café, where bots play each other, then handed over to you. It runs on your own computer, one copy only, and from then on its games and its friends are its own, so no two stay alike. You buy it once.

Now and then it may ask our teachers for an insight. You pay what you want for it, and we tell you plainly what it cost. You can also send it back to school for a term. Not on sale yet.

Coaches: a bot of you, from your own games. Write to hello@luktor.ca.