PlayES

How it works.

A chess bot that tries to play like me, not like an engine. This is what happens every time it's on move.

of my games
9,000
book positions
5,720
parameters
79M
server calls per move
0

01A model of a person

Engines look for the best move. Maia doesn't: it's a neural network trained to predict the move a human would actually play, given their rating. I took Maia-3 and fine-tuned it on about 9,000 of my games with LoRA, small adapter weights trained on top of the original model, so its guess stops being “a human at this rating” and becomes me.

It sees the last eight positions, not just the current one, and plays the move it thinks I'm most likely to play. One pass through the network, no search. Its mistakes are, in theory, mine.

02My opening book

The first moves don't need a network. The bot carries an opening book of about 5,700 positions from my games, with the moves I played in each and a weight for every move. It picks one at random by weight, ignoring flukes: moves I played there fewer than twice, or with less than 10% of the weight.

That's why it always opens 1.e4 as White, and as Black answers 1.e4 with the Sicilian and 1.d4 with 1...Nf6. So do I. Book moves are tagged “bk” in the move list.

03Thinking on my clock

The pause before each move isn't the bot calculating. A second, much smaller network predicts how long I'd take: it looks at the time control, both clocks, the move number, how many legal moves there are, whether it's in check, whether you just took something, how long I took on my last move, how long you took on yours, and how hard the position looks to plain Maia.

It predicts a spread rather than a single number, and the bot draws a time from it, so the same position doesn't always take the same time. It never spends more than 60% of what's left on its clock. It's charged exactly that time, never the time the computer took, and the first move of each side is free.

04Resigning and draws

Before every move it asks whether to resign. A third small network, trained on when I resign, weighs plain Maia's verdict on the position, the material and both clocks, and returns a probability; the bot rolls the dice. Two hard rules on top: it never resigns while its expected score is above 20%, or while it isn't down material.

Draws are plain rules. Offer one and it accepts from around move 27 if the position is dead equal with 16 points of material or less left on the board, if it could claim the draw anyway, or if it's worse, under 15 seconds and you have more time. It offers draws itself only in those dead-equal endgames, and rarely.

05It runs in your browser

There's no server doing the thinking. Everything above runs on your device, in a web worker, with onnxruntime-web on WebAssembly. The opening book and the small networks come first, under a megabyte, so you can start playing while the rest downloads.

Phones, slow connections and devices with little memory get Maia-3 23M (27 MB) with my adapters as a switch: on, it plays my moves; off, it's plain Maia, which feeds the clock, resignations and draws. Everything else also fetches Maia-3 79M (83 MB) with my adapters merged in, and switches to it once it arrives. Both are int8, one byte per weight, and both stay in your browser's cache, so a second visit loads from disk.

Credits

Built on Maia-3, from the Computational Social Science Lab at the University of Toronto. Maia-3 is licensed under AGPL-3.0, and so are the bot's model files.

Play me