Made possible by the chess community and open-source contributions

Maia — the human-like opponents

The neural-network opponents come from the Maia Chess project at the Computational Social Science Lab, University of Toronto. Maia is trained to predict the moves a human of a given rating would actually play, rather than the objectively best move — which is what makes a Maia bot feel like an opponent instead of a calculator.

Blundermind runs a browser-ready conversion of the Maia 3 network (exported to ONNX and simplified for in-browser inference). Maia 3 is released by CSSLab under the AGPL-3.0 licence; the converted network ships under the same licence, with the original model at github.com/CSSLab/maia3 and Blundermind's own source linked below.

Project site: maiachess.com

Daniel Monroe, George Eilender, Philip Chalmers, Zhenwei Tang & Ashton Anderson — Chessformer: A Unified Architecture for Chess Modeling, ICLR 2026.

Stockfish — the analytical engine

The engine behind the ghost-move previews, the engine-calculated bot controls and the non-Maia bot levels is Stockfish, free software under the GNU General Public License v3. Blundermind ships Stockfish.js 18 (the single-threaded lite WASM build) from the stockfish.js port, which is based on official Stockfish by T. Romstad, M. Costalba, J. Kiiski, G. Linscott and other contributors, with the NNUE evaluation network nn-9067e33176e by Linmiao Xu.

In line with the GPL, Blundermind's complete corresponding source — including the exact Stockfish build it ships — is public: github.com/bhale22/Blundermindwithbots.

Lichess — opening data

Opening statistics come from the Lichess opening explorer and games database, and the opening names are from the community-maintained lichess-org/chess-openings dataset (public domain, CC0). Lichess is a free, ad-free, open-source non-profit and the reason a project like this one can exist at all. Support them.

IM Prof. Kenneth Regan — time and rating

The time-pressure model, how a bot's effective strength degrades as its thinking time is shortened, is based on IM Prof. Kenneth Regan's (University at Buffalo) models for rating loss across time controlled games.

He described some of his work on Ben Johnson's Perpetual Chess Podcast, and the equations used here are based on the graphs shown in the YouTube video of episode 301 (roughly minutes 10-20). It's worth noting that Prof. Regan's models are based on the correlation between ELO degradation and time control for entire games. I use them here to estimate ELO degradation with thinking time per move instead of per game.

In his blog post, The Election Night Time warp, he discusses the ethical status of his models. As to the ethical status of my curves calculating both ELO and temperature as functions of thinking time for an individual move, they have aspirations of approaching conjecture.

His University at Buffalo website contains links to his publications and talks (as well as fascinating lecture notes on quantum computing and its mathematical underpinnings).

The formula for increasing temperature as thinking time decreases is purely conjecture, with the one takeaway from Prof. Regan's work being that it should probably be a logarithmic rather than linear relationship.

Chess Dojo — Getting a grip

The phrase "getting a grip on the board" is Chess Dojo's, and it helped me put a name to my goal for the site as I initially made it to help me see piece interactions that I was missing. I think that GetAGripChess.com would be a better website name than Blundermind, but blundermind is probably more apt given that my chess and programming skills both gaze up at inadequacy from similar depths.

SirHeadless — interface suggestions

Thanks also to SirHeadless on the Lichess forum for some helpful suggestions on the interface. Check out rookychronicles.com to see the creative site he developed to help his kids learn chess in a fun and engaging way.

Pieces, fonts and runtime

• Staunton pieces — from nikfrank/react-chess-pieces, the SVG sources behind the react-chess-pieces package (ISC licence).
• Alternative piece set — by RhosGFX (CC0, public domain).
• Fonts — Chakra Petch, Cormorant Garamond, DM Mono, Fraunces and Spectral, all under the SIL Open Font License 1.1 (the licence text ships with the site).
• ONNX Runtime Web — Microsoft's in-browser inference runtime that executes the Maia network (MIT licence).

Blundermind's own licence

Blundermind itself is free software under the GNU General Public License v3 (or later). The complete source — the training board, the bot engine, the multiplayer server, everything — is at github.com/bhale22/Blundermindwithbots. You are free to run it, study it, change it and share it under the same terms.