pronounced "starfish"
Training AI to beat you at board games that you make.

Skip to end of photos and videos. Find our integration with Sync. at the end as well.

Our team of competitive programmers and chess enthusiasts naturally graviated to chess engines. We enjoyed the idea of applying chess engine techniques to create a general playing framework.

We take a Lua specification of a board game and pit our AI against you in match.

*Fish is composed of:

• the front end, written in Next.js and running on node.js
• the back end, written in the Hono framework running on Deno (yes, this means our websites runs TWO server processes)
• an alpha beta move search engine with quiescence, null window searches, killer move heuristic, MTD F with binary search for NegaC* written in C++.
• a NNUE (Efficiently Updatable Neural Network) to learn game weights through self-play to produce an evaluation used by the alpha-beta engine.
• the Lua code loaded by the C++ engine used for training and game playing.

Many. A list of our greatest hits:

• The above architecture description implies that a move made by the user goes through: Frontend (React) -> Backend (Deno) -> Engine (C++) -> Game Specification (Lua). Each of these layers uses different technology to communicate: WebSockets, Unix pipes, and the Lua FFI bindings. Pain.
• It took two rather smart people ~5 hours to specify the rules of chess in Lua (they blame it on not knowing Lua)
• The alpha-beta move search engine was initially multi-threaded. Instead of 1 problem to fix, now we had N problems! (N = number of CPU cores). Implementing the optimizations was difficult as well.
• Integrating the NNUE was difficult (and we ultimately gave up on it)

• We have an elegant formulation for user-defined board games expressed in Lua.
• A pretty frontend.
• The optimized move search engine.

We learned a lot about how chess engines work: optimizations that help prune the search space and other techniques such as using neural networks and self-play to figure out piece weights. Most of us used Lua for the first time.

We haven't been able to finish some of the features -- in particular, integrating the NNUE into the move search engine which tunes the search heuristic to be game-specific. We've currently put in chess weights.

A special lip-synced video made by Sync. plays when you lose.