Climbchess
About Climbchess
Climbchess is an adaptive chess trainer that teaches you the patterns chess engines actually think about — extracted directly from a neural network rather than guessed at by a human author.
How it works
We trained sparse autoencoders (SAEs) on the residual stream of Leela Chess Zero (Lc0), a top open-source chess neural network. The SAEs decompose Lc0's internal representations into 320 interpretable concepts — back-rank weakness, knight outposts, pawn-storm structures, king-safety patterns, and hundreds more.
Each puzzle in the library is tagged with the concepts the network is thinking about when it solves it. You don't drill random tactics — you drill the specific pattern your network is weakest at, with adaptive Elo per concept.
Who built it
Climbchess is built and maintained by Michael Usov at MRU Consulting (Australia). It grew out of mechanistic interpretability research on Lc0's residual stream.
What it costs
The core trainer is free. We may add paid tiers in the future for cloud sync upgrades or curated curricula; existing free features will stay free.
Feedback
Use the floating "💬 Feedback" button anywhere in the app to send a message. We read every one. Or email [email protected].