Engines, NNUE & vibe coding
Workshop notes: how clrsrc comes about, what happens during NNUE training, and how to build an engine with AI assistance.
AI notice: The posts here are created with the support of AI systems; I review them editorially and take responsibility for them before publication.
Articles
-
clrsrc 1.3.0: sharper search, cleaner endgames
A pure search-and-endgame release: the neural net stays unchanged, all improvements sit in the search - singular extensions switched off, clean endgame conversion, correct mate scoring.
-
The load-bearing bug
A fresh-eyes audit of the whole codebase shipped two fixes - and one surprising lesson: the intuitive bugfix made it weaker, while switching the technique off gained +20.9 Elo. Measure, don't assume.
-
The bot in your pocket
A small Android app shows the games of clrsrc_lc0 live and notifies you of every new and finished game - open source on GitHub, with an APK to download.
-
Seven at one blow
+104 blitz Elo in one week - the sum of seven individually SPRT-gated improvements: a new NNUE generation, search, book, endgame, time management and infrastructure.
-
Caught and contained
After an update, our Lichess bot ran into Lichess's request rate limit - how we cleared the self-inflicted lock with monitoring, source-code reading and a three-layer fix.
-
A data center built from leftovers
Desktop, old laptops, smartphones, a rented VPS: how this becomes a distributed datagen cluster - hardened against heat and battery drain, split across three operating systems.
-
Why I build my own chess engine
There's Stockfish - so why clrsrc? On learning, building it yourself and the appeal of doing it from scratch.