Visions! omens! hallucinations! miracles! ecstasies!

THE REFERENT

You’ve reached the website of Cosmo, a computer person with an interest in artificial intelligence & low-level programming, particularly deep learning and chess engines.

I care very much about the correctness and performance of software, and hope that you do too. I have a continuing fondness for the Rust programming language, and use it extensively in my projects. I work on web services and command-line tools, accidentally standardise data formats, and reïmplement the occasional ML paper.

You may also enjoy the works of my most wonderful friends, at sialia.dev, ettolrach.com, and typ.dev.

PRINCIPAL ENDEAVOUR

ASSORTED MISADVENTURES

• AlphaZero, RL, and SPSA
  On the training objective of AlphaZero, and the use of SPSA to directly optimise win-rate in lc0.
• Better activation functions for NNUE
  Results from Swish and SwiGLU in NNUE.
• Chess networks use a Gated Nonlinear Unit
  In which I analogise pairwise multiplication to the Gated Linear Unit.
• Training dynamics of target weighting
  Does it matter whether you predict the true future, or a stronger version of yourself?
• Deep NNUE
  Important architectural improvements in Viridithas 14.
• More experimentation in NNUE
  King buckets, output buckets, and Elo scaling with network size.
• Experimenting with NNUE configurations
  Results from activation functions, learning rates, and batch sizes.
• NNUE performance improvements
  How to make NNs run fast in chess engines.
• Contra “Grandmaster-Level Chess Without Search”
  A critique of DeepMind's searchless chess paper, with comparison to Lc0.
• An overview of computer chess techniques
  Alpha-beta, PVS, iterative deepening, move ordering, search reductions, etc.
• and more at the Sitemap.