Let’s run a tiny chess neural network by hand!
summary
Claim: a tiny chess neural net that fits on two pieces of paper using a sparse board encoding, incremental updates and clamp activation and integer arithmetic is constructed and actually run by hand. The build reuses piece vectors, subtracts and adds them for moves, and simulates one-layer evaluation with quantized weights and SIMD-friendly tricks. It takes hours per evaluation, produces a measurable effect — roughly +4 percentage points versus weak opponents — useful but impractically slow.
AI-written summary · about
chapters (7)
- 0:00Motivation and constraints
- 1:32How chess engines work
- 3:28Designing a manual model
- 5:07Understanding NNUE
- 7:12Simplifying the architecture
- 8:37Testing the performance
- 9:52Executing by hand
related talks
08mMinimax search and the structure of cognition!
18mJavaScript and Artificial Intelligence
44mFast Neural Networks… a no brainer!
18mBuilding a Chess Coach
33mAlpha2048 - how I taught my computer to play 1024
01h 04mWriting a Neural Net from Scratch
41mBrain-inspired computational primitives for reliable neuromorphic intelligence
50mBuilding a Pong playing AI in just 1 hour (plus 4 days training...)