133 GPT-6 agents spend 6 days tuning Stockfish NNUE weights, gaining +17.66 Elo in 256-game match
imjustnewatai · x · 2026-11-12
A user ran 133 GPT-6 Astra agents for 6 days to edit Stockfish's NNUE weights and released an experimental build pairing Shawn Xu's test1758 network with a history-floor executable.
- In a 256-game development match against the official November 6 build, it scored 134.5–121.5 (51 wins, 167 draws, 38 losses)
- Estimated gain: +17.66 Elo, with a nominal paired-bootstrap 95% interval of +1.36 to +34.04
- Conditions: 128 development openings, both colors, 10s+0.1s/move, single thread, 64 MiB hash
- Official Stockfish adopted the same test1758 weights on November 10; a gain over the highest-rated Stockfish remains unverified
The Windows BMI2/AVX2 build and match PGNs are available on GitHub for reproduction in any chess GUI.
More from coding & agent
- StalkChain: an MCP server giving Claude live crypto wallet data, PnL, and FOMO trader flows — Altruistic_Ranger_57 · 2026-11-12
- Box CTO Ben Kus: Upgrading to Frontier Models Requires Constant Re-Architecture — Sam_Witteveen · 2026-11-12
- Voice AI engineering checklist: 7 systems production-ready voice agents need — goyalshaliniuk · 2026-11-12
- Grok Bot 1.0 joins video calls as an actual AI teammate — XFreeze · 2026-11-12
- 7 Engineering Decisions Behind Production-Ready Voice Agents, From Architecture to Evals — goyalshaliniuk · 2026-11-12
- A task-based decision guide: which Claude tool to use for ask, think, speed, reason, delegate, ship — rubenhassid · 2026-11-12