Astra (Extra High) loses to chess.com's 1300 ELO bot Nelson in live test
MikePFrank · x · 2026-09-06
Mike Frank tested Astra unaided at the Extra High setting and found it loses to chess.com's Nelson bot (ELO 1,300), then retried with Ultra mode. He also let the model instrument itself with a single-move board visualization capability to help it avoid obvious blunders — a vivid case of a frontier model struggling with spatial/state reasoning.
Related event: Astra Loses to 1300-ELO Chess Bot but Fights Back with Self-Built Tools(6 posts)→
More from Fun
- Joke: OpenAI should make Codex and ChatGPT point to the same product, Hardee's-style — soleio · 2026-09-06
- Dev dubs GPT-10 Astra in Codex "Astral Codex Ten" — jam3scampbell · 2026-09-06
- Employee takes CEO to first rave to "help his training data" with The Chainsmokers — ashleyyjoelle · 2026-09-06
- Dev builds surprisingly fun isometric puzzle-combat game entirely with GPT-6Astra High — IanArawjo · 2026-09-06
- Remember when we trained RNNs on MIDI? An AI veteran reminisces — cephaloform · 2026-09-06
- Dev hears their name on a real N64 as experimental support ships on GitHub — turtlesoupy · 2026-09-06