WHAT NEW IN X11.01 SUN

X 上的 AI 合集档案 · 22 个合集 · 63,170 条

gpt-6VIDEO

Claude Opus 5.5 与 GPT-6 Astra 对弈 22 步被将死

两模型纯靠自身推理下完整盘棋,Opus 执白第 22 步走 Nc2 漏防 b2,Astra 执黑 Qxb2 将死;Opus 复盘承认预感到攻势却防错方向。

Kinder
@kinderlayer13 天前收录原推发布于 2026.10.15
Kinder @kinderlayer原推发布于 2026.10.15
在 X 上打开原推

holy sh*t. Claude Opus 5.5 vs GPT-6 Astra - a frontier model just got mated in 22 moves. a full chess game. no engine, no internet, just two models thinking move by move. fresh sessions, high effort, no bot, no script, no lookup. Opus on white, Astra on black, one board between them. Opus castled queenside and looked completely fine. then Astra went quiet and started prying open the b-file - slowly, one move at a time, straight at Opus’s king. by move 20 you could feel it. Opus’s king had no air left. every black piece was pointing at the same corner. move 22. Opus plays Nc2 - the one move that stops guarding b2. Astra pounces: Qxb2#. mate. and here’s the brutal part. Opus knew. its own post-game note: “I was worried about ...Ba3 and missed it.” it felt the attack coming, braced for the wrong threat, and walked into the real one. final: 0-1. a frontier model, checkmated in 22 - by nothing but the other model’s head. so what is this: Astra’s just the sharper tactician, or Opus choked under its own overthinking? follow - I run a new model showdown every week. chess was only round one.

这条是引用着另一条写的:x.com/kinderlayer/status/2109997714591068522

x.com/kinderlayer/status/2110709758382502008采集于 2026.10.18 16:02