WHAT NEW IN X11.03 TUE

X 上的 AI 合集档案 · 22 个合集 · 66,612 条

gpt-6THREAD

GPT-6 Astra 最低推理档也带思维链

作者发现 Astra 无法关闭 CoT,最低档为 low,但乘法准确率仍达 99.6%。

Maksym Andriushchenko
@maksym_andr2026.09.17收录原推发布于 2026.09.17
THREAD作者分 2 层发完
  1. Fun fact: GPT-6-Astra cannot be even run without CoT! The lowest possible effort is 'low' which leads to almost perfect multiplication accuracy, 99.6%!

    GPT-6 Astra 最低推理档也带思维链(第 1 层)
  2. Can frontier LLMs multiply numbers without using external tools? Absolutely. Claude Opus 5 gets 100% accuracy on multiplications up to 20x20 digits (1200/1200 correct). This accuracy does require max reasoning, however! It's an extremely simple eval, but it becomes relevant again for measuring no-CoT capabilities, particularly of recurrent-depth architectures (such as GPT-6-Astra).

    GPT-6 Astra 最低推理档也带思维链(第 2 层)
Maksym Andriushchenko @maksym_andr原推发布于 2026.09.17
在 X 上打开原推
x.com/maksym_andr/status/2100275591811412019采集于 2026.09.17 10:17