Positive users praise SAO optimization for stable 1000-step training and benchmark wins over GRPO, while negative users criticize persistent service instability, reconnections, and weak inference infrastructure.
Based on 20 visible X reactions from 34 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@jietang The 1,000 stable steps may matter more than the benchmark wins. Better reasoning compounds only if optimization can keep running.
5:07 AM · Jul 14, 2026@the_alex_boudot @jietang yeah the thousand stable steps part matters
5:28 AM · Jul 14, 2026@roaring_doggie @jietang love this! incredibly insightful
12:24 AM · Jul 14, 2026@teortaxesTex Ya been doing this first since intellect 3, I think a part of the convo of do big labs use GRPO mistakenly group in the use of group advantages with the actual algo
9:59 AM · Jul 14, 2026@teortaxesTex Ya been doing this first since intellect 3, I think a part of the convo of do big labs use GRPO mistakenly group in the use of group advantages with the actual algo
9:59 AM · Jul 14, 2026Positive users praise SAO optimization for stable 1000-step training and benchmark wins over GRPO, while negative users criticize persistent service instability, reconnections, and weak inference infrastructure.
Based on 20 visible X reactions from 34 accounts; directional sample.
Ask a question below.
Published answers will appear here.