返回探索

Jev硬事实核查复测:与GPT-5.4无显著差异,成本仅五十分之一

社区多次复测显示,Jev在硬事实核查上与GPT-5.4基本持平(3.3个百分点差距,不显著),成本约为后者的五十分之一。

I said Jev beat GPT-5.4 on hard fact-checks by 11.1 points. That was one run. Rerun five times, I can't tell them apart: 3.3 points, not significant. Jev costs a fiftieth as much. Raw runs: https://github.com/TheWayWithin/jev-bench… https://jamiewatters.work/journey/jev-ties-gpt-5-4…

Jev硬事实核查复测:与GPT-5.4无显著差异,成本仅五十分之一 1
· 0 次赞在 X 打开