WHAT NEW IN X11.03 TUE

X 上的 AI 合集档案 · 22 个合集 · 66,612 条

gemini-4POST

实测 Gemini 4 识别伪造 API 方法的能力

用混入不存在方法的误导性 API 文档测试 Gemini 4,它识别出大部分假方法并给出替代方案,只在一个假方法上仍被误导。

Haider.
@haider120 天前收录原推发布于 2026.10.14
Haider. @haider1原推发布于 2026.10.14
在 X 上打开原推

gemini 4 has definitely made massive progress in reducing hallucinations instead of flashy demos, i tried it on misleading API documentation with a few nonexistent methods mixed in to see if it could spot unsupported claims rather than invent solutions surprisingly it caught most fake methods and suggested alternatives instead of hallucinating working code it still confidently accepted one nonexistent method but incredible progress overall

x.com/haider1/status/2110074599174295620采集于 2026.10.14 10:16