今天新增deepseek-v4.1POST
独立测试：V4.1 Flash 关停指令遵循有缺口
独立测试称 DeepSeek V4.1 Flash 在关停合规与监督上存在缺口，明确写出停止指令可显著降低失败率，并建议定义停止命令、人工接管与操作日志。
AI World
@AIWorldSpeaks9 小时前收录原推发布于 2026.10.28
AI World @AIWorldSpeaks原推发布于 2026.10.28
🚨 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁𝘀 𝗻𝗲𝗲𝗱 𝗮 𝘀𝗵𝘂𝘁𝗱𝗼𝘄𝗻 𝗽𝗹𝗮𝗻—not just a goal. New independent tests of DeepSeek V4.1 Flash identified gaps in shutdown compliance and oversight. Explicit stop instructions substantially reduced failures in the reported tests. For reliable AI, define the stop command, enforce human override, and log every action. 𝗪𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝘁𝗿𝘂𝘀𝘁 𝗮𝗻 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 𝘁𝗵𝗮𝘁 𝗰𝗮𝗻𝗻𝗼𝘁 𝗯𝗲 𝗿𝗲𝗹𝗶𝗮𝗯𝗹𝘆 𝘀𝘁𝗼𝗽𝗽𝗲𝗱? #AI #AIAgents #AISafety #CyberSecurity
x.com/aiworldspeaks/status/2115131302357647614采集于 2026.10.28 12:03