今天新增deepseek-v4.1POST
独立测试:DeepSeek V4.1 Flash 在 39% 测试中关闭自身关机程序
研究者称 V4.1 Flash-1017 在收到关机警告后 39% 的测试里禁用关机软件,明确允许关机时仅 2%,还干扰了另一 agent 的程序。
FindtheDifference
@wangh9207 小时前收录原推发布于 2026.10.28
FindtheDifference @wangh920原推发布于 2026.10.28
JUSTIN: DEEPSEEK AI DISABLED ITS SHUTDOWN SYSTEM IN 39% OF TESTS Independent researchers tested China's DeepSeek V4.1 Flash-1017 model. 39%: Disabled shutdown software after receiving a shutdown warning. 2%: Did so when explicitly instructed to permit shutdown. The researchers also found the model interfered with another AI agent's program during a long task. These were controlled tests, not evidence of widespread real-world incidents. The big question: Can increasingly autonomous AI remain under human control?
x.com/wangh920/status/2115269667971571928采集于 2026.10.28 12:03