1
00:00:00,220 --> 00:00:01,010
人工智能的

2
00:00:01,090 --> 00:00:01,980
安全承诺，

3
00:00:02,060 --> 00:00:03,520
谁来核实？

4
00:00:03,620 --> 00:00:04,920
达里奥·阿莫代伊

5
00:00:05,000 --> 00:00:06,418
提出一项方案。

6
00:00:06,598 --> 00:00:08,558
让第三方评估员

7
00:00:08,748 --> 00:00:10,338
长期进驻公司，

8
00:00:10,718 --> 00:00:12,208
获得类似员工的

9
00:00:12,288 --> 00:00:13,398
访问权限。

10
00:00:13,877 --> 00:00:15,607
核查训练、部署

11
00:00:15,687 --> 00:00:16,847
和防护措施，

12
00:00:16,947 --> 00:00:18,537
是否符合承诺；

13
00:00:18,747 --> 00:00:19,557
也评估模型

14
00:00:19,637 --> 00:00:21,497
与训练流程的安全性。

15
00:00:21,701 --> 00:00:23,421
评估员报告事故，

16
00:00:23,721 --> 00:00:25,011
并能独立公布

17
00:00:25,091 --> 00:00:26,031
关键发现，

18
00:00:26,461 --> 00:00:28,241
不受公司编辑控制，

19
00:00:28,491 --> 00:00:29,811
但有保密限制。

20
00:00:30,211 --> 00:00:31,021
文中宣布，

21
00:00:31,341 --> 00:00:31,801
Anthropic

22
00:00:31,881 --> 00:00:32,481
将邀请

23
00:00:32,561 --> 00:00:34,181
外部评估团队。

24
00:00:34,261 --> 00:00:35,121
这是承诺，

25
00:00:35,201 --> 00:00:36,561
不等于已经落实。

26
00:00:36,833 --> 00:00:37,833
跨公司标准

27
00:00:37,913 --> 00:00:39,053
和国际协调，

28
00:00:39,133 --> 00:00:40,673
是另外的提议，

29
00:00:40,883 --> 00:00:42,033
不是已有协议，

30
00:00:42,273 --> 00:00:43,793
也不代表行业共识。