1
00:00:00,218 --> 00:00:00,435
How can

2
00:00:00,435 --> 00:00:01,599
AI safety promises

3
00:00:01,599 --> 00:00:02,512
be checked?

4
00:00:02,512 --> 00:00:04,187
Dario Amodei worries

5
00:00:04,187 --> 00:00:04,798
that safeguards

6
00:00:04,798 --> 00:00:05,409
could lag

7
00:00:05,409 --> 00:00:06,724
behind increasingly capable

8
00:00:06,724 --> 00:00:07,521
AI.

9
00:00:07,521 --> 00:00:08,819
He proposes outside

10
00:00:08,819 --> 00:00:11,534
evaluators embedded inside companies.

11
00:00:11,625 --> 00:00:12,387
With ongoing,

12
00:00:12,387 --> 00:00:14,421
employee-like access to tools,

13
00:00:14,421 --> 00:00:15,275
workspaces,

14
00:00:15,275 --> 00:00:15,543
and

15
00:00:15,543 --> 00:00:16,145
staff.

16
00:00:16,145 --> 00:00:16,480
They would

17
00:00:16,480 --> 00:00:17,845
check safety commitments

18
00:00:17,845 --> 00:00:19,738
against actual practices.

19
00:00:19,738 --> 00:00:20,366
Including how

20
00:00:20,366 --> 00:00:21,613
models are trained,

21
00:00:21,613 --> 00:00:22,267
tested,

22
00:00:22,267 --> 00:00:23,283
and released.

23
00:00:23,283 --> 00:00:23,652
They would

24
00:00:23,652 --> 00:00:24,740
report incidents and

25
00:00:24,740 --> 00:00:26,256
publish key findings,

26
00:00:26,256 --> 00:00:27,311
with limited redactions

27
00:00:27,311 --> 00:00:28,952
for protected information.

28
00:00:28,952 --> 00:00:30,585
Amodei announces Anthropic’s

29
00:00:30,585 --> 00:00:31,456
commitment.

30
00:00:31,456 --> 00:00:32,938
Wider company standards

31
00:00:32,938 --> 00:00:34,179
and international

32
00:00:34,286 --> 00:00:34,789
coordination

33
00:00:34,789 --> 00:00:35,844
remain proposals,

34
00:00:35,844 --> 00:00:37,710
with verification challenges.

35
00:00:37,710 --> 00:00:38,137
His pacing

36
00:00:38,137 --> 00:00:39,201
proposal does not

37
00:00:39,201 --> 00:00:40,382
mean halting all

38
00:00:40,382 --> 00:00:41,889
technical progress.

39
00:00:41,889 --> 00:00:42,316
Safety research

40
00:00:42,316 --> 00:00:43,044
and improvements

41
00:00:43,044 --> 00:00:44,268
would continue.