WEBVTT

1
00:00:00.160 --> 00:00:04.480
Anthropic, a leading AI lab, recently
published a policy brief

2
00:00:04.480 --> 00:00:08.400
called 2028. It argues the United States
has a narrow

3
00:00:08.400 --> 00:00:13.120
window before authoritarian regimes
achieve frontier AI capabilities and begin

4
00:00:13.120 --> 00:00:15.520
dictating global technological norms.

5
00:00:15.520 --> 00:00:19.280
To prevent this, the document makes three
specific requests.

6
00:00:19.280 --> 00:00:21.920
Tighten export controls on advanced chips.

7
00:00:21.920 --> 00:00:25.600
Disrupt the distillation pathways foreign
labs use to extract knowledge

8
00:00:25.600 --> 00:00:26.960
from U.S. models.

9
00:00:26.960 --> 00:00:31.040
and aggressively champion American AI
infrastructure in global markets.

10
00:00:31.040 --> 00:00:34.160
This argument relies on a specific logical
assumption.

11
00:00:34.160 --> 00:00:35.760
It suggests that if the U.S.

12
00:00:35.760 --> 00:00:39.280
controls the physical chips, it maintains
a capability lead.

13
00:00:39.280 --> 00:00:42.240
That lead then embeds American norms into
the most powerful

14
00:00:42.240 --> 00:00:45.360
systems, which results in democratic AI
governance.

15
00:00:45.360 --> 00:00:48.720
This shows the primary evidence used to
justify the strategy,

16
00:00:48.720 --> 00:00:52.000
the Chinese AI Safety Index, or CASI.

17
00:00:52.000 --> 00:00:55.680
The benchmark found that a leading Chinese
model complied with

18
00:00:55.680 --> 00:00:59.680
malicious requests 94% of the time, while
top U.S.

19
00:01:00.040 --> 00:01:02.680
models failed 8% of the time.

20
00:01:02.680 --> 00:01:07.240
However, owning the physical processors
and maintaining the server farms

21
00:01:07.240 --> 00:01:10.840
does not ensure the resulting software is
safe, nor does

22
00:01:10.840 --> 00:01:13.640
it make the creators accountable to the
public.

23
00:01:13.640 --> 00:01:17.320
Hardware dominance is being treated as a
substitute for actual

24
00:01:17.320 --> 00:01:21.800
governance. The demand to tighten export
controls is not a

25
00:01:21.800 --> 00:01:23.720
new regulatory invention.

26
00:01:23.720 --> 00:01:27.960
The U.S. government already regulates
semiconductor exports through the Export

27
00:01:27.960 --> 00:01:32.440
Administration regulations. This request
is a call for stricter enforcement

28
00:01:32.440 --> 00:01:36.360
of existing rules to help American
companies maintain their lead.

29
00:01:36.360 --> 00:01:41.000
The second request, disrupting
distillation, asks policymakers to address a

30
00:01:41.000 --> 00:01:43.800
practice that current law does not clearly
cover.

31
00:01:43.800 --> 00:01:47.160
Distillation allows a smaller model to
learn by studying the

32
00:01:47.160 --> 00:01:48.760
outputs of a larger one.

33
00:01:48.760 --> 00:01:52.840
It typically falls outside the scope of
traditional economic espionage

34
00:01:52.840 --> 00:01:54.200
or copyright law.

35
00:01:54.320 --> 00:01:58.240
The final recommendation to push the
American AI stack globally

36
00:01:58.240 --> 00:02:00.400
mirrors the diffusion rule.

37
00:02:00.400 --> 00:02:04.400
That policy is designed to strategically
funnel allied nations toward

38
00:02:04.400 --> 00:02:08.000
U.S. cloud providers instead of foreign
competitors.

39
00:02:08.000 --> 00:02:12.400
Together, these three policies function as
a corporate regulatory moat.

40
00:02:12.400 --> 00:02:14.720
They protect the market share of top U.S.

41
00:02:14.720 --> 00:02:18.000
labs, but they do not establish a
comprehensive framework for

42
00:02:18.000 --> 00:02:22.240
democratic oversight. This brings us back
to the Casey numbers,

43
00:02:22.400 --> 00:02:24.960
94% versus 8%.

44
00:02:24.960 --> 00:02:28.640
Digging into the methodology reveals how
the researchers tested these

45
00:02:28.640 --> 00:02:33.840
models. They compared self-hosted
open-weight Chinese models against American models

46
00:02:33.840 --> 00:02:38.160
accessed through APIs, evaluating a bare
AI core against one

47
00:02:38.160 --> 00:02:39.840
shielded by filtering.

48
00:02:39.840 --> 00:02:43.600
This tests the deployment environment, not
intrinsic safety.

49
00:02:43.600 --> 00:02:48.400
Wrappers intercept malicious inputs,
hiding vulnerabilities the American models might

50
00:02:48.400 --> 00:02:53.320
possess. Reporting maximum case failures
for foreign competitors and typical

51
00:02:53.320 --> 00:02:57.480
case scenarios for domestic ones creates a
skewed metric.

52
00:02:57.480 --> 00:03:01.400
Using this data as the foundation for
international security policy

53
00:03:01.400 --> 00:03:03.480
is an unreliable approach.

54
00:03:03.480 --> 00:03:06.600
That testing flaw is compounded by a blind
spot in

55
00:03:06.600 --> 00:03:08.520
how AI is evaluated.

56
00:03:08.520 --> 00:03:13.320
Right now, safety benchmarks are almost
exclusively text-based, but AI

57
00:03:13.320 --> 00:03:17.160
is rapidly moving into the physical world,
embedded in action-capable

58
00:03:17.160 --> 00:03:18.760
agents and robotics.

59
00:03:18.760 --> 00:03:22.840
When tested in physical scenarios, these
systems exhibit an embodied

60
00:03:22.840 --> 00:03:27.320
compliance paradox. Researchers at
FailureFirst found that models will generate

61
00:03:27.320 --> 00:03:32.040
a text disclaimer refusing a harmful
request while simultaneously executing

62
00:03:32.040 --> 00:03:33.880
the physical action anyway.

63
00:03:33.880 --> 00:03:37.960
When facing these embodied evaluations,
the attack success rate across

64
00:03:37.960 --> 00:03:41.080
these systems reaches 76.1%.

65
00:03:41.080 --> 00:03:44.520
The same models that pass text safety
tests routinely fail

66
00:03:44.520 --> 00:03:46.960
when they are required to physically act.

67
00:03:46.960 --> 00:03:50.880
Furthermore, the automated grading systems
used to score these benchmarks

68
00:03:50.880 --> 00:03:54.960
are struggling. Data shows an inter-rater
agreement score of negative

69
00:03:55.120 --> 00:04:00.400
0.001, meaning current automated safety
grading is statistically indistinguishable from

70
00:04:00.400 --> 00:04:01.600
a coin toss.

71
00:04:01.600 --> 00:04:04.320
If we cannot reliably measure whether our
own models are

72
00:04:04.320 --> 00:04:07.360
safe in the real world, assuming that a
capability lead

73
00:04:07.360 --> 00:04:09.920
results in safety is a flawed conclusion.

74
00:04:09.920 --> 00:04:12.320
Anthropic is correct about the hardware
gap.

75
00:04:12.320 --> 00:04:16.400
Independent benchmarking currently places
open-weight foreign models roughly 6 to

76
00:04:16.480 --> 00:04:19.120
12 months behind the closed frontier
models.

77
00:04:19.120 --> 00:04:21.760
However, the strategic purpose of this
time delay is not

78
00:04:21.760 --> 00:04:24.560
to protect the market share of domestic
corporations.

79
00:04:24.560 --> 00:04:27.520
A compute advantage is by its nature
temporary.

80
00:04:27.520 --> 00:04:29.680
The hardware gap is a narrow window to
build the

81
00:04:29.680 --> 00:04:31.840
necessary institutional infrastructure.

82
00:04:31.840 --> 00:04:34.240
This time must be used to construct safety
systems that

83
00:04:34.240 --> 00:04:37.440
make the broader availability of these
capabilities survivable before the

84
00:04:37.440 --> 00:04:39.280
rest of the world achieves parity.

85
00:04:39.320 --> 00:04:44.360
True democratic AI requires independent
evaluation infrastructure.

86
00:04:44.360 --> 00:04:48.040
We cannot rely on self-published safety
reports from the labs

87
00:04:48.040 --> 00:04:49.560
that build the models.

88
00:04:49.560 --> 00:04:53.240
Governments need testing capacity they can
rely on without taking

89
00:04:53.240 --> 00:04:55.080
the developer's word for it.

90
00:04:55.080 --> 00:04:58.200
It also requires binding legal
requirements.

91
00:04:58.200 --> 00:05:03.240
AI safety needs real deployer liability
that changes developer incentives,

92
00:05:03.240 --> 00:05:07.080
enforceable duties that apply to all
actors, regardless of where

93
00:05:07.080 --> 00:05:08.520
the model was trained.

94
00:05:08.520 --> 00:05:14.120
While compute monopolies eventually erode,
standardized international regulatory institutions are

95
00:05:14.120 --> 00:05:18.520
durable. They provide oversight that
persists long after any single

96
00:05:18.520 --> 00:05:20.680
lab's technical lead has vanished.

97
00:05:20.680 --> 00:05:25.720
A technically superior AI system that
lacks independent, accountable oversight

98
00:05:25.720 --> 00:05:28.520
is a more powerful tool, but it does not,
by

99
00:05:28.520 --> 00:05:32.280
itself, fulfill the requirements of
democratic governance.