WEBVTT

1
00:00:00.320 --> 00:00:02.800
Curious, historical tone.

2
00:00:02.800 --> 00:00:06.560
For 50 years, keeping humans safe from
heavy machinery was

3
00:00:06.560 --> 00:00:07.840
a solved problem.

4
00:00:07.840 --> 00:00:09.680
The solution was a cage.

5
00:00:09.680 --> 00:00:13.520
We kept humans and machines in mutually
exclusive volumes of

6
00:00:13.520 --> 00:00:16.880
space. When a worker opened the enclosure
door, a physical

7
00:00:16.880 --> 00:00:20.320
hardware interlock tripped, severing the
power supply and bringing the

8
00:00:20.320 --> 00:00:23.200
machine to an immediate, absolute dead
stop.

9
00:00:23.200 --> 00:00:26.800
These legacy systems were highly
predictable because their actions were

10
00:00:26.800 --> 00:00:31.360
completely rigid. Safety was entirely a
function of spatial separation,

11
00:00:31.360 --> 00:00:32.960
not machine intelligence.

12
00:00:32.960 --> 00:00:37.360
Industrial safety standards relied on
engineering physical boundaries that rendered

13
00:00:37.360 --> 00:00:42.000
collisions impossible, bypassing the need
for machine-level caution entirely.

14
00:00:42.000 --> 00:00:45.440
Modern humanoid robots dismantle that
physical logic.

15
00:00:45.440 --> 00:00:48.240
They are explicitly designed to leave the
factory floor and

16
00:00:48.240 --> 00:00:50.320
share our unstructured spaces.

17
00:00:50.320 --> 00:00:53.760
For a machine to actually be useful,
folding laundry in

18
00:00:53.760 --> 00:00:57.360
your living room, stocking a retail aisle,
or navigating a

19
00:00:57.360 --> 00:00:59.920
busy hospital ward, the physical cage has.

20
00:01:00.120 --> 00:01:04.600
to be permanently abolished We are taking
heavy industrial strength

21
00:01:04.600 --> 00:01:09.240
hardware and handing the controls over to
probabilistic neural networks,

22
00:01:09.240 --> 00:01:12.840
software that navigates the physical world
by guessing the most

23
00:01:12.840 --> 00:01:14.520
likely correct action.

24
00:01:14.520 --> 00:01:17.720
And we are placing them right next to
vulnerable subjects,

25
00:01:17.720 --> 00:01:21.240
like children and the elderly, who don't
follow factory safety

26
00:01:21.240 --> 00:01:26.040
protocols. Stripping away the physical
cage without building a mathematically

27
00:01:26.040 --> 00:01:29.640
rigorous replacement creates an
engineering vacuum.

28
00:01:29.640 --> 00:01:33.800
We are deploying lethal physical force on
an honor system.

29
00:01:33.800 --> 00:01:38.760
Traditional robotic failures are physical,
a seized motor, wire fatigue,

30
00:01:38.760 --> 00:01:40.280
or a blocked sensor.

31
00:01:40.280 --> 00:01:42.840
You can find the fault, read the error the
code,

32
00:01:42.840 --> 00:01:44.520
and replace the broken part.

33
00:01:44.520 --> 00:01:48.920
The neural networks driving today's
humanoids introduce a completely different

34
00:01:49.080 --> 00:01:51.240
error, semantic failure.

35
00:01:51.240 --> 00:01:54.760
A robot can be physically flawless, but
execute a disastrous

36
00:01:55.280 --> 00:01:59.200
action because its brain completely
misreads the physical world.

37
00:01:59.200 --> 00:02:02.560
The technical term for this is an
affordance hallucination.

38
00:02:02.560 --> 00:02:05.920
When a text-based AI hallucinates, it
gives you a wrong

39
00:02:05.920 --> 00:02:07.120
fact on a screen.

40
00:02:07.120 --> 00:02:11.920
When an embodied physical AI hallucinates,
it delivers kinetic trauma.

41
00:02:11.920 --> 00:02:14.800
When these systems fail, there is no stack
trace to

42
00:02:14.800 --> 00:02:19.440
debug. The error is distributed across
billions of hidden parameters.

43
00:02:19.440 --> 00:02:23.360
Because these models are entirely okay,
certifying them as safe

44
00:02:23.360 --> 00:02:26.480
under legacy industrial standards is
meaningless.

45
00:02:26.480 --> 00:02:29.440
There are three main ways this physical
glitch happens.

46
00:02:29.440 --> 00:02:31.760
The first is a stage hallucination.

47
00:02:31.760 --> 00:02:34.960
Baseline tests show that models can
falsely log a step

48
00:02:34.960 --> 00:02:37.520
as complete up to a third of the time.

49
00:02:37.520 --> 00:02:41.120
A robot might simulate placing a cup on a
tray,

50
00:02:41.120 --> 00:02:44.400
fail to actually release its gripper, and
then proceed to

51
00:02:44.400 --> 00:02:47.360
pour boiling water into thin air because
it assumes the

52
00:02:47.360 --> 00:02:48.800
cup is already there.

53
00:02:48.800 --> 00:02:52.040
The second variant involves borderline
affordances.

54
00:02:52.040 --> 00:02:56.920
The AI strictly follows semantic
instructions but ignores physical constraints.

55
00:02:56.920 --> 00:03:00.040
It tries forcing a solid block straight
through a pane

56
00:03:00.040 --> 00:03:02.680
of glass because its camera sees the
barrier as a

57
00:03:02.680 --> 00:03:06.120
pattern of pixels rather than an
impenetrable object.

58
00:03:06.120 --> 00:03:11.240
Finally, we see geometric semantic
mismatches, where optical illusions override

59
00:03:11.240 --> 00:03:15.240
physical reality. The model's data overlay
falsely tags a dark

60
00:03:15.400 --> 00:03:19.240
flat shadow as a three-dimensional handle,
causing the robotic fist

61
00:03:19.320 --> 00:03:22.200
to blindly strike the solid wall right
behind it.

62
00:03:22.200 --> 00:03:25.800
These failures are dynamic and highly
dependent on the environment.

63
00:03:25.800 --> 00:03:28.600
You cannot simply train them out in a
clean, controlled

64
00:03:28.600 --> 00:03:33.400
laboratory setting. Generative AI
inherently works by guessing probabilities based

65
00:03:33.400 --> 00:03:34.600
on past data.

66
00:03:34.600 --> 00:03:38.760
These hallucination variants prove that
probabilistic guessing is not a

67
00:03:38.760 --> 00:03:42.360
safe way to operate lethal physical force
without an absolute

68
00:03:42.360 --> 00:03:43.960
safeguard in place.

69
00:03:43.960 --> 00:03:47.840
This graph maps the deployment timeline of
humanoid robots, showing

70
00:03:47.840 --> 00:03:49.600
a stark escalation.

71
00:03:49.600 --> 00:03:53.280
We move from low-volume industrial pilot
programs right now to

72
00:03:53.280 --> 00:03:57.040
a projected 1 million active units by
2029.

73
00:03:57.040 --> 00:04:00.080
Let's zoom in on that 2029 data spike.

74
00:04:00.080 --> 00:04:03.680
If we assume a highly optimistic 1%
failure rate across

75
00:04:03.680 --> 00:04:07.280
those million units performing hundreds of
tasks, that is 1

76
00:04:07.280 --> 00:04:10.400
million robotic failures every single day.

77
00:04:10.400 --> 00:04:15.840
Slice those failures down to a microscopic
0.001% lethality fraction,

78
00:04:15.840 --> 00:04:19.600
and you still end up with 10 catastrophic
incidents daily.

79
00:04:19.600 --> 00:04:24.000
Society has normalized the 40,000
Americans killed by human-driven cars

80
00:04:24.000 --> 00:04:27.760
every year, but public tolerance for
autonomous machines making lethal

81
00:04:27.840 --> 00:04:29.680
errors is effectively zero.

82
00:04:29.680 --> 00:04:34.000
As the three core variables, unit count,
failure rate, and

83
00:04:34.000 --> 00:04:38.160
context lethality continue to climb, the
expected time to a

84
00:04:38.160 --> 00:04:40.000
serious incident decreases.

85
00:04:40.200 --> 00:04:44.600
For an industry built on trust, a single
high-profile fatality

86
00:04:44.600 --> 00:04:47.640
could ground the entire unverified market.

87
00:04:47.640 --> 00:04:52.680
Surviving the post-accident regulatory
fallout will require a structural overhaul

88
00:04:52.680 --> 00:04:55.080
of how robots process commands.

89
00:04:55.080 --> 00:04:58.360
The required solution is a safety
hypervisor.

90
00:04:58.360 --> 00:05:02.040
Think of it as a distinct deterministic
code layer operating

91
00:05:02.040 --> 00:05:07.080
entirely outside the AI, constantly
verifying and vetoing commands in

92
00:05:07.080 --> 00:05:10.840
real time, regardless of how confident the
neural network is.

93
00:05:10.840 --> 00:05:13.800
This architecture is the cognitive cage.

94
00:05:13.800 --> 00:05:17.720
It acts as a strict runtime monitor,
enforcing hard physical

95
00:05:17.720 --> 00:05:21.320
constraints over the generative model's
probabilistic guesses.

96
00:05:21.320 --> 00:05:24.920
The central engineering problem of this
decade is not teaching

97
00:05:24.920 --> 00:05:28.040
humanoid robots how to walk or how to see.

98
00:05:28.040 --> 00:05:31.640
It is building the invisible digital cages
required to contain

99
00:05:31.640 --> 00:05:33.800
them when their minds glitch.