WEBVTT

1
00:00:00.240 --> 00:00:03.520
Open a product page for a new AI model
today,

2
00:00:03.520 --> 00:00:07.760
and you will immediately encounter a very
specific vocabulary.

3
00:00:07.760 --> 00:00:14.240
Systems are described as reasoning,
understanding, thinking, and occasionally hallucinating.

4
00:00:14.240 --> 00:00:18.800
These descriptions have moved far beyond
casual tech industry shorthand.

5
00:00:18.800 --> 00:00:22.160
They operate as formal claims made in
press releases, in

6
00:00:22.160 --> 00:00:26.960
boardroom presentations, and in sworn
testimony before lawmakers by executives

7
00:00:26.960 --> 00:00:29.840
who assert they are building a digital
brain for the

8
00:00:29.840 --> 00:00:33.120
world. A problem arises when we apply the
exact same

9
00:00:33.120 --> 00:00:36.080
vocabulary to a server guessing the next
word in a

10
00:00:36.080 --> 00:00:39.440
sequence and a human mind processing
reality.

11
00:00:39.440 --> 00:00:42.560
It begins to convince the public that
these systems possess

12
00:00:42.560 --> 00:00:44.800
actual biological sentience.

13
00:00:44.800 --> 00:00:49.600
This shared vocabulary effectively blurs
what researchers call the organismic

14
00:00:49.600 --> 00:00:53.440
line, the boundary between systems that
have real physical stakes

15
00:00:53.440 --> 00:00:55.840
and systems that merely simulate them.

16
00:00:55.840 --> 00:00:58.640
To understand how that line got blurred,
we have to

17
00:00:58.640 --> 00:00:59.920
look at the math used to define.

18
00:01:00.360 --> 00:01:05.240
intelligence Neuroscientist Carl Fristen
developed a theory called the free

19
00:01:05.240 --> 00:01:08.680
energy principle. At a high level, it
states that all

20
00:01:08.680 --> 00:01:13.400
self-organizing systems act to minimize
surprise and maintain their internal

21
00:01:13.400 --> 00:01:17.240
states. This mathematical framework
creates an illusion.

22
00:01:17.240 --> 00:01:19.880
Left, an AI lowers its error rate.

23
00:01:19.880 --> 00:01:22.600
Right, an organism avoids starvation.

24
00:01:22.600 --> 00:01:27.320
The formula treats them identically,
removing lethal physical consequences.

25
00:01:27.320 --> 00:01:30.840
The animal dies, the AI simply outputs a
larger loss

26
00:01:30.840 --> 00:01:35.080
number. AI pioneer Jeffrey Hinton focuses
on a technical divide

27
00:01:35.080 --> 00:01:38.680
between these two worlds, distinguishing
between digital hardware and what

28
00:01:38.680 --> 00:01:40.920
he calls mortal computation.

29
00:01:40.920 --> 00:01:44.280
Immortal digital systems allow data to be
perfectly copied to

30
00:01:44.280 --> 00:01:47.560
new hardware, even if the original machine
is destroyed.

31
00:01:47.560 --> 00:01:51.400
But mortal biological systems hold
knowledge inseparable from their physical

32
00:01:51.400 --> 00:01:55.520
device. When the organic structure dies,
its knowledge dies alongside

33
00:01:55.520 --> 00:01:58.640
it. In a digital framework, mortality is
often viewed as

34
00:01:58.640 --> 00:02:01.680
a computational inefficiency or a tax on
power.

35
00:02:01.680 --> 00:02:06.080
But this perspective ignores how deeply a
system's behavior changes

36
00:02:06.080 --> 00:02:09.680
when its learned structure cannot be
separated from a physical

37
00:02:09.680 --> 00:02:11.600
body that can be destroyed.

38
00:02:11.600 --> 00:02:15.280
An organism does not predict its
environment because it is

39
00:02:15.280 --> 00:02:17.600
mathematically elegant to do so.

40
00:02:17.600 --> 00:02:20.880
It predicts because failing to predict is
fatal.

41
00:02:20.880 --> 00:02:26.240
Neuroscientist Anil Seth calls human
perception a controlled hallucination that

42
00:02:26.240 --> 00:02:29.520
is carefully yoked to the real world by
constant error

43
00:02:29.520 --> 00:02:33.920
prediction. The AI industry quickly
co-opted Seth's phrase.

44
00:02:33.920 --> 00:02:39.760
They began calling large language model
confabulations hallucinations, completely unmooring

45
00:02:39.760 --> 00:02:42.960
the concept from its necessary biological
caveats.

46
00:02:42.960 --> 00:02:46.400
The consequences for errors across these
two systems could not

47
00:02:46.400 --> 00:02:47.360
be more different.

48
00:02:47.360 --> 00:02:51.240
If a biological brain hallucinates a
predator, the survival threat

49
00:02:51.240 --> 00:02:55.000
is immediate. When a server rack outputs
an incorrect word,

50
00:02:55.000 --> 00:02:57.720
there is zero physical danger to the
machine.

51
00:02:57.720 --> 00:03:02.920
When the tech industry excuses glitches as
hallucinations, they borrow

52
00:03:02.920 --> 00:03:07.640
the profound flavor of human consciousness
while discarding the metabolic

53
00:03:07.640 --> 00:03:09.400
cost of being wrong.

54
00:03:09.400 --> 00:03:12.360
Which brings us back to the AI products
being sold

55
00:03:12.360 --> 00:03:16.040
today. When a product page claims a model
thinks or

56
00:03:16.040 --> 00:03:19.640
reasons, it implies an organismic
commitment.

57
00:03:19.640 --> 00:03:23.720
It promises a level of intent that silicon
cannot fulfill.

58
00:03:23.720 --> 00:03:28.920
It executes sequential token prediction,
processing text inputs into statistical

59
00:03:28.920 --> 00:03:31.880
outputs through a transformer
architecture.

60
00:03:31.880 --> 00:03:35.640
Running safely on an external grid, it has
no intrinsic

61
00:03:35.640 --> 00:03:39.560
stakes, optimizing equations without
facing the threat of its own

62
00:03:39.560 --> 00:03:43.880
destruction. The human brain is not
running a detached mathematical

63
00:03:43.880 --> 00:03:48.240
loss function. It is running a strict
life-or-death body budget.

64
00:03:48.240 --> 00:03:51.760
Ignoring this line has severe downstream
consequences.

65
00:03:51.760 --> 00:03:55.600
It distorts how lawmakers attempt to
regulate these systems, and

66
00:03:55.600 --> 00:03:58.480
it changes what we forget about our
humanity when we

67
00:03:58.480 --> 00:04:00.400
treat software as a mirror.

68
00:04:00.400 --> 00:04:03.280
The next time a tech CEO claims their
model is

69
00:04:03.280 --> 00:04:07.920
hallucinating or thinking, ask yourself
what the system is physically

70
00:04:07.920 --> 00:04:09.840
paying for its errors.

71
00:04:09.840 --> 00:04:13.120
Because if the answer is nothing, it is
just math,

72
00:04:13.120 --> 00:04:14.800
not a mind.