WEBVTT

1
00:00:00.000 --> 00:00:04.160
We have integrated systems built from
billions and sometimes trillions

2
00:00:04.320 --> 00:00:08.320
of individual parameters into the core of
our digital infrastructure.

3
00:00:08.320 --> 00:00:12.720
Neural networks now drive vehicles,
diagnose medical conditions, and generate

4
00:00:12.720 --> 00:00:17.920
complex code. Despite building these
massive systems, the engineers programming

5
00:00:17.920 --> 00:00:21.120
them do not actually possess a
mathematical blueprint for how

6
00:00:21.120 --> 00:00:25.040
they function. Developers build them
through trial and error, tweaking

7
00:00:25.040 --> 00:00:28.160
inputs, running the data, and observing
what happens on the

8
00:00:28.160 --> 00:00:31.120
other side. Ask a researcher why we don't
have a

9
00:00:31.120 --> 00:00:34.800
rigid mathematical explanation for these
networks, and they will usually

10
00:00:34.800 --> 00:00:36.560
offer one of two answers.

11
00:00:36.560 --> 00:00:39.280
Either the math is hopelessly complex
because there are too

12
00:00:39.280 --> 00:00:42.800
many variables, or a blueprint simply is
not necessary because

13
00:00:42.800 --> 00:00:46.000
the current tinkering process gets us the
results we want.

14
00:00:46.000 --> 00:00:49.040
But relying on trial and error means we
cannot guarantee

15
00:00:49.040 --> 00:00:51.200
how a system will behave under pressure.

16
00:00:51.200 --> 00:00:54.960
Without mathematical proof of how a
network learns, we cannot

17
00:00:54.960 --> 00:00:59.280
reliably ensure it will operate safely,
align with human instructions,

18
00:00:59.280 --> 00:01:03.000
or run without wasting massive amounts of
computing power.

19
00:01:03.000 --> 00:01:07.160
A growing coalition of scientists is
rejecting this status quo.

20
00:01:07.160 --> 00:01:09.800
They argue that we can map these systems,
and they

21
00:01:09.800 --> 00:01:13.320
have published a framework detailing an
emerging scientific theory of

22
00:01:13.320 --> 00:01:16.680
deep learning. They call it learning
mechanics.

23
00:01:16.680 --> 00:01:20.440
To safely scale artificial intelligence,
we have to stop guessing

24
00:01:20.440 --> 00:01:23.160
at its behavior and start treating it like
a physical

25
00:01:23.160 --> 00:01:27.080
science. Inside a modern neural network, a
dense web of

26
00:01:27.080 --> 00:01:31.240
parameters constantly adjusts and shifts
every time the system processes

27
00:01:31.240 --> 00:01:35.400
new information. Calculating the exact
path of every single weight

28
00:01:35.400 --> 00:01:37.160
during training is impossible.

29
00:01:37.160 --> 00:01:41.800
The sheer volume of moving parts
overwhelms standard mathematical models.

30
00:01:41.800 --> 00:01:45.640
Physicists use statistical physics for
billions of moving parts.

31
00:01:45.640 --> 00:01:50.440
Instead of tracking every atom, they
calculate macroscopic observables, like

32
00:01:50.440 --> 00:01:52.280
measuring a room's temperature.

33
00:01:52.280 --> 00:01:55.000
Learning mechanics applies this logic to
AI.

34
00:01:55.280 --> 00:02:00.000
Instead of mapping specific weights,
researchers track coarse aggregate statistics

35
00:02:00.000 --> 00:02:01.840
to measure the training process.

36
00:02:01.840 --> 00:02:04.960
By stepping back from the microscopic
details to observe the

37
00:02:04.960 --> 00:02:10.480
broader trends, the unpredictable black
box reveals rigid mechanical laws.

38
00:02:10.480 --> 00:02:13.840
This science is actively being built upon
five distinct pillars

39
00:02:13.840 --> 00:02:15.280
of current research.

40
00:02:15.280 --> 00:02:18.320
First, researchers build idealized
settings.

41
00:02:18.320 --> 00:02:22.400
These are perfectly solvable, scaled-down
toy models that allow scientists

42
00:02:22.400 --> 00:02:25.680
to test the math before applying it to
massive, real-world

43
00:02:25.680 --> 00:02:29.200
systems. Second, they study tractable
limits.

44
00:02:29.200 --> 00:02:33.680
By pushing mathematical boundaries to
their extreme edges, researchers isolate

45
00:02:33.680 --> 00:02:36.960
the specific phenomena driving how a
network learns.

46
00:02:36.960 --> 00:02:40.800
This graph shows chaotic data clusters
mathematically resolving into a

47
00:02:40.800 --> 00:02:44.640
single geometric curve, representing
macroscopic laws.

48
00:02:44.640 --> 00:02:50.000
Fourth, scientists isolate
hyperparameters, variables like learning rates.

49
00:02:50.280 --> 00:02:55.400
or batch sizes By mathematically
disentangling these messy variables from

50
00:02:55.400 --> 00:02:59.240
the training process, we are left with the
purest, simplest

51
00:02:59.240 --> 00:03:03.960
systems. Finally, we are identifying
universal behaviors.

52
00:03:03.960 --> 00:03:07.400
These are exact learning patterns that
occur across all neural

53
00:03:07.400 --> 00:03:11.480
networks, entirely regardless of their
specific architecture.

54
00:03:11.480 --> 00:03:15.240
Together, these five pillars prove deep
learning is driven by

55
00:03:15.240 --> 00:03:17.640
falsifiable quantitative rules.

56
00:03:17.640 --> 00:03:20.520
They take an unpredictable art and turn it
into a

57
00:03:20.520 --> 00:03:25.320
predictable machine. As this transition
from empirical observation to hard

58
00:03:25.320 --> 00:03:29.960
science takes hold, the landscape of AI
development shifts entirely.

59
00:03:29.960 --> 00:03:33.960
It opens the door for mechanistic
interpretability, the attempt to

60
00:03:33.960 --> 00:03:37.960
look directly inside hidden neural layers
and reverse engineer the

61
00:03:37.960 --> 00:03:39.960
logic of an AI's decision.

62
00:03:39.960 --> 00:03:44.360
Learning mechanics provides universal
laws, creating a framework that allows

63
00:03:44.360 --> 00:03:48.160
interpretability to focus on specific
hidden layers.

64
00:03:48.160 --> 00:03:52.480
Understanding these global rules lets us
narrow down which internal

65
00:03:52.480 --> 00:03:56.240
processes matter, using the physics of the
whole to verify

66
00:03:56.240 --> 00:04:00.000
the specific circuit responsible for a
network's choice.

67
00:04:00.000 --> 00:04:03.920
By uncovering the underlying physics of
deep learning, we move

68
00:04:03.920 --> 00:04:07.520
from empirical guesswork to mathematical
certainty.

69
00:04:07.520 --> 00:04:10.800
We are no longer just observing the output
of synthetic

70
00:04:10.800 --> 00:04:14.640
minds. We are calculating exactly how they
work.