Artificial Intelligence Explained Gently For Sleep

1
00:00:00,080 --> 00:00:03,720
Good evening and welcome back to
Sleep Wise.

2
00:00:04,400 --> 00:00:10,160
Tonight we begin at the newest
edge of an old human wish, the

3
00:00:10,160 --> 00:00:15,560
wish to build a thinking tool,
then set it down the way you set

4
00:00:15,560 --> 00:00:19,160
down.
A heavy day outside the world

5
00:00:19,160 --> 00:00:24,240
can be loud with certainty, but
here, in the soft light of

6
00:00:24,240 --> 00:00:30,000
night, we can move gently and
reverse from the present back

7
00:00:30,200 --> 00:00:33,400
through the quiet centuries that
led us here.

8
00:00:33,720 --> 00:00:37,800
Picture a modern room after
midnight, a laptop lid half

9
00:00:37,800 --> 00:00:44,440
closed, a phone face down on a
bedside table somewhere far away

10
00:00:44,880 --> 00:00:49,040
in a data center cooled by fans
and careful engineering.

11
00:00:49,720 --> 00:00:54,720
Machines are awake in a
different way, not with desire

12
00:00:54,720 --> 00:01:00,800
or intention, but with activity.
Tiny calculations ripple through

13
00:01:00,800 --> 00:01:05,480
chips like the faint hum you
barely notice until the house

14
00:01:05,480 --> 00:01:10,760
grows still.
And in many homes, people type a

15
00:01:10,760 --> 00:01:17,960
question into a simple box and
words arrive back fluent, calm,

16
00:01:18,560 --> 00:01:24,320
sometimes surprising.
In practical terms, what we call

17
00:01:24,360 --> 00:01:29,360
artificial intelligence to day
is not one single invention.

18
00:01:30,000 --> 00:01:33,480
It is an umbrella for many
methods that help computers

19
00:01:33,480 --> 00:01:37,640
recognize patterns and make
useful decisions.

20
00:01:38,440 --> 00:01:41,960
Sorting photos.
Translating languages.

21
00:01:42,480 --> 00:01:46,360
Recommending songs.
Detecting fraud.

22
00:01:46,800 --> 00:01:50,800
Adjusting a thermostat.
Steering the robot arm.

23
00:01:51,800 --> 00:01:56,240
Over the last decade, these
systems have spread because two

24
00:01:56,240 --> 00:02:02,040
things became abundant data and
the ability to compute on that

25
00:02:02,040 --> 00:02:05,240
data quickly.
When a computer learns from

26
00:02:05,240 --> 00:02:10,080
examples, it does not memorize
them the way a person remembers

27
00:02:10,560 --> 00:02:16,240
A childhood St.
Instead, it adjusts internal

28
00:02:16,240 --> 00:02:21,400
numbers so that overtime it
becomes better at map as inputs

29
00:02:21,400 --> 00:02:27,280
to outputs an image to a label,
a sentence to another sentence,

30
00:02:27,600 --> 00:02:31,800
a sound to a transcript.
You can think of this learning

31
00:02:32,120 --> 00:02:37,040
as a long practice session.
The system sees an example,

32
00:02:37,280 --> 00:02:43,480
makes a guess, measures how
wrong it was, and nudges itself

33
00:02:43,480 --> 00:02:46,960
slightly toward being less wrong
next time.

34
00:02:47,840 --> 00:02:53,120
That cycle repeats millions,
sometimes billions of times

35
00:02:53,720 --> 00:02:57,360
until the model becomes reliable
at a task.

36
00:02:58,160 --> 00:03:03,320
The models themselves can be
many shapes, decision trees,

37
00:03:03,680 --> 00:03:10,160
linear models, neural networks.
And the training can happen in

38
00:03:10,160 --> 00:03:14,640
the cloud on specialized
hardware, with enormous batches

39
00:03:14,640 --> 00:03:17,280
of examples flowing through like
a river.

40
00:03:18,080 --> 00:03:22,400
But tonight, we won't stay in
the present too long.

41
00:03:23,240 --> 00:03:28,560
We'll let it fade behind us like
city lights shrinking in a rear

42
00:03:28,560 --> 00:03:32,280
view mirror.
We'll drift backward through the

43
00:03:32,280 --> 00:03:38,040
era of smartphones and cloud
services, back through the rise

44
00:03:38,040 --> 00:03:42,880
of powerful chips, back through
earlier waves of machined

45
00:03:42,880 --> 00:03:48,160
learning, and further still to
the first dreams of mechanical

46
00:03:48,160 --> 00:03:54,200
calculation, the first symbolic
logics, the first attempts to

47
00:03:54,200 --> 00:03:58,840
measure uncertainty, and the
ancient fascination with

48
00:03:58,840 --> 00:04:05,560
automata devices that moved as
if they had a tiny life inside

49
00:04:05,560 --> 00:04:09,360
them.
For now, let your breathing find

50
00:04:09,360 --> 00:04:12,760
its own pace.
Let the idea of these bright

51
00:04:12,760 --> 00:04:17,959
machines feel distant and quiet,
like a constellation.

52
00:04:17,959 --> 00:04:22,240
You don't need the name.
We are simply walking backward

53
00:04:22,240 --> 00:04:27,920
along the path, step by step,
until the night grows older and

54
00:04:27,920 --> 00:04:32,800
simpler and the world slows.
And so do we.

55
00:04:33,200 --> 00:04:38,520
The present begins to soften at
the edges, and we step back into

56
00:04:38,520 --> 00:04:43,800
the recent decade, when AI
stopped feeling like a distant

57
00:04:43,800 --> 00:04:48,680
research word and started
feeling like a quiet utility.

58
00:04:49,760 --> 00:04:55,280
It appears in small places
first, the way night air slips

59
00:04:55,280 --> 00:05:00,000
in through a cracked window.
Your photo app groups similar

60
00:05:00,000 --> 00:05:04,400
faces.
A spam filter keeps your inbox

61
00:05:04,400 --> 00:05:11,600
from filling, A map suggests the
quickest turn before you decide.

62
00:05:12,600 --> 00:05:17,960
None of this asks for applause.
Yet together it changes the

63
00:05:17,960 --> 00:05:23,440
texture of ordinary life.
Underneath there is usually a

64
00:05:23,440 --> 00:05:29,280
trained model, a system of
adjustable parameters fitted to

65
00:05:29,280 --> 00:05:33,480
examples.
In image recognition, the input

66
00:05:33,480 --> 00:05:38,920
is a grid of pixel values and
the output is a label or a set

67
00:05:38,920 --> 00:05:44,040
of probabilities.
In speech recognition, the input

68
00:05:44,040 --> 00:05:49,240
is a waveform shaped into
features and the output is text.

69
00:05:49,800 --> 00:05:54,800
In recommendations, the input is
a history of views, clicks, and

70
00:05:54,800 --> 00:05:59,720
pauses, and the output is an
ordered list designed to be

71
00:05:59,720 --> 00:06:02,960
useful.
The mathematics differs from

72
00:06:02,960 --> 00:06:06,520
task to task, but the rhythm is
similar.

73
00:06:06,760 --> 00:06:11,560
The system is shown an example.
It produces an output, and a

74
00:06:11,560 --> 00:06:15,640
measure of error is computed by
comparing that output to a

75
00:06:15,640 --> 00:06:19,160
target.
Then the parameters are nudged

76
00:06:19,240 --> 00:06:21,520
slightly toward better
performance.

77
00:06:21,840 --> 00:06:26,640
The adjustments are small,
almost unimaginably small, but

78
00:06:26,640 --> 00:06:30,880
they add up.
After millions of examples, the

79
00:06:30,880 --> 00:06:35,720
model begins to capture patterns
that are hard to write as

80
00:06:35,720 --> 00:06:40,480
explicit rules.
It can notice that certain

81
00:06:40,480 --> 00:06:46,360
arrangements of edges often form
a face, that certain rhythms

82
00:06:46,360 --> 00:06:51,520
often form a word, that certain
sequences of choices often form

83
00:06:51,520 --> 00:06:56,320
a preference.
Importantly, these systems do

84
00:06:56,320 --> 00:07:00,000
not store a single tidy
explanation.

85
00:07:00,600 --> 00:07:07,320
They store tendencies encoded as
numbers, which you only see when

86
00:07:07,320 --> 00:07:11,880
the system makes a prediction.
Once deployed, the model meets

87
00:07:11,880 --> 00:07:15,760
the real world, and the real
world is never as neat as a

88
00:07:15,760 --> 00:07:18,720
training set.
Lighting shifts.

89
00:07:19,520 --> 00:07:23,680
Accents vary.
People phrase requests in

90
00:07:23,880 --> 00:07:28,400
unexpected ways.
Engineers watch performance,

91
00:07:28,880 --> 00:07:34,680
collect new examples and retrain
because learning systems can

92
00:07:34,680 --> 00:07:38,320
drift.
As the world drifts in this way,

93
00:07:38,880 --> 00:07:44,880
modern AI became less like a one
time invention and more like

94
00:07:44,960 --> 00:07:51,080
ongoing maintenance, a garden
that must be tended with careful

95
00:07:51,080 --> 00:07:55,080
attention to reliability and
fairness.

96
00:07:55,440 --> 00:08:02,440
This is why the last decade felt
like an acceleration 2 Resources

97
00:08:02,440 --> 00:08:05,960
became abundant at once.
First, data.

98
00:08:06,800 --> 00:08:12,040
More of life was digitized, and
everyday actions left traces

99
00:08:12,040 --> 00:08:16,040
that could be gathered into
training sets, photos,

100
00:08:16,040 --> 00:08:22,000
recordings, transactions, sensor
readings, searches, and logs.

101
00:08:22,840 --> 00:08:29,040
Second, compute chips grew
faster and more specialized, and

102
00:08:29,040 --> 00:08:33,960
cloud infrastructure made large
training runs possible without

103
00:08:33,960 --> 00:08:36,200
owning a private room of
servers.

104
00:08:36,559 --> 00:08:41,840
And now, as we continue our
backward walk, we let the

105
00:08:41,840 --> 00:08:46,440
conveniences fade into their
hidden machinery.

106
00:08:47,480 --> 00:08:52,840
We follow the path from polished
apps to the era of smartphones

107
00:08:52,840 --> 00:08:57,560
and cloud services, then further
into the hardware that made

108
00:08:57,560 --> 00:09:03,240
modern learning practical, and
further still toward older

109
00:09:03,240 --> 00:09:08,440
methods, smaller data sets, and
earlier dreams of teaching

110
00:09:08,440 --> 00:09:12,840
machines at all.
A little farther back, the story

111
00:09:12,840 --> 00:09:16,680
takes on the feel of a pocket
sized revolution.

112
00:09:17,600 --> 00:09:23,880
Before AI became a headline, the
smartphone became a habit.

113
00:09:24,280 --> 00:09:28,480
It gathered the world into a
single object you carried from

114
00:09:28,480 --> 00:09:33,840
room to room, from street to
cafe, from morning to night.

115
00:09:34,360 --> 00:09:41,720
In doing so, it also gathered
data, images, motion, location,

116
00:09:42,120 --> 00:09:46,920
touch and voice.
A phone is a small bundle of

117
00:09:46,920 --> 00:09:51,160
sensors, and sensors turned life
into numbers.

118
00:09:51,560 --> 00:09:55,400
Those numbers did not
automatically become training

119
00:09:55,400 --> 00:09:58,600
material, but they made it
possible.

120
00:09:59,600 --> 00:10:04,960
As services moved online,
organizations built systems to

121
00:10:04,960 --> 00:10:09,800
store and organize immense
volumes of information.

122
00:10:10,200 --> 00:10:15,120
They built pipelines to clean
it, label parts of it, remove

123
00:10:15,120 --> 00:10:18,760
duplicates, and handle missing
values.

124
00:10:19,760 --> 00:10:26,560
They learned to split data into
training sets and test sets so

125
00:10:26,560 --> 00:10:31,960
that a model could be evaluated
honestly, not rewarded for

126
00:10:31,960 --> 00:10:35,200
simply repeating what it had
already seen.

127
00:10:35,640 --> 00:10:39,720
They also learned, often through
public debate and internal

128
00:10:39,720 --> 00:10:44,560
friction, that privacy and
consent are not optional

129
00:10:44,560 --> 00:10:49,440
details.
Data has owners, contexts and

130
00:10:49,440 --> 00:10:53,480
consequences.
Connectivity changed the feel of

131
00:10:53,480 --> 00:10:58,200
what a device could be.
When networks became fast and

132
00:10:58,200 --> 00:11:04,000
reliable, a phone could send a
request to distant servers and

133
00:11:04,000 --> 00:11:07,800
receive a result.
In the time it takes to blink,

134
00:11:08,880 --> 00:11:14,520
intelligence began to feel like
a living layer over the device.

135
00:11:14,760 --> 00:11:19,320
Some predictions ran locally for
speed and battery life.

136
00:11:20,040 --> 00:11:25,160
Others ran in the cloud, where
larger machines could do heavier

137
00:11:25,160 --> 00:11:29,200
computation.
This division shaped modern

138
00:11:29,200 --> 00:11:34,720
products, lightweight models on
the device, and heavier training

139
00:11:34,720 --> 00:11:39,160
and experimentation on
centralized infrastructure.

140
00:11:39,560 --> 00:11:42,200
The edge mattered because of
constraints.

141
00:11:42,440 --> 00:11:47,080
Phones have limited battery,
limited heat budget, and limited

142
00:11:47,080 --> 00:11:51,680
memory.
So engineers learn to compress

143
00:11:51,680 --> 00:11:58,120
models, prune unnecessary parts,
and simplify calculations,

144
00:11:58,880 --> 00:12:03,680
making systems small enough to
run smoothly without draining a

145
00:12:03,680 --> 00:12:06,440
device.
They also learn to measure

146
00:12:06,440 --> 00:12:13,400
latency, because a model that is
accurate but slow can feel worse

147
00:12:13,400 --> 00:12:17,280
than a model that is slightly
less accurate.

148
00:12:17,800 --> 00:12:22,280
But instant app stores and
frequent updates created a new

149
00:12:22,280 --> 00:12:25,480
rhythm.
Products could change weekly,

150
00:12:25,480 --> 00:12:29,280
sometimes daily.
That encouraged constant

151
00:12:29,280 --> 00:12:32,760
evaluation.
In the real world, teams watched

152
00:12:32,760 --> 00:12:37,160
how often a feature was used,
whether it reduced errors,

153
00:12:37,400 --> 00:12:41,000
whether it increased
satisfaction, and whether it

154
00:12:41,000 --> 00:12:46,680
behaved consistently across
languages, regions and devices.

155
00:12:47,000 --> 00:12:52,000
In factual terms, this period
taught the world how to

156
00:12:52,440 --> 00:12:58,520
operationalize machine learning,
deploy models, monitor them,

157
00:12:59,000 --> 00:13:06,480
collect fresh examples, retrain
and roll improvements out again.

158
00:13:06,840 --> 00:13:13,000
Now we step backward once more.
We leave the glass screens and

159
00:13:13,000 --> 00:13:18,640
wireless signals behind, and we
approach the era when the hidden

160
00:13:18,720 --> 00:13:22,840
engines of progress became
unmistakable.

161
00:13:23,600 --> 00:13:29,080
The racks, the cooling, and the
chips built to repeat

162
00:13:29,080 --> 00:13:35,400
mathematics at enormous scale.
Now the not air cools and we

163
00:13:35,400 --> 00:13:39,000
find ourselves closer to the
machine rooms.

164
00:13:39,400 --> 00:13:43,520
Before modern products could
rely on learning systems

165
00:13:43,520 --> 00:13:49,200
everywhere, training had to
become feasible at scale.

166
00:13:49,520 --> 00:13:54,840
One of the quiet enablers was
specialized hardware, especially

167
00:13:54,840 --> 00:13:58,840
the graphics processing unit,
the GPU.

168
00:13:59,600 --> 00:14:04,120
It began as a tool for drawing
images quickly, shading,

169
00:14:04,160 --> 00:14:07,080
lighting, rendering, the
illusion of depth.

170
00:14:07,840 --> 00:14:13,120
But it turned out to be good at
something else too, repeating

171
00:14:13,120 --> 00:14:19,680
the same numerical operations
many times in parallel without

172
00:14:19,680 --> 00:14:23,080
slowing down.
Many learning methods, and

173
00:14:23,080 --> 00:14:27,480
especially neural networks, rely
on long chains of arithmetic

174
00:14:27,760 --> 00:14:31,680
that can be expressed as vast
collections of simple steps.

175
00:14:32,240 --> 00:14:39,680
Multiply, add, compare, repeat.
In earlier eras, training larger

176
00:14:39,680 --> 00:14:44,840
models on general purpose CPUs
could be painfully slow.

177
00:14:45,560 --> 00:14:49,720
But GPU's were built to handle
thousands of similar

178
00:14:49,720 --> 00:14:54,600
computations at once because
graphics demanded it.

179
00:14:55,520 --> 00:15:00,280
Researchers realized they could
borrow that speed for training.

180
00:15:01,080 --> 00:15:06,480
A run that once took weeks could
be shortened, and experiments

181
00:15:06,480 --> 00:15:11,840
that once felt impossible began
to fit into human schedules.

182
00:15:12,200 --> 00:15:16,360
Around this hardware grew the
infrastructure of the modern

183
00:15:16,360 --> 00:15:20,080
era.
Data centers, warehouse sized

184
00:15:20,080 --> 00:15:25,760
rooms filled with racks, cooling
systems, backup power, careful

185
00:15:25,760 --> 00:15:32,200
networking, and teams who think
in terms of airflow and uptime.

186
00:15:33,040 --> 00:15:37,640
These places can sound dramatic,
but they are also strangely

187
00:15:37,640 --> 00:15:40,000
patient.
They do not hurry.

188
00:15:40,640 --> 00:15:45,480
They run jobs, store results,
run the next jobs.

189
00:15:46,120 --> 00:15:51,240
They are the modern equivalent
of mills, turning raw material

190
00:15:51,240 --> 00:15:55,880
into something shaped.
Here, the raw material is

191
00:15:55,880 --> 00:16:00,240
information and the turning is
computation.

192
00:16:00,600 --> 00:16:04,480
Hardware progress is not only
about speed.

193
00:16:05,160 --> 00:16:10,400
Memory bandwidth matters because
models and data must be moved

194
00:16:10,400 --> 00:16:14,320
efficiently.
Interconnects matter because

195
00:16:14,320 --> 00:16:18,000
training can be split across
multiple machines.

196
00:16:18,880 --> 00:16:24,040
Reliability matters because long
runs are costly to interrupt.

197
00:16:24,440 --> 00:16:29,800
And energy matters because
computation becomes heat, and

198
00:16:29,800 --> 00:16:36,720
heat must be managed.
The story of AI is in part a

199
00:16:36,720 --> 00:16:41,760
story of engineering limits
being pushed gently outward,

200
00:16:42,160 --> 00:16:45,400
year by year, as we drift
backward.

201
00:16:45,760 --> 00:16:52,040
You can let the words chips and
compute become softer, like

202
00:16:52,040 --> 00:16:57,040
distant weather.
The details matter, but they do

203
00:16:57,040 --> 00:17:01,760
not ask you to stay alert.
We are simply noticing the

204
00:17:01,840 --> 00:17:06,760
engines beneath the surface and
then letting them fade as we

205
00:17:06,760 --> 00:17:11,640
step into an earlier time, when
data sets were smaller, training

206
00:17:11,640 --> 00:17:15,480
was slower, and progress
depended heavily on how

207
00:17:15,480 --> 00:17:20,920
carefully it was measured.
So we step back again toward the

208
00:17:20,920 --> 00:17:25,760
era of benchmarks and shared
yardsticks, where the field

209
00:17:25,760 --> 00:17:30,200
learned to agree quietly on what
it meant to improve.

210
00:17:30,600 --> 00:17:35,360
We arrive in a period when the
field learned to trust its

211
00:17:35,360 --> 00:17:39,320
measurements.
Progress in machine learning is

212
00:17:39,320 --> 00:17:44,320
not only about building models.
It is also about deciding how to

213
00:17:44,320 --> 00:17:48,600
judge them.
In earlier decades, researchers

214
00:17:48,600 --> 00:17:54,480
often worked on small data sets,
sometimes collected by hand, and

215
00:17:54,480 --> 00:17:57,560
compared results in scattered
ways.

216
00:17:57,880 --> 00:18:03,680
Overtime, shared benchmarks
became common, standardized sets

217
00:18:03,680 --> 00:18:07,400
of examples that many teams
could use so that improvements

218
00:18:07,400 --> 00:18:11,200
were visible, repeatable, and
comparable.

219
00:18:11,600 --> 00:18:15,520
A benchmark might be a
collection of images with

220
00:18:15,520 --> 00:18:21,320
labels, recorded speech with
transcripts, or text paired with

221
00:18:21,320 --> 00:18:25,160
answers.
The point is not that a

222
00:18:25,160 --> 00:18:30,840
benchmark represents all of
reality, it never does, but that

223
00:18:30,840 --> 00:18:36,080
it becomes a lighthouse.
If you can improve performance

224
00:18:36,080 --> 00:18:41,320
on the same set under similar
conditions, you can tell whether

225
00:18:41,320 --> 00:18:45,720
a new method is truly better or
merely different.

226
00:18:46,600 --> 00:18:49,960
This practice also made the
field more honest.

227
00:18:50,840 --> 00:18:56,040
It reduced the temptation to
claim progress without evidence,

228
00:18:56,720 --> 00:19:01,120
and it encouraged careful
evaluation across multiple

229
00:19:01,120 --> 00:19:04,680
tasks.
Along with benchmarks came the

230
00:19:04,680 --> 00:19:07,160
quieter work of data
preparation.

231
00:19:07,840 --> 00:19:13,840
Data is rarely ready as it is.
It arrives messy duplicates,

232
00:19:14,040 --> 00:19:20,720
errors, strange formats, missing
labels and biases embedded in

233
00:19:20,720 --> 00:19:23,640
what was collected and what was
ignored.

234
00:19:24,440 --> 00:19:27,760
Cleaning and curating data sets
became a craft.

235
00:19:28,480 --> 00:19:33,040
So did deciding how to split
data into training, validation

236
00:19:33,040 --> 00:19:38,680
and test sets, and how to avoid
leakage where hints from the

237
00:19:38,680 --> 00:19:44,160
test set sneak into training and
make results look better than

238
00:19:44,160 --> 00:19:47,400
they truly are?
Metrics became part of the

239
00:19:47,400 --> 00:19:52,560
shared language.
Accuracy, precision, recall,

240
00:19:53,040 --> 00:19:58,160
error rates, and other measures
help teams describe performance

241
00:19:58,160 --> 00:20:02,200
clearly.
Some tasks needed new metrics

242
00:20:02,480 --> 00:20:06,920
and some needed human evaluation
especially.

243
00:20:07,280 --> 00:20:11,200
When outputs were subjective or
when correctness was hard to

244
00:20:11,200 --> 00:20:17,640
define, but the impulse was the
same, create a common map of

245
00:20:17,640 --> 00:20:20,920
progress.
Tonight you can imagine these

246
00:20:20,920 --> 00:20:24,880
data sets as quiet piles of
photographs, boxes of

247
00:20:24,880 --> 00:20:28,480
recordings, and stacks of
printed pages.

248
00:20:29,000 --> 00:20:33,080
Not as stories you must read,
but as raw material.

249
00:20:34,360 --> 00:20:39,320
The work of building benchmarks
is not glamorous, yet it shapes

250
00:20:39,320 --> 00:20:45,280
what gets built, because what
gets measured gets optimized.

251
00:20:46,160 --> 00:20:52,200
And as we step backward again,
those piles become smaller, the

252
00:20:52,200 --> 00:20:56,880
stacks thinner, and the
experiments more dependent on

253
00:20:56,880 --> 00:21:00,920
handcrafted features and
classical methods.

254
00:21:01,280 --> 00:21:05,840
We are approaching the era when
machine learning often meant

255
00:21:05,840 --> 00:21:11,920
carefully chosen inputs, simpler
models, and a different kind of

256
00:21:11,920 --> 00:21:15,840
patience.
Now we step into the time when

257
00:21:16,080 --> 00:21:22,160
machine learning often meant a
narrower set of tools and a

258
00:21:22,160 --> 00:21:27,640
different balance between human
design and automated learning.

259
00:21:28,040 --> 00:21:32,720
Before deep neural networks
dominated many applications,

260
00:21:33,520 --> 00:21:38,560
practitioners frequently relied
on models like logistic

261
00:21:38,560 --> 00:21:45,440
regression, decision trees,
random forests, and support

262
00:21:45,440 --> 00:21:49,240
vector machines.
These methods could be powerful,

263
00:21:49,720 --> 00:21:54,960
but they usually depended on
features, carefully crafted

264
00:21:54,960 --> 00:22:00,520
representations of the input.
In a vision task, features might

265
00:22:00,520 --> 00:22:05,360
describe edges, corners,
textures, or color histograms.

266
00:22:05,960 --> 00:22:10,600
In speech, features might
capture frequency patterns over

267
00:22:10,600 --> 00:22:14,960
time.
In text, features might include

268
00:22:14,960 --> 00:22:17,560
counts of words or short
phrases.

269
00:22:18,360 --> 00:22:23,280
Creating these representations
required domain knowledge and

270
00:22:23,280 --> 00:22:27,480
experimentation.
The model then learned on top of

271
00:22:27,480 --> 00:22:31,240
those features, adjusting its
parameters to separate

272
00:22:31,240 --> 00:22:38,040
categories or predict outcomes.
In this era, a large portion of

273
00:22:38,040 --> 00:22:42,040
the intelligence was in the
pipeline, the preprocessing, the

274
00:22:42,040 --> 00:22:46,520
feature engineering, the careful
selection of signals that

275
00:22:46,520 --> 00:22:49,920
mattered.
These approaches were not crude.

276
00:22:50,440 --> 00:22:54,560
They were often elegant and
statistically grounded.

277
00:22:55,240 --> 00:22:58,680
They taught the field how to
think in terms of

278
00:22:59,080 --> 00:23:04,000
generalization.
A model should perform well not

279
00:23:04,000 --> 00:23:09,240
only on the data it is seen, but
also on new data drawn from the

280
00:23:09,240 --> 00:23:14,560
same underlying process.
They encouraged regularization

281
00:23:15,000 --> 00:23:18,840
methods to avoid over fitting by
limiting complexity.

282
00:23:19,720 --> 00:23:23,760
They encouraged cross
validation, testing models

283
00:23:23,760 --> 00:23:28,320
across different splits of data
to estimate robustness.

284
00:23:29,200 --> 00:23:34,320
And they encouraged humility
because small changes in the

285
00:23:34,320 --> 00:23:40,880
data could change results.
Computing constraints shaped the

286
00:23:40,880 --> 00:23:45,280
style of work.
Training had to be efficient,

287
00:23:45,640 --> 00:23:51,000
and memory had to be respected.
That led to an emphasis on

288
00:23:51,000 --> 00:23:55,360
convex optimization in some
settings, problems where a

289
00:23:55,360 --> 00:23:59,800
single best solution could be
found reliably.

290
00:24:00,640 --> 00:24:05,400
It also led to practical
heuristics and to a culture of

291
00:24:05,400 --> 00:24:09,360
careful baselines.
Compare your new idea to

292
00:24:09,360 --> 00:24:12,520
something simple and strong
before celebrating.

293
00:24:12,880 --> 00:24:17,920
As we drift backward, we can
feel the field in this period as

294
00:24:17,920 --> 00:24:23,560
a workshop of careful craft.
Less spectacle, more tuning,

295
00:24:24,080 --> 00:24:29,040
less scale, more thought about
what to include and what to

296
00:24:29,040 --> 00:24:33,000
ignore.
And yet, beneath the surface,

297
00:24:33,400 --> 00:24:36,280
older neural ideas never
vanished.

298
00:24:36,840 --> 00:24:41,280
They waited quietly for the
conditions that would let them

299
00:24:41,280 --> 00:24:46,360
return.
So we step back again toward the

300
00:24:46,360 --> 00:24:52,120
years when neural networks began
to re emerge, then further to

301
00:24:52,120 --> 00:24:57,200
the techniques that made them
trainable, and to the earlier

302
00:24:57,200 --> 00:25:02,040
hopes and disappointments that
shaped the field's memory.

303
00:25:02,440 --> 00:25:07,080
We move backward into the early
2000 tens when a familiar

304
00:25:07,080 --> 00:25:12,840
approach, neural networks, began
to feel newly capable.

305
00:25:13,200 --> 00:25:18,360
Neural networks had existed for
decades, sometimes celebrated,

306
00:25:18,640 --> 00:25:24,120
sometimes dismissed, often
emphasized in theory and limited

307
00:25:24,360 --> 00:25:28,680
in practice.
But as larger data sets became

308
00:25:28,680 --> 00:25:34,320
available and faster hardware
became common, deeper networks,

309
00:25:34,880 --> 00:25:39,840
networks with many layers
started to perform strikingly

310
00:25:39,840 --> 00:25:44,680
well on tasks like image
recognition and speech.

311
00:25:45,040 --> 00:25:50,040
In simple terms, a neural
network learns by adjusting

312
00:25:50,040 --> 00:25:55,520
weights, numbers that determine
how strongly 1 unit influences

313
00:25:55,520 --> 00:26:00,120
another during training, the
network processes.

314
00:26:00,120 --> 00:26:05,880
An example produces an output,
compares it to a target, and

315
00:26:05,880 --> 00:26:10,480
then uses an error signal to
nudge many weights in small

316
00:26:10,480 --> 00:26:14,320
increments.
The key is that this error

317
00:26:14,320 --> 00:26:19,080
signal can be propagated
backward through the layers so

318
00:26:19,080 --> 00:26:25,160
that each weight is updated in
the direction that reduces

319
00:26:25,160 --> 00:26:28,760
error.
It is a method of learning by

320
00:26:28,760 --> 00:26:35,280
tiny corrections repeated
countless times, like smoothing

321
00:26:35,280 --> 00:26:41,080
a stone by steady water.
This period also sharpened

322
00:26:41,080 --> 00:26:46,320
practical techniques, better
ways to initialize weights,

323
00:26:46,960 --> 00:26:51,840
better activation functions,
better ways to prevent

324
00:26:51,840 --> 00:26:57,360
overfitting, and better
optimization methods that made

325
00:26:57,360 --> 00:27:01,640
training stable.
Researchers learned to use large

326
00:27:01,640 --> 00:27:07,920
batches of data to normalize
inputs, to manage exploding or

327
00:27:07,920 --> 00:27:13,560
vanishing gradients, and to
select architectures suited to

328
00:27:13,560 --> 00:27:20,200
the structure of the problem.
The results felt like a hinge.

329
00:27:21,120 --> 00:27:25,800
Systems that once seemed brittle
began to feel practical.

330
00:27:26,040 --> 00:27:30,800
At the same time, costs and
constraints became part of the

331
00:27:30,800 --> 00:27:34,800
conversation.
Larger models demanded more

332
00:27:34,800 --> 00:27:41,320
energy and more compute.
Access was uneven, and data

333
00:27:41,320 --> 00:27:46,360
reflected the world as it is
uneven, imperfect, sometimes

334
00:27:46,360 --> 00:27:50,120
unfair.
So models could inherit those

335
00:27:50,120 --> 00:27:56,080
patterns unless care was taken.
The field began to broaden its

336
00:27:56,080 --> 00:28:02,600
view, not only asking does it
work, but also for whom and

337
00:28:02,600 --> 00:28:07,480
under what conditions.
Tonight we do not need to solve

338
00:28:07,480 --> 00:28:12,320
those questions.
We only need to notice the shape

339
00:28:12,480 --> 00:28:17,600
of the era, an old idea
returning with new strength,

340
00:28:18,480 --> 00:28:24,840
powered by compute, guided by
benchmarks, and made practical

341
00:28:25,520 --> 00:28:30,560
by patient iteration.
From here, the path continues

342
00:28:30,560 --> 00:28:36,560
backward into the core training
technique itself, how learning

343
00:28:36,840 --> 00:28:41,720
is turned into mathematics, and
into earlier decades.

344
00:28:41,720 --> 00:28:48,240
When optimism cooled into long
winters, we step back again

345
00:28:48,720 --> 00:28:53,200
toward the engine of training
and the history that gathered

346
00:28:53,200 --> 00:28:56,600
around it.
Now we step back to the

347
00:28:56,600 --> 00:29:01,400
mechanism that quietly underlies
much of modern learning,

348
00:29:02,200 --> 00:29:05,680
optimization.
It is a simple word for a deep

349
00:29:05,680 --> 00:29:10,000
habit, adjusting something until
it fits.

350
00:29:11,080 --> 00:29:15,080
In machine learning, the
something is a set of

351
00:29:15,080 --> 00:29:21,000
parameters, and the fit is
measured by a loss function, a

352
00:29:21,000 --> 00:29:25,880
numerical score that represents
how far a model's outputs are

353
00:29:25,880 --> 00:29:29,240
from the targets we want.
Training is the process of

354
00:29:29,240 --> 00:29:33,360
reducing that loss.
One of the central ideas is

355
00:29:33,360 --> 00:29:38,280
gradient descent.
Imagine you're on a landscape at

356
00:29:38,280 --> 00:29:43,480
night, and you cannot see the
whole terrain, but you can sense

357
00:29:43,480 --> 00:29:47,200
which direction slopes downward
from where you stand.

358
00:29:47,920 --> 00:29:53,160
You take a small step down, then
another, then another.

359
00:29:53,800 --> 00:30:00,400
Each step is tiny because large
steps might overshoot overtime.

360
00:30:00,800 --> 00:30:05,320
You descend toward a valley,
toward a set of parameters that

361
00:30:05,320 --> 00:30:10,200
produces better predictions.
This is not guaranteed to find

362
00:30:10,200 --> 00:30:15,240
the absolute best solution in
every case, but it is often good

363
00:30:15,240 --> 00:30:19,600
enough and remarkably effective
at scale.

364
00:30:20,040 --> 00:30:24,800
For neural networks, the
practical question was how to

365
00:30:24,800 --> 00:30:29,760
compute gradients efficiently
when there are many layers.

366
00:30:30,520 --> 00:30:34,640
The answer is a method that
computes the contribution of

367
00:30:34,680 --> 00:30:39,400
each parameter to the loss by
working backward through the

368
00:30:39,400 --> 00:30:43,480
network, reusing intermediate
computations.

369
00:30:44,080 --> 00:30:48,440
This makes training feasible
even when there are millions of

370
00:30:48,440 --> 00:30:52,760
parameters.
Without this efficiency, deep

371
00:30:52,760 --> 00:30:58,000
networks would be too slow to
train and too expensive to

372
00:30:58,000 --> 00:31:01,560
experiment with.
Optimization also requires

373
00:31:01,560 --> 00:31:07,840
choices, learning rate, batch
size, momentum, regularization.

374
00:31:08,560 --> 00:31:11,640
Each one shapes how training
behaves.

375
00:31:12,680 --> 00:31:16,920
Too fast and the model may
bounce around without settling.

376
00:31:17,800 --> 00:31:22,200
Too slow and training may take
impractically long.

377
00:31:23,040 --> 00:31:27,280
Too flexible and the model may
memorize quirks of the data.

378
00:31:28,160 --> 00:31:32,480
Too constrained and it may never
learn enough.

379
00:31:33,560 --> 00:31:38,400
Much of machine learning
practice is the calm, repetitive

380
00:31:38,400 --> 00:31:43,560
tuning of these settings, guided
by validation, results and

381
00:31:43,560 --> 00:31:47,680
experience.
As we drift backward, this world

382
00:31:47,680 --> 00:31:53,080
of gradients and losses begins
to thin into earlier eras.

383
00:31:53,840 --> 00:31:57,960
We will soon reach the decades
when the field leaned heavily on

384
00:31:57,960 --> 00:32:03,320
hand crafted features and
simpler models, and then further

385
00:32:03,320 --> 00:32:09,640
back into a time of bold early
promises and sharp limitations,

386
00:32:10,120 --> 00:32:14,360
when neural ideas were
introduced, challenged, and

387
00:32:14,360 --> 00:32:19,400
sometimes set aside.
So we step back again, toward

388
00:32:19,400 --> 00:32:24,560
those earlier hopes, and toward
the long, quiet winters that

389
00:32:24,560 --> 00:32:30,240
taught researchers patience.
We step backward into the

390
00:32:30,240 --> 00:32:37,200
periods often called AI winters,
when enthusiasm cooled and

391
00:32:37,200 --> 00:32:43,080
funding became harder to find.
These winters were not times

392
00:32:43,080 --> 00:32:47,360
when nothing happened.
They were times when progress

393
00:32:47,360 --> 00:32:51,480
was slower and expectations were
corrected.

394
00:32:52,480 --> 00:32:57,160
Early systems had promised a
great deal, but they struggled

395
00:32:57,160 --> 00:33:00,880
in the wild.
They required too much hand

396
00:33:00,880 --> 00:33:04,320
tuning.
They failed when the environment

397
00:33:04,320 --> 00:33:08,240
changed.
They could not easily scale to

398
00:33:08,240 --> 00:33:11,800
the full messiness of human
language or perception.

399
00:33:12,200 --> 00:33:15,800
One reason was that early
approaches often leaned on

400
00:33:15,800 --> 00:33:19,960
explicit rules.
If you can write down the rules

401
00:33:19,960 --> 00:33:23,560
of a domain, a computer can
follow them quickly.

402
00:33:24,240 --> 00:33:28,760
But the world is full of
exceptions, edge cases, and

403
00:33:28,760 --> 00:33:34,280
tacit knowledge.
Rules multiply, then collide.

404
00:33:34,840 --> 00:33:37,840
Maintaining them becomes its own
problem.

405
00:33:38,560 --> 00:33:43,000
Another reason was compute.
Many methods that looked good on

406
00:33:43,000 --> 00:33:48,400
paper or on small toy problems
became infeasible when data grew

407
00:33:48,680 --> 00:33:54,240
and tasks became realistic.
A third reason was data itself.

408
00:33:54,920 --> 00:33:59,120
Many domains simply did not have
large, well labeled data sets

409
00:33:59,120 --> 00:34:02,600
available.
In these quieter years, a

410
00:34:02,600 --> 00:34:05,240
different kind of work
continued.

411
00:34:06,040 --> 00:34:09,679
Researchers refined statistical
methods.

412
00:34:10,480 --> 00:34:13,600
They developed better evaluation
practices.

413
00:34:14,400 --> 00:34:17,199
They built data sets where they
could.

414
00:34:18,000 --> 00:34:23,000
They explored learning in
constrained settings, they

415
00:34:23,000 --> 00:34:28,120
improved algorithms for search
and planning in specific

416
00:34:28,120 --> 00:34:33,719
domains, and they learned the
value of bass lines and

417
00:34:33,719 --> 00:34:38,320
reproducibility.
The field matured even when it

418
00:34:38,320 --> 00:34:42,080
was not in fashion.
It is also in these eras that

419
00:34:42,080 --> 00:34:45,000
you can feel the rhythm of
science more broadly.

420
00:34:45,679 --> 00:34:49,280
A new idea arrives.
Bright and confident.

421
00:34:50,000 --> 00:34:56,080
Reality pushes back.
The idea is reshaped or set

422
00:34:56,080 --> 00:34:59,280
aside or combined with other
ideas.

423
00:35:00,000 --> 00:35:04,480
Sometimes it returns later, when
the conditions are right.

424
00:35:05,120 --> 00:35:07,680
This is not failure in the
dramatic sense.

425
00:35:07,680 --> 00:35:13,640
It is endurance, the slow
tightening of a craft as we

426
00:35:13,640 --> 00:35:17,680
drift backward.
We can imagine the laboratories

427
00:35:17,680 --> 00:35:22,760
of that time, fluorescent
lights, whiteboards, modest

428
00:35:22,760 --> 00:35:27,960
machines, small data sets and
long patience.

429
00:35:28,600 --> 00:35:32,080
The world outside might have
moved on to other obsessions,

430
00:35:32,480 --> 00:35:38,280
but inside the work continued
measuring, testing, revising,

431
00:35:39,000 --> 00:35:44,480
and further back still, we find
the early neural inspirations

432
00:35:44,480 --> 00:35:50,720
that sparked both optimism and
disappointment, perceptrons,

433
00:35:51,120 --> 00:35:56,120
simple learning rules, and the
first attempts to mimic the

434
00:35:56,120 --> 00:36:01,080
brain with mathematics.
So we step back again toward the

435
00:36:01,080 --> 00:36:06,600
beginning of learning machines
and toward the early decades of

436
00:36:06,600 --> 00:36:12,280
computation that made such
dreams imaginable.

437
00:36:12,640 --> 00:36:17,320
Now we are closer to the
beginning of the learning

438
00:36:17,320 --> 00:36:21,800
machine dream.
In the mid 20th century,

439
00:36:22,280 --> 00:36:26,160
researchers began proposing
systems that could adjust

440
00:36:26,160 --> 00:36:32,440
themselves based on experience.
One early family of ideas was

441
00:36:32,440 --> 00:36:38,240
the perceptron, a simple model
that takes inputs, multiplies

442
00:36:38,240 --> 00:36:44,400
them by weights, sums them, and
produces an output based on

443
00:36:44,400 --> 00:36:46,800
whether the sum crosses a
threshold.

444
00:36:47,640 --> 00:36:51,400
It is modest by modern
standards, yet it carries an

445
00:36:51,400 --> 00:36:56,280
important promise.
The weights can be learned from

446
00:36:56,280 --> 00:37:00,120
examples.
If a system can change itself in

447
00:37:00,120 --> 00:37:04,560
response to data, then
intelligence might be something

448
00:37:04,560 --> 00:37:09,400
you can train rather than
something you must hand code.

449
00:37:10,080 --> 00:37:14,800
That was the hope, and
perceptrons could learn some

450
00:37:14,800 --> 00:37:19,280
kinds of patterns, especially
patterns that can be separated

451
00:37:19,280 --> 00:37:24,960
cleanly in a mathematical sense.
But they also had clear

452
00:37:24,960 --> 00:37:28,640
limitations.
Some seemingly simple patterns

453
00:37:28,920 --> 00:37:32,520
cannot be learned by a single
layer perception.

454
00:37:33,480 --> 00:37:39,480
This led to criticism,
disillusionment, and a shift of

455
00:37:39,480 --> 00:37:42,320
attention toward other
approaches.

456
00:37:42,600 --> 00:37:47,960
Still, the perceptron era
mattered deeply.

457
00:37:48,760 --> 00:37:53,480
It introduced the language of
weights and learning rules.

458
00:37:54,520 --> 00:37:59,240
It encouraged thinking about
generalization and noise.

459
00:37:59,520 --> 00:38:02,400
It raised the question of
representation.

460
00:38:02,760 --> 00:38:07,920
What form should inputs take so
that a learning system can use

461
00:38:07,920 --> 00:38:11,120
them?
And it forced researchers to

462
00:38:11,120 --> 00:38:16,680
confront the gap between
appealing metaphors of neurons

463
00:38:17,280 --> 00:38:21,120
and the hard requirements of
mathematics and data.

464
00:38:21,560 --> 00:38:25,960
Around the same time, the world
of computation was expanding

465
00:38:25,960 --> 00:38:28,880
rapidly.
Early electronic computers were

466
00:38:28,880 --> 00:38:33,360
being built and improved storage
became more reliable.

467
00:38:33,960 --> 00:38:39,440
Programming languages emerged.
The notion of a stored program,

468
00:38:39,760 --> 00:38:45,480
data and instructions living in
memory made computers flexible

469
00:38:45,680 --> 00:38:50,600
in a way that earlier
calculating devices were not.

470
00:38:51,680 --> 00:38:56,840
Problems could be expressed as
procedures, and procedures could

471
00:38:56,840 --> 00:39:00,520
be changed without rebuilding
the machine.

472
00:39:00,760 --> 00:39:06,680
As we drift backward from this
point, we begin to leave AI as a

473
00:39:06,680 --> 00:39:12,480
named field and return to the
broader foundations, computation

474
00:39:12,480 --> 00:39:17,800
itself, information as a
measurable quantity, and the

475
00:39:17,800 --> 00:39:24,160
mathematics of uncertainty.
We will pass through war time

476
00:39:24,160 --> 00:39:29,320
and post war engineering through
the careful articulation of what

477
00:39:29,320 --> 00:39:36,560
a universal machine means and
into the earlier centuries where

478
00:39:36,560 --> 00:39:41,080
logic and calculation were first
formalized.

479
00:39:41,440 --> 00:39:46,520
For now, let the earliest
learning machines feel like

480
00:39:46,520 --> 00:39:53,120
small lanterns, not spotlights.
They did not solve everything,

481
00:39:53,840 --> 00:39:58,680
but they lit a path, and we
continue stepping backward

482
00:39:59,080 --> 00:40:02,920
toward the deeper roots of the
idea that thought can be

483
00:40:02,920 --> 00:40:06,680
mechanized.
We stepped back into the era

484
00:40:06,880 --> 00:40:12,720
when computation itself was
being clarified, not as a single

485
00:40:12,720 --> 00:40:18,400
device, but as a concept.
Early electronic machines were

486
00:40:18,400 --> 00:40:21,840
powerful.
But the deeper breakthrough was

487
00:40:21,840 --> 00:40:25,760
the realization that many
different tasks could be

488
00:40:25,760 --> 00:40:28,920
expressed in the same abstract
language.

489
00:40:29,520 --> 00:40:36,800
Symbols manipulated by rules.
A computation is a procedure.

490
00:40:37,160 --> 00:40:42,640
A procedure can be written down,
and if it can be written down

491
00:40:42,640 --> 00:40:47,080
precisely, a machine can carry
it out.

492
00:40:47,560 --> 00:40:52,520
This is where the idea of a
universal machine becomes so

493
00:40:52,520 --> 00:40:56,680
important.
Rather than building a separate

494
00:40:56,680 --> 00:41:02,240
device for every task, you could
build 1 machine capable of

495
00:41:02,240 --> 00:41:06,440
reading a description of any
task and performing it.

496
00:41:06,800 --> 00:41:12,480
In practice this meant storing
instructions in memory alongside

497
00:41:12,480 --> 00:41:18,720
data so the machine could be
reconfigured by software.

498
00:41:19,600 --> 00:41:23,880
This is one reason computers
became general purpose tools.

499
00:41:24,440 --> 00:41:28,520
They were no longer just
calculators, they were engines

500
00:41:28,600 --> 00:41:33,520
for procedures.
Alongside computation grew

501
00:41:33,520 --> 00:41:38,400
information theory, a way to
quantify messages.

502
00:41:39,280 --> 00:41:44,640
If you treat communication as
signals that can be noisy, then

503
00:41:44,640 --> 00:41:49,920
you can ask how much information
is contained in a message, how

504
00:41:49,920 --> 00:41:55,720
efficiently it can be
compressed, and how reliably it

505
00:41:55,720 --> 00:42:00,480
can be transmitted.
This is not AI in the modern

506
00:42:00,480 --> 00:42:04,360
sense, but it becomes part of
the foundation.

507
00:42:05,160 --> 00:42:09,520
Compression and prediction are
closely linked because to

508
00:42:09,520 --> 00:42:12,560
compress well, you must capture
patterns.

509
00:42:13,080 --> 00:42:18,160
To capture patterns is, in a
sense, to predict, and beneath

510
00:42:18,160 --> 00:42:23,880
both computation and information
lies probability, the

511
00:42:23,880 --> 00:42:30,000
mathematics of uncertainty.
Long before computers, thinkers

512
00:42:30,000 --> 00:42:35,040
developed ways to reason about
chance and incomplete knowledge.

513
00:42:35,600 --> 00:42:40,400
They built tools for estimating
what is likely and for updating

514
00:42:40,400 --> 00:42:43,240
beliefs when new evidence
arrives.

515
00:42:44,440 --> 00:42:50,360
These tools, when combined with
computers, became statistical

516
00:42:50,360 --> 00:42:56,680
learning systems that do not
claim certainty but offer well

517
00:42:56,680 --> 00:43:01,400
calibrated guesses.
As we drift backward, the modern

518
00:43:01,400 --> 00:43:07,080
machinery becomes quieter.
We are leaving behind racks of

519
00:43:07,080 --> 00:43:11,960
servers and entering older
rooms, chalkboards, mechanical

520
00:43:11,960 --> 00:43:16,600
calculators and the early formal
languages of logic.

521
00:43:17,520 --> 00:43:23,000
Soon we will reach the 19th
century's dreams of programmable

522
00:43:23,000 --> 00:43:28,440
machines, then earlier centuries
where logic was turned into

523
00:43:28,520 --> 00:43:33,360
algebra and uncertainty was
given numbers for to night.

524
00:43:33,760 --> 00:43:40,040
Notice how gentle this chain is.
Each step is a small

525
00:43:40,040 --> 00:43:44,520
clarification.
What is a procedure?

526
00:43:44,800 --> 00:43:48,440
What is a message?
What is uncertainty?

527
00:43:49,640 --> 00:43:53,960
The world slows and the
questions become simpler.

528
00:43:54,400 --> 00:43:59,440
And in that simplicity, there is
a kind of rest.

529
00:43:59,760 --> 00:44:05,800
We continue backward toward the
older crafts that prepared the

530
00:44:05,800 --> 00:44:08,480
ground for everything that
followed.

531
00:44:08,960 --> 00:44:14,880
We step back to a time when the
question of intelligence felt

532
00:44:14,880 --> 00:44:20,520
less like a product feature and
more like a thought you could

533
00:44:20,520 --> 00:44:26,720
hold in your hand, turning it
slowly, seeing new edges in the

534
00:44:26,720 --> 00:44:29,760
lamp light.
In the middle of the 20th

535
00:44:29,760 --> 00:44:35,040
century, as electronics and
mathematics began to meet, Alan

536
00:44:35,040 --> 00:44:39,280
Turing asked a simple,
unsettling question.

537
00:44:40,320 --> 00:44:46,080
If a machine can carry out any
well defined procedure, how

538
00:44:46,080 --> 00:44:49,160
would we decide whether it is
thinking?

539
00:44:50,240 --> 00:44:55,320
He proposed an imitation game as
a practical lens, because we

540
00:44:55,320 --> 00:45:00,680
only ever judge a mind from the
outside by the words and actions

541
00:45:00,680 --> 00:45:04,440
it offers us.
Behind that question sit

542
00:45:04,440 --> 00:45:07,600
something even more
foundational.

543
00:45:08,200 --> 00:45:14,320
The idea of a universal machine.
Instead of building a separate

544
00:45:14,320 --> 00:45:19,480
device for every task, you could
build 1 machine that reads

545
00:45:19,480 --> 00:45:23,800
instructions and follows them
whatever those instructions

546
00:45:23,800 --> 00:45:28,200
describe.
In practical terms, that means a

547
00:45:28,200 --> 00:45:34,000
program can be stored, copied,
edited and run again.

548
00:45:34,480 --> 00:45:39,040
The machine becomes a general
performer of procedures, and the

549
00:45:39,040 --> 00:45:43,560
procedures become an invisible
architecture that can change

550
00:45:43,840 --> 00:45:46,440
without changing the physical
hardware.

551
00:45:47,240 --> 00:45:53,440
This is the quiet pivot from a
calculator to a computer.

552
00:45:53,840 --> 00:45:59,080
There is a sleepy clarity in
that a computer is a device for

553
00:45:59,240 --> 00:46:03,440
symbol manipulation.
It moves marks according to

554
00:46:03,440 --> 00:46:07,400
rules.
Yet that simplicity is powerful

555
00:46:07,840 --> 00:46:12,720
because so much of life can be
expressed as rules operating on

556
00:46:12,720 --> 00:46:19,120
symbols, arithmetic, accounting,
navigation, tables, scheduling,

557
00:46:19,560 --> 00:46:23,720
and code breaking.
Later, those same machines would

558
00:46:23,720 --> 00:46:27,240
be used to tune learning systems
from examples.

559
00:46:27,960 --> 00:46:34,360
But the first step is simpler,
to show that doing can be

560
00:46:34,360 --> 00:46:39,960
formalized, that a method can be
written down South precisely,

561
00:46:40,280 --> 00:46:42,760
that it runs without
interpretation.

562
00:46:43,120 --> 00:46:47,280
Still, the night invites A
softer distinction.

563
00:46:47,880 --> 00:46:52,040
Repeating a procedure is not the
same as understanding it.

564
00:46:52,920 --> 00:46:56,640
Turing did not claim that
machines already possessed an

565
00:46:56,640 --> 00:47:00,480
inner life.
He asked us to be honest about

566
00:47:00,480 --> 00:47:04,400
what we mean by mind when we
judge it.

567
00:47:04,760 --> 00:47:09,800
If a system produces the same
outward behavior as a thinking

568
00:47:09,800 --> 00:47:14,520
person in a particular setting,
where do we draw the line?

569
00:47:15,240 --> 00:47:22,160
Is intelligence an inner glow or
a pattern of responses shaped by

570
00:47:22,160 --> 00:47:28,920
context, memory and learning?
As we drift backward, picture

571
00:47:28,920 --> 00:47:35,040
paper stacked in neat piles, the
slow scratch of chalk, and a

572
00:47:35,040 --> 00:47:41,320
world realizing that procedures
can be written, stored and

573
00:47:41,320 --> 00:47:46,640
carried out at speed, We are
still in the age of circuits and

574
00:47:46,640 --> 00:47:50,760
urgency, and that is where our
steps carry us next.

575
00:47:51,000 --> 00:47:57,000
Softly, without hurry, we move
back into the war time and post

576
00:47:57,000 --> 00:48:02,240
war years when computation was
pulled forward by necessity.

577
00:48:02,600 --> 00:48:07,800
The story is often told with
grand language, but in the quiet

578
00:48:07,800 --> 00:48:12,520
truth of it, much of this era
was about logistics and

579
00:48:12,520 --> 00:48:15,920
patience.
Messages that needed to be

580
00:48:15,920 --> 00:48:20,160
decoded, trajectories that
needed to be calculated,

581
00:48:20,640 --> 00:48:25,080
supplies that needed to be
routed, patterns that needed to

582
00:48:25,080 --> 00:48:30,080
be detected in noise.
Human beings did what they could

583
00:48:30,080 --> 00:48:35,840
by hand, with tables and teams,
but speed mattered and error

584
00:48:35,840 --> 00:48:39,800
mattered, and so machines were
built to help.

585
00:48:40,160 --> 00:48:45,560
Early electronic computers were
large, hot, and temperamental.

586
00:48:46,080 --> 00:48:51,000
They filled rooms, demanded
careful maintenance, and could

587
00:48:51,000 --> 00:48:57,040
fail in ways that felt almost
biological, sudden, stubborn,

588
00:48:57,520 --> 00:49:01,320
unpredictable.
Yet they offered something

589
00:49:01,320 --> 00:49:06,680
precious Repeatability.
Once a procedure was set up

590
00:49:06,680 --> 00:49:12,680
correctly, the machine could
perform it over and over without

591
00:49:12,680 --> 00:49:18,040
growing tired, without losing
its place, without drifting into

592
00:49:18,040 --> 00:49:23,320
distraction.
This repeatability became a kind

593
00:49:23,320 --> 00:49:27,680
of trust.
A key practical shift was the

594
00:49:27,680 --> 00:49:33,000
stored program idea,
instructions and data living

595
00:49:33,000 --> 00:49:37,080
together in memory.
When you can store a program,

596
00:49:37,440 --> 00:49:41,040
you can change it without
rewiring a machine.

597
00:49:41,800 --> 00:49:45,520
Software becomes a new kind of
craft.

598
00:49:45,960 --> 00:49:52,440
It also becomes a new kind of
vulnerability because bugs are

599
00:49:52,440 --> 00:49:55,720
not mechanical, they are
conceptual.

600
00:49:56,560 --> 00:50:01,800
A single misplaced step can send
an entire calculation quietly

601
00:50:01,800 --> 00:50:05,640
off course.
This pushed early programmers

602
00:50:05,640 --> 00:50:12,200
toward discipline, checking,
testing, documenting, building

603
00:50:12,200 --> 00:50:15,440
confidence through careful
repetition.

604
00:50:15,760 --> 00:50:20,520
And beneath the engineering ran
a philosophical undercurrent.

605
00:50:21,280 --> 00:50:25,560
If a machine can follow a
procedure, Florida State, then

606
00:50:25,560 --> 00:50:31,800
the boundary between calculation
and reasoning begins to blur,

607
00:50:32,600 --> 00:50:38,720
not because a machine feels, but
because many tasks we once

608
00:50:38,720 --> 00:50:42,760
thought required intelligence
can be broken into steps.

609
00:50:43,080 --> 00:50:48,360
The more steps we can specify,
the more we can delegate.

610
00:50:49,360 --> 00:50:53,640
This does not reduce human
thought to mere procedure, but

611
00:50:53,640 --> 00:50:58,520
it does reveal that procedure
can go surprisingly far.

612
00:50:58,880 --> 00:51:04,160
As we drift backward, we feel
this era like a corridor between

613
00:51:04,160 --> 00:51:08,200
worlds.
On one side, modern computing,

614
00:51:08,600 --> 00:51:13,720
flexible and general.
On the other, older traditions

615
00:51:13,720 --> 00:51:18,240
of measurement and table making,
where humans built knowledge

616
00:51:18,240 --> 00:51:22,760
into lists, charts, and hand
computed reference books.

617
00:51:23,160 --> 00:51:28,080
We'll keep stepping away from
humming rooms of electronics

618
00:51:28,600 --> 00:51:32,240
toward the quieter idea that
information itself can be

619
00:51:32,240 --> 00:51:37,640
counted, and that messages can
be treated like measurable

620
00:51:37,640 --> 00:51:42,000
things.
Now we step back into a moment

621
00:51:42,560 --> 00:51:47,480
when information became
something you could measure with

622
00:51:47,480 --> 00:51:52,160
the same calm certainty as
weight or distance.

623
00:51:53,040 --> 00:51:57,560
Claude Shannon, working in the
middle of the 20th century,

624
00:51:58,160 --> 00:52:03,520
showed that messages can be
treated mathematically.

625
00:52:03,840 --> 00:52:10,040
A message in this view is not
defined by its meaning, but by

626
00:52:10,040 --> 00:52:14,640
its structure, by the set of
possible signals, and how

627
00:52:14,640 --> 00:52:17,760
surprising a particular signal
is.

628
00:52:18,320 --> 00:52:22,400
The more surprising, the more
information it carries.

629
00:52:23,280 --> 00:52:29,400
This is a strange idea at 1st,
and then it becomes beautifully

630
00:52:29,400 --> 00:52:32,600
simple.
With that shift came a new way

631
00:52:32,600 --> 00:52:37,720
to think about noise.
Every channel, wire, radio wave,

632
00:52:37,720 --> 00:52:44,960
paper, air adds distortions.
Information theory asks how much

633
00:52:44,960 --> 00:52:49,360
can you compress a message
without losing it, and how much

634
00:52:49,360 --> 00:52:53,760
redundancy do you need to add so
it survives the noise.

635
00:52:54,360 --> 00:52:59,840
It is a theory of limits and
possibilities, and in a quiet

636
00:52:59,840 --> 00:53:05,240
way it connects to learning.
Because compression and

637
00:53:05,240 --> 00:53:11,160
prediction are close cousins.
If you can predict what comes

638
00:53:11,160 --> 00:53:15,800
next, you do not need to
transmit it in full.

639
00:53:16,440 --> 00:53:21,160
If you cannot predict, you must
send more detail.

640
00:53:21,520 --> 00:53:27,560
This doesn't mean that meaning
disappears, it means that before

641
00:53:27,560 --> 00:53:33,320
meaning there is pattern.
A language has regularities.

642
00:53:33,920 --> 00:53:40,120
A melody has regularities, A
stream of numbers has

643
00:53:40,120 --> 00:53:44,600
regularities.
Once you see inflammation as

644
00:53:44,600 --> 00:53:50,400
pattern plus surprise, you begin
to notice that many problems are

645
00:53:50,400 --> 00:53:54,680
really about capturing
structure, finding what is

646
00:53:54,880 --> 00:54:00,600
stable beneath what is noisy.
The practical results were

647
00:54:00,600 --> 00:54:04,280
immense.
Communication systems improved,

648
00:54:04,640 --> 00:54:09,560
data storage improved, codes and
checksums improved, and a

649
00:54:09,560 --> 00:54:14,920
conceptual bridge was built
between the world of signals and

650
00:54:14,920 --> 00:54:20,440
the world of computation.
Bits became a common currency,

651
00:54:20,800 --> 00:54:25,520
and with a common currency,
different domains could speak to

652
00:54:25,520 --> 00:54:29,760
one another.
A message could be an image, a

653
00:54:29,760 --> 00:54:32,880
sentence, a measurement, a
command.

654
00:54:33,720 --> 00:54:37,640
Everything becomes encodable.
To night.

655
00:54:38,160 --> 00:54:44,000
We let this become a gentle
thought, the world as a stream

656
00:54:44,000 --> 00:54:49,640
of signals and the human mind as
something that has always tried

657
00:54:49,640 --> 00:54:53,520
to find the signal inside the
noise.

658
00:54:53,920 --> 00:54:58,120
We are walking backward.
So soon we will leave the

659
00:54:58,120 --> 00:55:03,120
electronics behind and move
toward the older mathematics

660
00:55:03,120 --> 00:55:07,720
that taught people to live with
uncertainty, probability,

661
00:55:08,080 --> 00:55:14,320
statistics, and the careful art
of inference long before data

662
00:55:14,720 --> 00:55:17,520
meant the modern kind of data at
all.

663
00:55:17,880 --> 00:55:22,960
We step back into the older
habit of making sense of

664
00:55:22,960 --> 00:55:26,920
uncertainty.
Before computers, before

665
00:55:26,920 --> 00:55:32,120
circuits, before stored
programs, people still had to

666
00:55:32,120 --> 00:55:35,960
decide with incomplete
information.

667
00:55:36,760 --> 00:55:41,600
Merchants weighed risk.
Ship captains weighed weather.

668
00:55:42,000 --> 00:55:45,440
Governments counted populations
imperfectly.

669
00:55:46,240 --> 00:55:50,080
Physicians made judgments from
partial symptoms.

670
00:55:50,800 --> 00:55:55,840
The world was never fully known,
so mathematics learned to make

671
00:55:55,840 --> 00:56:00,560
room for maybe.
Statistics, in its broadest

672
00:56:00,560 --> 00:56:04,560
sense, is a language for
describing variation.

673
00:56:05,120 --> 00:56:09,960
Not just the average, but the
spread, not just the typical,

674
00:56:10,440 --> 00:56:15,760
but the outliers.
Over centuries, people develop

675
00:56:15,760 --> 00:56:20,720
tools for summarizing data and
drawing conclusions from

676
00:56:20,720 --> 00:56:26,560
samples, learning that a small
set of observations can hint at

677
00:56:26,560 --> 00:56:30,920
a larger truth but can also
mislead.

678
00:56:31,280 --> 00:56:36,080
The discipline grew alongside
practical needs, astronomy,

679
00:56:36,200 --> 00:56:41,160
navigation, insurance, public
health and economics.

680
00:56:42,240 --> 00:56:46,160
When you measure repeatedly, you
see error.

681
00:56:46,480 --> 00:56:50,160
When you see error, you want a
way to account for it.

682
00:56:50,480 --> 00:56:54,120
One important idea is
estimation.

683
00:56:54,840 --> 00:57:00,000
If you have noisy measurements,
what is the best guess of the

684
00:57:00,080 --> 00:57:04,240
underlying value?
Another is confidence.

685
00:57:04,880 --> 00:57:10,280
How sure can you be, given the
number of observations in their

686
00:57:10,280 --> 00:57:14,200
variability?
These ideas were refined by

687
00:57:14,200 --> 00:57:19,320
generations of mathematicians
and scientists, and later became

688
00:57:19,640 --> 00:57:25,000
essential to machine learning,
because learning is, at heart,

689
00:57:25,400 --> 00:57:30,360
the act of fitting a model to
data that is never perfect.

690
00:57:30,720 --> 00:57:33,200
And there is a humbling lesson
here.

691
00:57:33,560 --> 00:57:38,920
The more carefully you measure,
the more you notice uncertainty

692
00:57:39,520 --> 00:57:43,880
not as a failure, but as a
feature of reality.

693
00:57:44,240 --> 00:57:46,720
The world is not a clean
equation.

694
00:57:47,200 --> 00:57:52,440
It is an ongoing process with
randomness, hidden causes, and

695
00:57:52,440 --> 00:57:57,520
limits to observation.
Statistics teaches you to build

696
00:57:57,520 --> 00:58:03,520
models that are not rigid claims
but structured guesses, useful,

697
00:58:03,680 --> 00:58:08,240
revisable, and aware of their
own fragility as we drift

698
00:58:08,240 --> 00:58:12,120
backward.
Imagine candlelight over a desk,

699
00:58:12,480 --> 00:58:18,440
a hand drawing columns of
figures, a careful mind checking

700
00:58:18,440 --> 00:58:21,840
sums twice.
Long before machines could

701
00:58:21,840 --> 00:58:26,280
learn, humans were already
practicing a kind of learning,

702
00:58:26,760 --> 00:58:30,760
updating their beliefs when
evidence changed, refining

703
00:58:30,760 --> 00:58:37,480
methods, improving predictions.
In a moment, we'll step into the

704
00:58:37,480 --> 00:58:42,960
age when measurement expanded,
when States and institutions

705
00:58:42,960 --> 00:58:47,760
began to collect numbers at
scale, creating the early raw

706
00:58:47,760 --> 00:58:50,280
material from modern data
thinking.

707
00:58:50,680 --> 00:58:54,560
Now we step back into an age of
counting.

708
00:58:55,120 --> 00:58:58,760
Not the fast counting of
computers, but the slow

709
00:58:58,760 --> 00:59:03,600
institutional counting of
societies as they grew more

710
00:59:03,600 --> 00:59:08,120
complex.
Censuses, tax records, shipping

711
00:59:08,120 --> 00:59:11,080
logs, mortality tables, weather
Diaries.

712
00:59:11,600 --> 00:59:16,200
These were early data systems
created for practical reasons,

713
00:59:16,720 --> 00:59:20,680
governance, trade, planning and
survival.

714
00:59:21,360 --> 00:59:26,360
Each record was a small act of
translation, turning lived

715
00:59:26,360 --> 00:59:32,880
reality into symbols on paper.
This era matters because it

716
00:59:32,880 --> 00:59:35,800
changed what people believed was
possible.

717
00:59:36,360 --> 00:59:40,560
When you keep records across
years, patterns emerge.

718
00:59:40,760 --> 00:59:44,600
Prices rise and fall.
Birth rates shift.

719
00:59:45,000 --> 00:59:50,640
Diseases move through cities.
Harvests correlate with

720
00:59:50,640 --> 00:59:54,720
rainfall.
Trade routes respond to war and

721
00:59:54,720 --> 00:59:58,800
policy.
The patterns are never perfect,

722
00:59:59,320 --> 01:00:04,880
but they are visible enough to
guide decisions in a quiet way.

723
01:00:05,440 --> 01:00:10,720
Large scale record keeping
becomes a tool for prediction.

724
01:00:11,080 --> 01:00:13,800
It also shapes the culture of
measurement.

725
01:00:14,440 --> 01:00:18,880
Standard units are adopted.
Instruments improve.

726
01:00:19,480 --> 01:00:25,400
Time keeping becomes more
precise, maps become more

727
01:00:25,440 --> 01:00:29,520
accurate.
The world becomes increasingly

728
01:00:29,520 --> 01:00:35,240
legible to itself, not in a
complete sense, but in a way

729
01:00:35,240 --> 01:00:40,240
that supports coordination.
And coordination is one of the

730
01:00:40,240 --> 01:00:42,840
hidden roots of later
computation.

731
01:00:43,600 --> 01:00:49,560
If many people agree on formats
and standards, information can

732
01:00:49,560 --> 01:00:56,160
travel farther and be reused.
You can feel tonight how gentle

733
01:00:56,160 --> 01:01:00,920
this is.
A Ledger is not a machine, but

734
01:01:00,920 --> 01:01:06,280
it is a memory.
A table is not an algorithm, but

735
01:01:06,280 --> 01:01:10,040
it is a method.
A standardized form is not

736
01:01:10,040 --> 01:01:15,040
intelligence, but it is a kind
of shared structure.

737
01:01:15,360 --> 01:01:20,000
These are the quiet scaffolds
that let later systems scale.

738
01:01:20,560 --> 01:01:26,400
When computers arrive, they will
inherit not only mathematical

739
01:01:26,400 --> 01:01:32,840
ideas but administrative habits.
Categorizing, indexing,

740
01:01:33,000 --> 01:01:38,480
summarizing, revising.
As we continue backward, we

741
01:01:38,480 --> 01:01:44,480
approach a time when probability
itself was being formalized,

742
01:01:45,200 --> 01:01:49,880
when chance stopped being merely
a feeling and became something

743
01:01:49,880 --> 01:01:52,880
you could write down and reason
about.

744
01:01:53,280 --> 01:01:59,200
That shift begins with games,
with wagers, with the simple

745
01:01:59,200 --> 01:02:02,040
human desire to know what is
likely.

746
01:02:02,600 --> 01:02:08,320
Even when the future refuses to
be certain, we step back into

747
01:02:08,320 --> 01:02:14,120
the early mathematics of chance,
when probability began to take a

748
01:02:14,120 --> 01:02:19,520
recognizable shape.
The setting is surprisingly

749
01:02:19,520 --> 01:02:24,760
ordinary.
Games, wagers, questions of

750
01:02:24,760 --> 01:02:29,880
fairness, and the restless
curiosity of people who wanted

751
01:02:29,880 --> 01:02:33,600
to understand what random might
mean.

752
01:02:34,040 --> 01:02:38,760
When you roll dice or draw
cards, you cannot predict the

753
01:02:38,760 --> 01:02:43,200
next outcome, but you can
predict the balance of outcomes.

754
01:02:43,720 --> 01:02:49,800
Over time, that balance, this
steadying tendency, became a

755
01:02:49,800 --> 01:02:53,800
kind of comfort for the mind
that likes order.

756
01:02:54,200 --> 01:02:58,280
Probability gives you a way to
speak about the future without

757
01:02:58,280 --> 01:03:02,400
pretending to control it.
It does not say what will

758
01:03:02,400 --> 01:03:07,680
happen, but what tends to happen
and how strongly.

759
01:03:08,440 --> 01:03:13,720
It teaches you to separate
possibility from likelihood.

760
01:03:14,040 --> 01:03:18,760
It also teaches you to update
when new evidence arrives.

761
01:03:19,080 --> 01:03:22,240
The best view of the world
should change.

762
01:03:22,800 --> 01:03:26,960
This is not indecision, it is
discipline.

763
01:03:27,000 --> 01:03:31,400
A belief that never updates is
not stable.

764
01:03:31,880 --> 01:03:35,920
It is frozen even before later
refinements.

765
01:03:36,160 --> 01:03:40,240
This early era created tools
that would echo forward.

766
01:03:40,920 --> 01:03:45,880
Expected value, for example, is
a way of weighing outcomes by

767
01:03:45,880 --> 01:03:52,080
their probabilities, turning
uncertainty into a single useful

768
01:03:52,080 --> 01:03:57,480
number for decision making.
The law of large numbers is the

769
01:03:57,480 --> 01:04:04,240
quiet reassurance that over many
trials, averages become stable.

770
01:04:05,160 --> 01:04:08,440
These ideas travel far beyond
games.

771
01:04:08,960 --> 01:04:13,800
They shape insurance, finance,
science, and eventually the

772
01:04:13,800 --> 01:04:18,200
mathematical foundations of
learning systems that choose

773
01:04:18,200 --> 01:04:21,880
parameters to minimize expected
error.

774
01:04:22,160 --> 01:04:27,440
And there is a deep link between
probability and prediction.

775
01:04:28,280 --> 01:04:32,520
To predict well is to assign
reasonable probabilities.

776
01:04:33,040 --> 01:04:38,720
To learn from data is to adjust
those probabilities as patterns

777
01:04:38,720 --> 01:04:42,720
become clearer.
Modern machine learning often

778
01:04:42,720 --> 01:04:47,840
looks sophisticated, but beneath
it is this ancient bargain.

779
01:04:48,360 --> 01:04:53,400
We accept uncertainty, and we
build methods for navigating it.

780
01:04:53,840 --> 01:04:58,680
As we drift backward, we can
imagine the gentle clatter of

781
01:04:58,680 --> 01:05:04,000
dice in a wooden cup, the
scratch of ink as someone writes

782
01:05:04,000 --> 01:05:09,640
down outcomes, the patience of
repeating an experiment again

783
01:05:09,640 --> 01:05:13,960
and again.
The night invites the same

784
01:05:13,960 --> 01:05:17,400
patients.
Your mind does not need

785
01:05:17,400 --> 01:05:22,400
certainty right now.
It only needs a soft path to

786
01:05:22,400 --> 01:05:26,800
follow.
And that path leads further back

787
01:05:27,400 --> 01:05:33,080
into the roots of logic, the
older dream that reasoning

788
01:05:33,080 --> 01:05:37,280
itself can be written in
symbols.

789
01:05:37,680 --> 01:05:42,840
Now we step back into the long
tradition of formal reasoning,

790
01:05:43,480 --> 01:05:48,840
where thinkers try to describe
logic as a structure independent

791
01:05:48,840 --> 01:05:53,640
of any particular person.
Long before binary circuits and

792
01:05:53,640 --> 01:05:59,200
programming languages, there
were arguments, syllogisms, and

793
01:05:59,200 --> 01:06:03,600
the careful separation of valid
reasoning from persuasion.

794
01:06:04,680 --> 01:06:10,680
The setting could be a shaded
colonnade, a monastery library,

795
01:06:11,320 --> 01:06:17,720
a small classroom with a slate
board, places where ideas were

796
01:06:17,720 --> 01:06:24,280
refined through repetition.
Logic at its simplest, is the

797
01:06:24,280 --> 01:06:30,760
study of what follows from what.
If certain premises are true,

798
01:06:31,280 --> 01:06:34,480
what conclusions must also be
true?

799
01:06:35,080 --> 01:06:40,560
The power of logic is that it
aims to be impersonal.

800
01:06:41,320 --> 01:06:45,280
It is not about who speaks, but
about structure.

801
01:06:45,640 --> 01:06:49,520
This aspiration would later
become essential for

802
01:06:49,520 --> 01:06:54,280
computation.
Because a computer does not know

803
01:06:54,280 --> 01:07:00,720
what a sentence means in the
human sense, it can only follow

804
01:07:01,240 --> 01:07:06,360
formal rules.
Logic is one of the oldest

805
01:07:06,360 --> 01:07:10,760
attempts to make rules for
thought explicit.

806
01:07:11,120 --> 01:07:17,160
Over centuries, logical systems
were debated, extended, and

807
01:07:17,160 --> 01:07:20,760
sometimes tangled in
philosophical disputes.

808
01:07:21,640 --> 01:07:27,440
But the enduring contribution
was the idea that reasoning can

809
01:07:27,440 --> 01:07:31,600
be represented.
You can write down a form.

810
01:07:31,960 --> 01:07:36,120
You can check whether an
argument matches the form.

811
01:07:36,760 --> 01:07:40,120
You can separate the content
from the structure.

812
01:07:40,520 --> 01:07:46,560
That separation is a quiet kind
of abstraction, and abstraction

813
01:07:46,840 --> 01:07:52,800
is the air computation breathes.
This doesn't mean human thinking

814
01:07:52,800 --> 01:07:57,800
is nothing but logic.
Humans also imagine, feel,

815
01:07:57,880 --> 01:08:04,880
intuit, and improvise.
Yet logic provides a skeleton, a

816
01:08:04,880 --> 01:08:11,080
way to test consistency, to
avoid contradictions, to make a

817
01:08:11,080 --> 01:08:14,480
chain of reasoning stable enough
to build upon.

818
01:08:15,480 --> 01:08:22,319
Science leans on it, law leans
on it, mathematics leans on it,

819
01:08:22,920 --> 01:08:28,560
and later computer science will
lean on it too, even when the

820
01:08:28,560 --> 01:08:33,120
systems being built are
probabilistic rather than

821
01:08:33,120 --> 01:08:37,880
strictly logical.
As we drift backward, logic

822
01:08:37,880 --> 01:08:42,520
begins to look less like a
modern discipline and more like

823
01:08:42,520 --> 01:08:47,399
a craft of language.
Careful definitions, careful

824
01:08:47,399 --> 01:08:53,200
distinctions, careful steps.
It is slow work.

825
01:08:54,000 --> 01:09:02,640
It belongs to quiet rooms, and
from here our path leads to even

826
01:09:02,720 --> 01:09:10,080
older aids for thinking, tools
for calculation, tables and

827
01:09:10,080 --> 01:09:16,479
devices that help people extend
the reach of their minds beyond

828
01:09:16,479 --> 01:09:22,279
what memory alone could hold.
We step back into the world of

829
01:09:22,279 --> 01:09:27,640
calculation before electricity,
when the mind reached for tools

830
01:09:27,720 --> 01:09:31,240
the way a hand reaches for a
railing in the dark.

831
01:09:32,120 --> 01:09:36,680
Long before programmable
computers, there were abacuses,

832
01:09:36,800 --> 01:09:42,439
counting boards, and written
algorithms, procedures carried

833
01:09:42,439 --> 01:09:46,080
in words.
There were logarithm tables that

834
01:09:46,080 --> 01:09:51,479
turned multiplication into
addition, and navigation tables

835
01:09:51,479 --> 01:09:55,760
that helped ships find their way
across open water.

836
01:09:56,840 --> 01:10:01,400
These were not machines that
thought, but they were

837
01:10:01,400 --> 01:10:06,000
extensions of thought, quiet
prosthetics for memory and

838
01:10:06,000 --> 01:10:09,320
arithmetic.
The practical idea here is

839
01:10:09,320 --> 01:10:13,000
simple.
If a task is repeated often

840
01:10:13,000 --> 01:10:18,040
enough, you can codify it.
You can create a method that

841
01:10:18,040 --> 01:10:23,680
anyone can follow step by step
and get the same answer.

842
01:10:24,000 --> 01:10:27,680
This is the beginning of an
algorithmic mindset.

843
01:10:28,320 --> 01:10:34,840
It does not require circuits.
It requires clarity, and clarity

844
01:10:35,280 --> 01:10:38,840
can be written on paper in many
places.

845
01:10:39,200 --> 01:10:42,560
Calculation was not an abstract
hobby.

846
01:10:43,000 --> 01:10:47,560
It was survival.
Merchants needed reliable

847
01:10:47,560 --> 01:10:51,760
arithmetic.
Engineers needed measurements.

848
01:10:52,440 --> 01:10:57,360
Astronomers needed predictions
of planetary positions.

849
01:10:58,160 --> 01:11:03,800
Governments needed accounts.
These needs shaped a culture

850
01:11:03,800 --> 01:11:08,200
that valued accuracy and
repeatability, and that valued

851
01:11:08,200 --> 01:11:12,720
the creation of tools, mental
and physical, that reduced the

852
01:11:12,720 --> 01:11:16,520
chance of error.
You can imagine the texture of

853
01:11:16,520 --> 01:11:22,200
this era, the grain of wood
beneath an Abacus, the smell of

854
01:11:22,240 --> 01:11:27,440
ink, the thinness of paper worn
by repeated handling.

855
01:11:28,360 --> 01:11:32,800
People built tables the way
later engineers would build

856
01:11:32,800 --> 01:11:37,840
software libraries, shared
resources that make future work

857
01:11:37,840 --> 01:11:40,640
easier.
And when a table contains enough

858
01:11:40,640 --> 01:11:44,880
of the world, a kind of
compression happens.

859
01:11:45,680 --> 01:11:51,200
Complex reality becomes a page
you can consult.

860
01:11:51,560 --> 01:11:57,000
As we drift backward,
calculation becomes slower still

861
01:11:57,560 --> 01:12:00,720
and instruments become more
mechanical.

862
01:12:01,320 --> 01:12:06,800
Clocks, gears, escapements,
devices built to keep steady

863
01:12:06,800 --> 01:12:12,680
time and steady motion.
Precision engineering begins to

864
01:12:12,680 --> 01:12:16,400
grow.
It is a quiet prelude to the age

865
01:12:16,640 --> 01:12:21,760
when someone will dream not only
of calculating but of building a

866
01:12:21,760 --> 01:12:26,080
machine that can follow
sequences of instructions,

867
01:12:26,760 --> 01:12:31,880
programmable machinery waiting
just ahead of us on the backward

868
01:12:31,880 --> 01:12:36,160
road.
Now we step back into the craft

869
01:12:36,160 --> 01:12:40,840
of precision, where gears and
springs taught people that

870
01:12:40,840 --> 01:12:46,040
regularity can be engineered.
Time keeping is one of the

871
01:12:46,080 --> 01:12:49,400
oldest demands for dependable
mechanism.

872
01:12:50,320 --> 01:12:55,520
A good clock is a promise.
The world can be measured

873
01:12:55,520 --> 01:13:01,400
consistently, and once you can
build consistent motion, you can

874
01:13:01,400 --> 01:13:04,600
begin to build repeatable
operations.

875
01:13:05,120 --> 01:13:11,160
Turns, clicks, steps, carries.
The physical world becomes a

876
01:13:11,160 --> 01:13:17,480
place where procedures can live.
This is not yet computation in

877
01:13:17,480 --> 01:13:22,560
the modern sense, but it is the
soil that computation grows

878
01:13:22,560 --> 01:13:26,680
from.
Mechanical devices can embody

879
01:13:26,680 --> 01:13:30,760
rules.
A gear ratio is a rule.

880
01:13:31,520 --> 01:13:37,600
An escapement is a rule.
A Cam that lifts a lever at a

881
01:13:37,600 --> 01:13:41,200
particular point in a rotation
is a rule.

882
01:13:41,840 --> 01:13:47,480
In the quiet language of
machinery, cause and effect are

883
01:13:47,480 --> 01:13:52,360
made dependable, and
dependability is what makes long

884
01:13:52,360 --> 01:13:57,200
chains of operations possible
without constant human

885
01:13:57,200 --> 01:14:00,720
correction.
Alongside clocks came other

886
01:14:00,720 --> 01:14:08,280
instruments, balances, sextants,
compasses, measuring chains,

887
01:14:08,760 --> 01:14:13,840
surveying tools.
Each one is a way of turning a

888
01:14:13,840 --> 01:14:17,120
messy world into a readable
quantity.

889
01:14:18,000 --> 01:14:23,080
Each one makes it easier to
coordinate across distance and

890
01:14:23,080 --> 01:14:26,680
time.
When you can trust measurements,

891
01:14:27,080 --> 01:14:32,000
you can trust plans.
When you can trust plans, you

892
01:14:32,000 --> 01:14:36,920
can build larger systems.
Modern computing inherits this

893
01:14:36,920 --> 01:14:41,640
sensibility.
Accuracy matters, repeatability

894
01:14:41,640 --> 01:14:46,160
matters, and error must be
accounted for.

895
01:14:46,600 --> 01:14:50,040
There is also something
psychological here.

896
01:14:50,800 --> 01:14:55,600
A well made mechanism invites a
certain kind of wonder.

897
01:14:56,120 --> 01:15:00,480
It moves with purpose, yet
without intention.

898
01:15:00,760 --> 01:15:07,920
It seems alive though it is not.
That feeling of lifelike motion

899
01:15:07,920 --> 01:15:14,080
emerging from parts is an old
human fascination, and it will

900
01:15:14,080 --> 01:15:18,920
soon lead us into the era of
ambitious mechanical calculation

901
01:15:19,240 --> 01:15:25,040
and early programmability, where
brass and steel begin to imitate

902
01:15:25,040 --> 01:15:30,400
not only motion but method.
For now, let the image be

903
01:15:30,400 --> 01:15:34,280
simple.
A quiet workshop, the faint

904
01:15:34,280 --> 01:15:40,360
sound of a ticking clock, metal
cooling after being shaped, a

905
01:15:40,360 --> 01:15:44,040
steady rhythm that does not
rush.

906
01:15:44,400 --> 01:15:48,720
We are moving backward and the
world is becoming older and

907
01:15:48,720 --> 01:15:53,200
slower, but also clearer in its
foundations.

908
01:15:54,240 --> 01:15:58,880
In the next stretch, we will
approach the 19th century's

909
01:15:58,880 --> 01:16:04,960
dream of programmable machinery
and the careful steps that made

910
01:16:04,960 --> 01:16:09,160
it imaginable.
We've arrived at a century that

911
01:16:09,160 --> 01:16:13,080
loved gears the way our own
century loves code.

912
01:16:13,760 --> 01:16:18,840
In the 1800s, factories were
learning rhythm, railways were

913
01:16:18,840 --> 01:16:22,760
learning schedules, and nations
were learning to count

914
01:16:22,760 --> 01:16:25,320
themselves with ledgers and
forms.

915
01:16:25,640 --> 01:16:30,120
In that same atmosphere of
precision, Charles Babbage

916
01:16:30,120 --> 01:16:35,280
imagined machines that could do
more than calculate a single

917
01:16:35,400 --> 01:16:39,200
table.
He wanted machinery that could

918
01:16:39,200 --> 01:16:46,200
follow a sequence of operations
reliably, repeatably, and with

919
01:16:46,200 --> 01:16:50,840
fewer human mistakes.
His early vision, the Difference

920
01:16:50,840 --> 01:16:56,680
Engine, was aimed at producing
mathematical tables, those long

921
01:16:56,680 --> 01:17:01,640
columns used for navigation and
engineering, without the slip of

922
01:17:01,640 --> 01:17:06,400
a tired hand.
But his larger dream was the

923
01:17:06,480 --> 01:17:12,240
analytical engine, a general
purpose calculating machine that

924
01:17:12,240 --> 01:17:14,560
could be directed by
instructions.

925
01:17:14,920 --> 01:17:21,320
The striking part is that the
idea of programmability appears

926
01:17:21,320 --> 01:17:26,760
here in Metal and motion.
Rather than rewiring the machine

927
01:17:26,760 --> 01:17:32,040
for each task, you could feed it
different instructions and it

928
01:17:32,040 --> 01:17:34,720
would perform different
procedures.

929
01:17:35,600 --> 01:17:41,400
That is a familiar idea now, but
then it was a quiet leap Babbage

930
01:17:41,560 --> 01:17:46,480
drew on the industrial
technologies of his day.

931
01:17:47,200 --> 01:17:51,560
Looms used punched cards to
control patterns in cloth,

932
01:17:52,320 --> 01:17:57,840
turning design into a sequence
of holes and absences.

933
01:17:58,240 --> 01:18:03,920
That same notion, encoding a set
of steps externally, then

934
01:18:03,920 --> 01:18:09,880
letting a machine execute them,
became one of the inspirations

935
01:18:09,880 --> 01:18:15,200
for early programmable devices.
The machine's gears would carry

936
01:18:15,200 --> 01:18:20,400
numbers, but the program would
be carried by the cards, an

937
01:18:20,400 --> 01:18:24,200
early separation between
hardware and instructions.

938
01:18:24,520 --> 01:18:30,040
Ada Lovelace, working with
Babbage's concepts, offered an

939
01:18:30,040 --> 01:18:36,920
insight that echoes forward.
She understood that if a machine

940
01:18:36,920 --> 01:18:41,960
can manipulate symbols according
to rules, then it may not be

941
01:18:41,960 --> 01:18:47,600
limited to arithmetic alone.
Numbers can represent other

942
01:18:47,600 --> 01:18:53,040
things, notes in music, letters
in a language, steps in a

943
01:18:53,040 --> 01:18:57,560
process.
In modern terms, this is the

944
01:18:57,560 --> 01:19:01,680
idea that computation is not
bound to one domain.

945
01:19:02,360 --> 01:19:07,800
It is a general method applied
to encoded representations.

946
01:19:08,160 --> 01:19:11,680
It's important to keep the scale
gentle.

947
01:19:12,400 --> 01:19:16,720
These machines were not fully
built in the way Babbage hoped,

948
01:19:17,360 --> 01:19:21,840
and they did not transform
society in his lifetime.

949
01:19:22,640 --> 01:19:27,800
But the concept they introduced
is foundational, a machine that

950
01:19:27,800 --> 01:19:34,320
can be instructed, a system that
can execute a method described

951
01:19:34,320 --> 01:19:39,440
in a form it can read.
Tonight you can imagine brass

952
01:19:39,440 --> 01:19:45,320
turning slowly, card stacked
like thin leaves, and a mind

953
01:19:45,320 --> 01:19:50,200
realizing that procedure itself
can be mechanized.

954
01:19:50,560 --> 01:19:55,720
As we continue backward will
meet an even more abstract

955
01:19:55,720 --> 01:19:59,840
shift.
Logic becoming algebra,

956
01:20:00,560 --> 01:20:05,880
reasoning becoming something you
can write as symbols that behave

957
01:20:06,280 --> 01:20:12,680
predictably.
The metal will fade and the idea

958
01:20:13,360 --> 01:20:17,120
will remain.
And in a sleep wise kind of way,

959
01:20:17,120 --> 01:20:20,360
there is comfort in how modest
this beginning is.

960
01:20:20,960 --> 01:20:25,880
No flashing screens, no rush,
only careful drawings and

961
01:20:25,880 --> 01:20:29,280
patient thought.
The future arrives here as a

962
01:20:29,280 --> 01:20:34,360
blueprint, not a spectacle.
You can let that be reassuring.

963
01:20:34,760 --> 01:20:40,480
Progress is often quiet at 1st,
and the night is good at holding

964
01:20:40,480 --> 01:20:45,120
quiet things.
The workshop air clears and we

965
01:20:45,120 --> 01:20:48,920
find ourselves in a different
kind of construction site.

966
01:20:49,520 --> 01:20:54,120
Not gears on a table, but
symbols on a page.

967
01:20:55,000 --> 01:21:00,800
In the 19th century, George
Boule pursued a strange, elegant

968
01:21:00,800 --> 01:21:05,920
ambition to treat logic as a
form of mathematics.

969
01:21:06,280 --> 01:21:12,960
Instead of arguing in sentences,
you could write reasoning in a

970
01:21:12,960 --> 01:21:19,280
symbolic language and manipulate
it with consistent rules.

971
01:21:19,880 --> 01:21:26,960
The way you manipulate numbers.
At first it can sound dry, true

972
01:21:26,960 --> 01:21:33,440
and false and or not, but the
leap is profound.

973
01:21:34,320 --> 01:21:39,360
If logical statements can be
represented in a formal system,

974
01:21:39,800 --> 01:21:44,120
then reasoning becomes something
you can calculate.

975
01:21:45,040 --> 01:21:49,160
Bool's algebra takes
propositions and gives them

976
01:21:49,160 --> 01:21:52,800
operations.
You can combine statements,

977
01:21:52,920 --> 01:21:58,000
negate them, simplify them, and
check whether one follows from

978
01:21:58,000 --> 01:22:01,040
another.
It is a way of turning the

979
01:22:01,040 --> 01:22:06,920
structure of thought into a kind
of machinery, except the

980
01:22:06,920 --> 01:22:12,000
machinery is abstract.
This is one of the deepest roots

981
01:22:12,000 --> 01:22:16,520
of later computing.
Electronic circuits can be built

982
01:22:16,720 --> 01:22:22,400
so that switches represent
states on and off, and

983
01:22:22,400 --> 01:22:28,040
combinations of switches
implement logical operations.

984
01:22:28,360 --> 01:22:33,760
Bool did not design modern
computers, but his framework

985
01:22:33,920 --> 01:22:37,680
made it natural to think of
logic as something that can be

986
01:22:37,680 --> 01:22:41,280
embodied in a system of
operations.

987
01:22:42,520 --> 01:22:48,320
The distance from a page of
symbols to a network of gates is

988
01:22:48,320 --> 01:22:53,920
shorter than it appears because
both are about reliable

989
01:22:53,920 --> 01:22:57,880
transformations.
There is also a more human

990
01:22:57,880 --> 01:23:02,480
lesson here.
Formal logic does not capture

991
01:23:02,560 --> 01:23:08,080
everything we do when we think,
but it offers a standard of

992
01:23:08,080 --> 01:23:13,160
clarity.
It separates persuasion from

993
01:23:13,160 --> 01:23:15,960
validity.
It lets you see when a

994
01:23:15,960 --> 01:23:20,760
conclusion is supported by
premises, and when it is merely

995
01:23:20,760 --> 01:23:23,440
surrounded by confident
language.

996
01:23:24,200 --> 01:23:30,800
In science and engineering, that
clarity becomes a kind of safety

997
01:23:30,800 --> 01:23:34,760
rail.
It keeps complex reasoning from

998
01:23:34,760 --> 01:23:39,480
collapsing under its own weight
in the broader backward story.

999
01:23:39,720 --> 01:23:43,160
This is where method becomes
portable.

1000
01:23:43,960 --> 01:23:48,200
A logical structure can be
copied, taught, checked, and

1001
01:23:48,320 --> 01:23:52,000
reused.
It does not depend on one

1002
01:23:52,000 --> 01:23:56,840
person's memory or intuition.
And once methods are portable,

1003
01:23:57,240 --> 01:24:03,000
they can be layered, one method
supporting another, until a

1004
01:24:03,000 --> 01:24:06,880
whole architecture of reasoning
emerges.

1005
01:24:07,240 --> 01:24:12,160
Tonight, notice how soothing
that portability can feel.

1006
01:24:12,960 --> 01:24:17,000
Your mind does not need to hold
everything at once.

1007
01:24:17,400 --> 01:24:21,880
It can rest on structures.
It can let one thought support

1008
01:24:21,880 --> 01:24:27,200
the next without strain.
And as we keep moving backward,

1009
01:24:27,480 --> 01:24:32,680
the symbols will loosen from
strict true and false and enter

1010
01:24:32,680 --> 01:24:39,040
the softer world of uncertainty,
where beliefs can be updated and

1011
01:24:39,040 --> 01:24:44,800
knowledge is expressed not as
certainty but as probability.

1012
01:24:45,080 --> 01:24:49,800
Imagine a quiet desk lamp
pooling light over a page of

1013
01:24:49,800 --> 01:24:54,920
tidy notation, the ink still
slightly raised where it has

1014
01:24:54,920 --> 01:24:59,200
dried.
In that small circle of light,

1015
01:24:59,840 --> 01:25:06,280
an idea is settling into place
that some parts of reasoning can

1016
01:25:06,280 --> 01:25:11,040
be made dependable.
Like a well built clock, the

1017
01:25:11,040 --> 01:25:16,000
world outside can be ambiguous,
but inside the symbol system

1018
01:25:16,520 --> 01:25:21,960
each step is crisp, and that
crispness will later help

1019
01:25:21,960 --> 01:25:27,920
machines act consistently.
Before we reach the oldest myths

1020
01:25:27,920 --> 01:25:34,440
and mechanisms, we pass through
a quiet doorway that many modern

1021
01:25:34,440 --> 01:25:39,880
fields share.
The mathematics of updating your

1022
01:25:39,880 --> 01:25:43,880
mind.
Thomas Bayes is often associated

1023
01:25:43,880 --> 01:25:49,840
with a simple, powerful idea.
When new evidence arrives, A

1024
01:25:49,840 --> 01:25:54,240
rational belief should change in
a specific way.

1025
01:25:55,160 --> 01:25:59,640
The details can be written as a
formula, but the spirit is

1026
01:25:59,640 --> 01:26:03,080
gentle.
Start with what you currently

1027
01:26:03,080 --> 01:26:07,520
believe, then revise it in
proportion to how strongly the

1028
01:26:07,520 --> 01:26:10,400
new evidence supports or
contradicts it.

1029
01:26:10,840 --> 01:26:16,080
Bayesian thinking did not appear
fully formed in one moment.

1030
01:26:16,760 --> 01:26:22,640
It sits within a broader history
of probability, statistics and

1031
01:26:22,680 --> 01:26:29,360
inference that grew from
astronomy, navigation, games of

1032
01:26:29,360 --> 01:26:34,160
chance, insurance, and
scientific measurement.

1033
01:26:34,520 --> 01:26:39,360
Over time, people realized that
uncertainty is not merely

1034
01:26:39,440 --> 01:26:44,360
ignorance, it is a measurable
feature of decision making.

1035
01:26:45,080 --> 01:26:49,960
If you can assign probabilities,
you can compare choices.

1036
01:26:50,640 --> 01:26:54,800
If you can update probabilities,
you can learn.

1037
01:26:55,200 --> 01:27:00,000
This is foundational to many
approaches in machine learning,

1038
01:27:00,640 --> 01:27:06,960
especially those that treat
models as hypotheses and data.

1039
01:27:07,320 --> 01:27:11,240
As evidenced, even when modern
systems are trained with

1040
01:27:11,240 --> 01:27:15,600
different techniques, the
conceptual link remains.

1041
01:27:16,200 --> 01:27:21,720
Learning is, at heart, the
improvement of expectations.

1042
01:27:22,120 --> 01:27:27,880
In the presence of uncertainty.
You do not need perfect

1043
01:27:27,880 --> 01:27:31,040
knowledge.
You need a disciplined way to

1044
01:27:31,040 --> 01:27:36,080
move from less informed guesses
to better informed ones.

1045
01:27:36,480 --> 01:27:40,320
Probability also teaches
humility.

1046
01:27:41,320 --> 01:27:45,720
A high probability outcome is
not guaranteed.

1047
01:27:46,680 --> 01:27:53,120
A low probability outcome is not
impossible in a world full of

1048
01:27:53,120 --> 01:27:58,920
randomness and hidden causes.
Probabilistic models offer a

1049
01:27:58,920 --> 01:28:05,680
kind of calm realism.
They let you say this is likely

1050
01:28:06,240 --> 01:28:09,240
without pretending to be
omniscient.

1051
01:28:10,160 --> 01:28:13,440
They also invite better
questions.

1052
01:28:13,960 --> 01:28:18,800
What evidence would change my
mind and how much?

1053
01:28:19,200 --> 01:28:23,760
Historically, these ideas
matured as people gathered more

1054
01:28:23,760 --> 01:28:26,800
measurements.
Astronomers compared

1055
01:28:26,800 --> 01:28:30,200
observations to predicted
orbits.

1056
01:28:30,800 --> 01:28:35,840
Surveyors dealt with error.
Physicians compared treatments.

1057
01:28:36,400 --> 01:28:40,880
Governments counted populations
and discovered drift.

1058
01:28:41,760 --> 01:28:47,840
Each domain needed tools to
separate signal from noise, and

1059
01:28:47,840 --> 01:28:53,000
probability supplied a language
for that separation.

1060
01:28:53,320 --> 01:28:57,840
As we move backward tonight, let
that language soften into

1061
01:28:57,840 --> 01:29:04,320
something almost soothing.
Updating is a kind of release.

1062
01:29:04,880 --> 01:29:07,800
You do not have to cling to an
earlier belief.

1063
01:29:08,280 --> 01:29:14,320
If the world shows you something
new, you can adjust, you can

1064
01:29:14,320 --> 01:29:18,800
loosen, you can let your mind
become lighter.

1065
01:29:19,280 --> 01:29:24,360
And with that lightness we can
drift further back, away from

1066
01:29:24,360 --> 01:29:29,640
equations and into the older
human fascination with lifelike

1067
01:29:29,640 --> 01:29:36,760
machines, automata that moved,
sang and startled people into

1068
01:29:36,760 --> 01:29:40,680
wonder.
In practical terms, a prior is

1069
01:29:40,680 --> 01:29:45,680
simply a starting point, and a
posterior is the new view.

1070
01:29:45,680 --> 01:29:51,280
After learning tonight, you can
treat your own thoughts the same

1071
01:29:51,280 --> 01:29:55,360
way.
Begin where you are, then gently

1072
01:29:55,360 --> 01:30:00,600
update toward rest, like a tide
correcting its line along the

1073
01:30:00,600 --> 01:30:04,240
shore.
Each small revision brings you

1074
01:30:04,240 --> 01:30:08,680
closer to stillness.
Long before anyone spoke of

1075
01:30:08,680 --> 01:30:15,400
algorithms, there were machines
made to astonish automata.

1076
01:30:16,080 --> 01:30:21,760
Devices that imitate life
through clever mechanisms appear

1077
01:30:21,760 --> 01:30:28,920
again and again across history.
In ancient Greece, engineers

1078
01:30:28,920 --> 01:30:35,000
like Hero of Alexandria
described pneumatic and

1079
01:30:35,000 --> 01:30:40,120
mechanical contraptions, doors
that opened seemingly by

1080
01:30:40,120 --> 01:30:46,120
themselves, figures that moved
in temples, little theatrical

1081
01:30:46,120 --> 01:30:50,680
scenes powered by weights, water
and hidden tubes.

1082
01:30:51,000 --> 01:30:55,080
Later centuries brought
clockwork birds, mechanical

1083
01:30:55,080 --> 01:30:59,640
musicians and intricate dolls
that could write or play

1084
01:30:59,640 --> 01:31:04,680
instruments, their motions
guided by cams and gears.

1085
01:31:05,080 --> 01:31:10,960
The point of an automaton was
not productivity, it was wonder.

1086
01:31:11,880 --> 01:31:16,720
It asked a quiet question
without speaking it aloud.

1087
01:31:17,480 --> 01:31:22,440
If movement can be produced by
parts, what else might be

1088
01:31:22,440 --> 01:31:26,560
produced by parts?
People knew these creations were

1089
01:31:26,560 --> 01:31:31,840
not alive and still.
The mind tends to lean toward

1090
01:31:31,840 --> 01:31:35,280
life when it sees coordinated
motion.

1091
01:31:35,840 --> 01:31:43,800
A bird that flutters, a head
that turns, a hand that lifts.

1092
01:31:44,640 --> 01:31:49,720
The illusion is thin, and yet it
is emotionally potent.

1093
01:31:50,160 --> 01:31:55,120
Technically, automata are
systems of stored behavior.

1094
01:31:55,920 --> 01:31:59,280
A Cam profile encodes emotion
over time.

1095
01:31:59,880 --> 01:32:03,880
A gear train transforms one
rhythm into another.

1096
01:32:04,160 --> 01:32:09,520
A spring stores energy, then
releases it in a controlled way.

1097
01:32:10,160 --> 01:32:15,800
In modern language, you might
call this a program, except the

1098
01:32:15,800 --> 01:32:22,080
program is carved into metal.
The behaviour repeats because

1099
01:32:22,280 --> 01:32:27,120
the structure repeats.
There is no learning here, no

1100
01:32:27,120 --> 01:32:32,680
adaptation, only design.
But the idea that behaviour can

1101
01:32:32,680 --> 01:32:38,280
be encoded is a crucial ancestor
of later computation.

1102
01:32:38,640 --> 01:32:41,600
Automata also influenced
culture.

1103
01:32:42,360 --> 01:32:47,680
They appeared in courts and
exhibitions, symbols of mastery

1104
01:32:47,680 --> 01:32:51,240
over matter.
They lived in stories as well.

1105
01:32:51,640 --> 01:32:57,440
Artificial servants, enchanted
statues, crafted companions.

1106
01:32:58,400 --> 01:33:04,240
Even when the details were
fantastical, the fascination was

1107
01:33:04,240 --> 01:33:08,360
real.
Humans have long imagined minds

1108
01:33:08,360 --> 01:33:12,120
and bodies as things that might
be assembled.

1109
01:33:12,520 --> 01:33:16,040
Tonight.
We can hold that fascination

1110
01:33:16,040 --> 01:33:19,960
gently without turning it into
drama.

1111
01:33:21,040 --> 01:33:25,920
Imagine a quiet hall after
visitors have gone home.

1112
01:33:26,760 --> 01:33:30,120
A mechanical bird sits still on
its perch.

1113
01:33:31,120 --> 01:33:37,280
A clockwork musician is silent.
The gears are at rest and the

1114
01:33:37,280 --> 01:33:42,320
illusion sleeps.
In that stillness, we can sense

1115
01:33:42,320 --> 01:33:47,200
why these devices mattered.
They trained the imagination to

1116
01:33:47,200 --> 01:33:50,400
see motion as a result of
method.

1117
01:33:50,720 --> 01:33:55,200
And as we drift backward one
last time, we leave the

1118
01:33:55,200 --> 01:33:58,760
workshops and exhibition halls
behind.

1119
01:33:59,560 --> 01:34:03,600
We return to the oldest
technology of all, shared

1120
01:34:03,600 --> 01:34:09,560
stories passed from voice to
voice, compressing experience

1121
01:34:09,560 --> 01:34:15,160
into language so that a
community can remember what one

1122
01:34:15,160 --> 01:34:21,040
person alone cannot.
In the 18th century, makers like

1123
01:34:21,040 --> 01:34:27,680
the Jacques Draws family built
celebrated automata, the writer,

1124
01:34:27,880 --> 01:34:32,720
the draftsman, the musician,
showing how far meticulous

1125
01:34:32,720 --> 01:34:36,040
mechanism could go when paired
with patience.

1126
01:34:36,480 --> 01:34:41,720
The lesson was not that metal
had become human, but that

1127
01:34:41,720 --> 01:34:46,280
sequence and timing could be
captured, preserved, and

1128
01:34:46,280 --> 01:34:50,200
replayed with exquisite
fidelity.

1129
01:34:50,600 --> 01:34:53,240
At the far end of our backward
walk.

1130
01:34:53,720 --> 01:34:57,960
The world grows older than
engines and older than ink.

1131
01:34:58,760 --> 01:35:04,400
Here, intelligence is not a
machine at all, but a habit

1132
01:35:05,080 --> 01:35:09,840
practiced by people who had to
read the environment carefully

1133
01:35:10,240 --> 01:35:13,720
before formal mathematics.
There were patterns.

1134
01:35:14,280 --> 01:35:17,680
The stars rose and fell in
familiar roots.

1135
01:35:18,360 --> 01:35:23,720
The moon waxed and thinned.
Seasons returned, not perfectly

1136
01:35:24,080 --> 01:35:27,720
but reliably enough to guide
planting and migration.

1137
01:35:28,040 --> 01:35:32,360
Wind changed with geography.
Tides followed rhythms.

1138
01:35:32,680 --> 01:35:39,400
Animals moved with weather.
To notice these regularities was

1139
01:35:39,400 --> 01:35:43,080
to survive in this earliest
layer.

1140
01:35:43,440 --> 01:35:46,080
Prediction lived inside the
body.

1141
01:35:46,560 --> 01:35:50,560
You learned the feel of
approaching rain in the air's

1142
01:35:50,560 --> 01:35:53,880
pressure.
You learned the smell of water

1143
01:35:53,880 --> 01:35:57,760
near by.
You learned which clouds meant a

1144
01:35:57,760 --> 01:36:01,520
storm and which meant nothing at
all.

1145
01:36:02,280 --> 01:36:06,920
Then, slowly, communities learn
to share these observations.

1146
01:36:07,680 --> 01:36:11,720
One person's experience became a
group's memory.

1147
01:36:12,560 --> 01:36:16,600
That is where language begins to
behave like a tool.

1148
01:36:17,200 --> 01:36:22,040
It carries more than emotion.
It carries compressed knowledge.

1149
01:36:22,320 --> 01:36:27,240
A story can hold a map.
A myth can hold a warning.

1150
01:36:27,800 --> 01:36:31,240
A proverb can hold a rule of
thumb.

1151
01:36:32,360 --> 01:36:35,480
Oral tradition is not random
entertainment.

1152
01:36:35,920 --> 01:36:41,120
It is a storage system tuned for
minds that remember through

1153
01:36:41,120 --> 01:36:46,880
rhythm and repetition.
In that sense, storytelling is

1154
01:36:46,880 --> 01:36:52,120
an ancient form of compression.
It removes unnecessary detail

1155
01:36:52,320 --> 01:36:58,240
and keeps what tends to matter,
the pattern beneath the noise.

1156
01:36:59,000 --> 01:37:04,120
This is one reason the human
mind is so at home in narrative.

1157
01:37:04,920 --> 01:37:10,640
We'd learned long ago to store
the world in sequences of words.

1158
01:37:11,000 --> 01:37:15,840
When you look back from modern
AI to this older place, a

1159
01:37:15,840 --> 01:37:21,320
continuity appears.
Machines today learn from data

1160
01:37:22,000 --> 01:37:28,600
by adjusting internal parameters
humans long ago learned from

1161
01:37:28,600 --> 01:37:33,320
experience by adjusting
expectations.

1162
01:37:33,680 --> 01:37:39,200
Both are forms of fitting ways
of becoming better at

1163
01:37:39,200 --> 01:37:45,240
anticipating what comes next.
The difference is not only power

1164
01:37:45,240 --> 01:37:51,560
or speed, but meaning.
Human learning is tangled with

1165
01:37:51,560 --> 01:37:55,560
desire, responsibility, and
care.

1166
01:37:56,400 --> 01:38:00,560
It changes who we are, not just
what we predict.

1167
01:38:00,960 --> 01:38:05,000
Tonight, we can let that be a
calming thought.

1168
01:38:05,840 --> 01:38:10,520
The newest tools are built on
the oldest materials, pattern,

1169
01:38:10,600 --> 01:38:15,560
language, memory, and the
patient urge to make sense of an

1170
01:38:15,560 --> 01:38:19,400
uncertain world.
And if the modern world

1171
01:38:19,400 --> 01:38:24,520
sometimes feels too bright with
its inventions, you can remember

1172
01:38:24,520 --> 01:38:30,000
how deep the roots really are,
and how quietly they began.

1173
01:38:30,320 --> 01:38:35,400
Somewhere long ago.
A small group sat near firelight

1174
01:38:35,720 --> 01:38:40,800
and listened as a voice traced
the shape of a river, the path

1175
01:38:40,800 --> 01:38:44,440
of a herd, the safest crossing
after rain.

1176
01:38:45,280 --> 01:38:48,600
The listener did not download
knowledge.

1177
01:38:49,000 --> 01:38:53,600
They absorbed it slowly with
trust and carried it forward.

1178
01:38:54,120 --> 01:38:57,840
Tonight, you can do the same
with calmness.

1179
01:38:58,200 --> 01:39:02,640
Let the timeline fade.
Let the gears stop turning.

1180
01:39:02,960 --> 01:39:07,240
Let the cymbals loosen and fall
like soft snow.

1181
01:39:08,480 --> 01:39:11,160
Your mind doesn't have to solve
anything.

1182
01:39:11,440 --> 01:39:16,360
It only has to rest, held by the
same gentle rhythms that held

1183
01:39:16,360 --> 01:39:21,240
our ancestors.
The turning sky, the returning

1184
01:39:21,240 --> 01:39:25,120
seasons, the slow breath in the
dark.

1185
01:39:25,520 --> 01:39:26,320
Good night.